250M params
System Gating Network] CL -->|"Route"| M1[Coding Specialist
27B params] CL -->|"Route"| M2[Insurance Specialist
7B params] CL -.->|"Not loaded"| M3[Jira Specialist
3B params] CL -.->|"Not loaded"| M4[General Knowledge
14B params] CL -.->|"Not loaded"| M5[...] CL -.->|"Not loaded"| M6[Model 12] M1 --> JP[Judgment Panel] M2 --> JP JP --> VP[Verification Panel] VP --> O2[Verified Response] style M3 fill:#ccc,stroke:#999,stroke-dasharray: 5 5 style M4 fill:#ccc,stroke:#999,stroke-dasharray: 5 5 style M5 fill:#ccc,stroke:#999,stroke-dasharray: 5 5 style M6 fill:#ccc,stroke:#999,stroke-dasharray: 5 5 end ``` This matters because it combines the efficiency gains of MoE with properties that intra-model MoE cannot provide: each specialist can be independently trained, fine-tuned, replaced, or upgraded without touching the rest of the system. A new domain (say, healthcare compliance) is added by training a new specialist and registering it in the catalogue -- not by retraining a trillion-parameter model. The cost of adding a 7B specialist is \$50K-\$500K. The cost of retraining a frontier model is \$500M or more. The research literature has begun to formalise this concept. Quirke et al.'s "Beyond Monoliths: Expert Orchestration for More Capable, Democratic, and Safe Language Models" (submitted to NeurIPS 2025) argues that expert orchestration delivers superior performance to monolithic models, with clearer evaluation metrics, narrower input spaces for testing, and more effective application of specialised human expertise. Chai et al.'s "An Expert is Worth One Token" (ACL 2024) demonstrated that representing expert LLMs as tokens in a meta-LLM's vocabulary -- literally lifting MoE routing from layers to entire models -- outperforms existing multi-LLM collaboration paradigms across six expert domains. ## Architecture Deep Dive ### The Cognition Stream Every component in Annie communicates through a single backbone: the Cognition Stream, implemented on Bellerophon BStream. This is a deliberate architectural choice. Rather than building custom orchestration logic to coordinate twelve models, a classifier, judgment panels, and verification panels, Annie treats the entire pipeline as a series of messages flowing through a durable, observable stream. Bellerophon provides several properties that would be expensive and error-prone to build from scratch: * **Durability**: Every message (prompt, classification, expert response, judgment, verification result) is persisted. Nothing is lost if a component crashes or is restarted. * **Replay**: Any point in the pipeline can be replayed from the stream. This makes debugging straightforward -- you can trace exactly what the classifier decided, what each expert produced, how the judgment panel scored them, and whether verification passed or failed. * **Backpressure**: If experts are slower than incoming prompts, the stream manages queueing and flow control rather than dropping requests or overwhelming GPU memory. * **Observability**: Every stage transition is a stream event. Latency, throughput, error rates, and routing decisions are measurable without custom instrumentation. * **Decoupling**: Components subscribe to the message types they care about. The classifier does not need to know which experts are loaded. Experts do not need to know about the judgment panel. This makes the system genuinely composable. The alternative -- direct API calls between components with custom retry logic, circuit breakers, and state management -- would produce a system that works identically in the happy path but fails unpredictably under load, is opaque to debugging, and requires custom code for every new interaction pattern. Every mature distributed system eventually builds or adopts a messaging backbone. Annie starts with one. ```mermaid theme={null} graph LR subgraph "Cognition Stream Pipeline" direction LR GW["Incoming Gateway
Keyword extraction
Shape analysis
Parameter capture"] CL["Classifier
~250M params
Type, complexity,
priority"] subgraph "Expert Pool (1-6 active)" EX1["Expert 1"] EX2["Expert 2"] EX3["Expert N"] end JP["Judgment Panel
2-4 domain models
Rubric scoring"] VP["Verification Panel
2-4 models
Prompt fulfilment
consensus"] OP["Outbound Processing
Deterministic
Format application"] GW -->|"BStream"| CL CL -->|"BStream"| EX1 CL -->|"BStream"| EX2 CL -->|"BStream"| EX3 EX1 -->|"BStream"| JP EX2 -->|"BStream"| JP EX3 -->|"BStream"| JP JP -->|"BStream"| VP VP -->|"Pass"| OP VP -.->|"Fail: Re-enter"| GW end SY["Synapse Context Engine
Prior interactions
Knowledge bases
Connected sources"] SY -.->|"Context injection"| EX1 SY -.->|"Context injection"| EX2 SY -.->|"Context injection"| EX3 RP["Rapport Model
Relationship state
Tone calibration"] RP -.->|"Tone signal"| OP ``` ### The Classification Layer The classifier is the smallest model in the system (\~250M parameters) and the most critical. It makes two decisions that determine everything downstream: how complex is this prompt, and what domain(s) does it belong to? **Fast path**: Low-complexity, general queries -- greetings, simple factual lookups, clarification questions -- skip the consensus pipeline entirely. A single appropriate expert responds directly. This is how Annie achieves sub-second latency for the majority of interactions. Research from FrugalGPT (Chen et al., 2023) and RouteLLM (UC Berkeley et al., 2024) demonstrates that intelligent routing can deliver 85-98% cost reduction while maintaining quality, precisely because most queries do not require the full weight of the system. **Full pipeline**: High-complexity knowledge work, domain-specific tasks, anything with consequences -- flows through expert processing, judgment, and verification. The latency cost is measured in seconds to minutes, but the output is consensus-verified. This is not a novel pattern. It is the same complexity-based routing that every well-designed customer service operation uses (tier 1 handles simple queries, escalates complex ones) and that every MoE model uses internally (route easy tokens cheaply, hard tokens to more experts). Annie applies it at the system level with the classifier as the gating function. The classifier's small size is a feature: it loads instantly, runs on minimal hardware, and can be retrained cheaply as the domain catalogue evolves. At 250M parameters, it adds negligible latency to every request while determining whether the response requires \$0.001 or \$0.10 of compute. ### Expert Orchestration Annie maintains a catalogue of twelve specialist models (today), of which one to six are loaded into GPU or CPU memory at any given time. The classifier's complexity and domain signals determine which experts engage and how many work in parallel on the same prompt. For a low-complexity coding question, a single coding specialist responds. For a high-complexity insurance coverage determination that touches regulatory compliance, multiple specialists may engage in parallel: the insurance domain expert, a regulatory specialist, and a general reasoning model. Each expert writes its response back to the Cognition Stream independently. **Why small specialists beat one big generalist on domain tasks**: The research is clear on this point. Microsoft's Phi-3 Technical Report (2024) showed that a 3.8B parameter model trained on curated data rivals Mixtral 8x7B and GPT-3.5 on standard benchmarks. Research published at arXiv (2505.24189, 2025) found that fine-tuning a small language model can outperform prompting a frontier LLM on domain-specific tasks. Apple's production architecture uses a \~3B on-device model with runtime-swappable LoRA adapters for task specialisation -- effectively system-level MoE deployed at billion-device scale. The efficiency argument is quantitative. Kondratyuk et al. (Google AI, 2020) demonstrated that on ImageNet, an ensemble of two EfficientNet-B5 models matches EfficientNet-B7 accuracy using approximately 50% fewer FLOPs, with the efficiency gap widening as models get larger. SLM-MUX (Wang et al., 2024) showed that just two small language models can outperform Qwen 2.5 72B on GPQA and GSM8K. The "Blending Is All You Need" study (Chai Research, 2024) demonstrated that an ensemble of three models totalling \~25B parameters outperformed ChatGPT (175B+) in real-world user retention and engagement over 30 days of A/B testing. **Model loading and unloading**: Not all twelve specialists need to be resident in memory simultaneously. The classifier's domain signals determine which models are needed. Frequently-used specialists remain loaded; niche specialists are loaded on demand. With modern quantisation (Q4/Q8) and frameworks like vLLM, a 7B model loads in seconds. The Cognition Stream's backpressure management ensures prompts queue gracefully during model loading rather than failing. **Tool use during expert processing**: During inference, Annie's specialist models operate in two tiers. Internal platform tools (Synapse context queries, knowledge base searches, internal platform APIs) execute immediately as part of expert processing. External tools (reading files on customer systems, calling customer APIs, writing and executing code in external environments, querying external databases) are returned as tool call requests in the expert's response -- similar to how frontier models return tool\_use blocks. These external requests are resolved at judgment time and executed by the client, not by Annie. All tool inputs and outputs flow through the same judgment and verification pipeline as generated text responses, ensuring that all external actions are evaluated and verified before reaching the user. ### Score-Based Consensus When multiple experts produce responses to the same prompt, the Judgment Panel evaluates them. This is not a simple majority vote. Two to four domain-specific models are selected based on the prompt's classification, and each scores the expert responses against rubrics -- structured evaluation criteria specific to the domain and task type. The highest-scoring response wins. This could be a single expert's response used verbatim, or an amalgamation that draws from multiple expert outputs. The judgment is published back to the Cognition Stream with full scoring details, making every decision auditable. **Why consensus beats single-pass inference**: The evidence is substantial. Wang et al.'s "Self-Consistency" paper (ICLR 2023) demonstrated that sampling multiple reasoning paths and selecting the most consistent answer improves accuracy by 17.9% on GSM8K, 11.0% on SVAMP, and 12.2% on AQuA. Verga et al.'s Panel of LLM Evaluators (PoLL) showed that a panel of three small, diverse models from different providers outperformed a single large judge across six datasets while reducing intra-model bias at significantly lower cost. The Language Model Council (Zhao et al., 2024) demonstrated that a council of LLMs that evaluate each other produces rankings more robust and less biased than any individual judge. **Why rubrics matter**: Research from "Beyond the Illusion of Consensus" (2026) revealed that model-level agreement masks fragile sample-level agreement, and that dynamically generated rubrics with domain-knowledge grounding increase agreement by 22-27% depending on domain. "Rubric Is All You Need" (ACM ICER 2025) confirmed that question-specific rubrics outperform generic evaluation criteria. Annie's rubrics are domain-specific by design -- an insurance coverage determination is judged against different criteria than a code review. **The independence problem -- and Annie's structural advantage**: A critical 2025 study ("Nine Judges, Two Effective Votes") found that nine frontier LLMs from seven model families provide only about two independent votes' worth of information because they make correlated errors. This dramatically limits the theoretical gains from consensus. Annie has a structural advantage here: its specialists are genuinely different models, trained on different data for different domains, with different architectures and parameter counts. A 3B insurance specialist and a 14B coding model and a 7B general knowledge model are far less likely to share correlated failure modes than three versions of GPT-4. The literature confirms this: diverse panels from different providers outperform homogeneous panels, and cross-model probes significantly enhance error detection where within-model consistency fails (Tan et al., EMNLP 2025). ### Closed-Loop Verification After the Judgment Panel selects or synthesises a response, the Verification Panel performs a distinct function: it compares the judged response against the original prompt and determines whether the response actually fulfils what was asked. Two to four models must reach consensus on this binary question. If verification passes, the response is published to the outbound stream. If it fails, the original prompt re-enters the pipeline from scratch, and the current result is discarded. **Why this matters for high-stakes domains**: In insurance, a coverage determination that misinterprets a policy exclusion costs real money. In government, an incorrect regulatory interpretation has legal consequences. In finance, a flawed risk assessment can trigger material losses. These are domains where getting the answer right matters more than getting it fast. The evidence supports verification as a powerful quality lever. Meta AI's Chain-of-Verification (2023) improved factual accuracy by 4-8% across question types. VeriFY (2026) achieved 9.7-53.3% hallucination reduction by teaching models to reason about factual uncertainty through consistency-based self-verification. A multi-modal fact-verification framework (2025) demonstrated 67% hallucination reduction without sacrificing response quality. "Sample, Scrutinize and Scale" (2025) showed that scaling verification at inference time allowed Gemini v1.5 to surpass o1-Preview performance on reasoning benchmarks. The generate-verify-refine loop -- treating the system's own output as a hypothesis to be tested rather than an answer to be delivered -- is emerging as a recognised pattern in the literature. Annie implements this structurally: generation (experts), verification (the panel), and refinement (re-entry on failure) are separate stages with separate models, connected through the Cognition Stream. **The discard-and-restart limitation**: Currently, when verification fails, the entire response is discarded and the prompt starts fresh. This is wasteful -- the failed response contains information about what went wrong that could guide a better second attempt. The improvement roadmap includes feeding verification failure signals back to experts as additional context on re-entry, enabling targeted refinement rather than blind restart. The current approach is conservative by design: it guarantees that a failed response cannot contaminate a second attempt, at the cost of redundant computation. ### Synapse and the Rapport Model **Synapse Context Engine** searches prior prompts, responses, knowledge bases, and connected sources to provide situational context to every expert. When an insurance specialist processes a coverage question, Synapse provides the relevant policy documents, prior interactions about the same account, and applicable regulatory guidance. This context is injected alongside the prompt, giving each expert full situational awareness rather than treating every interaction as isolated. **The Rapport Model** is one of the twelve specialist models, but it serves a unique function: it tracks the relationship state between Annie and each individual user. Rather than a binary toggle between "formal" and "casual," the Rapport Model understands where each relationship sits on a continuum and calibrates Annie's communication style accordingly. A new or unknown user receives clinical, professional responses. As interactions accumulate and rapport builds, Annie's full personality emerges -- warm, direct, bantery communication that reflects a genuine working relationship. A high-rapport user asking a simple question gets a different tone than a first-time user asking the same question, even though the factual content is identical. This is a meaningful differentiator. Research published in Nature (Scientific Reports, 2026) found that personalisation -- remembering preferences and prior interactions -- significantly enhances emotional bonding, satisfaction, and trust. Three experimental studies (n=643) on emotional AI found that empathic conversational agents are perceived as warmer and more competent, positively influencing satisfaction and word-of-mouth. McKinsey's personalisation research documents 5-15% revenue lift on average from personalised interactions. The Rapport Model makes Annie feel like a coworker who knows you, not a stateless API that treats every request as a fresh transaction. This is deliberate alignment with the "AI coworker" paradigm that 76% of executives already use to frame agentic AI (BCG, November 2025). ### Sleep Cycles and Continuous Learning Prompts and responses are never discarded. Every interaction -- the original prompt, the classification decision, each expert's response, the judgment scores, the verification outcome -- persists in the Cognition Stream. During off-peak periods (sleep cycles), this data is used to train and refine the specialist models. This creates a flywheel: more interactions produce more training data, which produces better specialists, which produce better responses, which build more rapport and trust, which drives more interactions. The flywheel spins faster with small models because fine-tuning a 7B model on domain-specific interaction data costs under \$5 per run and takes hours rather than weeks. A frontier model cannot be continuously refined from customer interactions at any reasonable cost. The sleep cycle pattern also means that Annie's specialists improve specifically on the domains and query types that her actual users care about, rather than improving on abstract benchmarks. A deployment serving insurance professionals will develop increasingly sharp insurance expertise. A deployment serving software engineers will sharpen on coding tasks. The same architecture adapts to radically different use cases through the data it processes. ## Architectural Decisions and Tradeoffs | Decision | Alternative Considered | Why Annie Chose This | Honest Tradeoff | | ----------------------------------------------- | ------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Multi-model ensemble** | Single large model (e.g., 70B+ generalist) | Ensembles of small models outperform single large models on domain tasks at lower cost (Kondratyuk et al. 2020, SLM-MUX 2024, Chai Research 2024). Independent training, replacement, and upgrade of individual specialists. | Higher system complexity. Coordination overhead. No single model's coherence across an extended multi-turn conversation -- context must be managed externally through Synapse. | | **Asynchronous pipeline** | Synchronous request-response | Quality verification requires multiple stages that take time. The "coworker" model aligns with how knowledge workers actually operate. The market is moving toward async agents (\$26B Devin valuation, GitHub Copilot agent mode). | Users accustomed to ChatGPT-speed responses may find latency jarring for complex queries. Requires UX that sets expectations and communicates progress. Simple queries must still be fast (solved by fast path). | | **Small specialists (250M-27B)** | One large generalist (70B+) | Specialists match or exceed generalist performance on defined domain tasks at 5-50x lower inference cost. Each fits on a consumer GPU. Training cost per specialist: \$2K-\$500K vs \$500M+ for frontier. | Weaker on novel, cross-domain reasoning that requires broad world knowledge. A 7B insurance specialist will not match Opus 4.8 on open-ended creative writing. Annie must know her boundaries. | | **Score-based consensus pipeline** | Single-pass inference | Multi-model consensus improves accuracy by 4-18% depending on task (Wang et al. 2023). Panels of diverse small models outperform single large judges (Verga et al. 2024). Domain-specific rubrics increase agreement by 22-27% (2026). | 2-5x compute cost per complex query compared to single-pass. Latency increases linearly with pipeline stages. Consensus quality depends on genuine model diversity -- homogeneous panels provide minimal benefit. | | **Bellerophon BStream backbone** | Custom orchestration / direct API calls | Durability, replay, backpressure, and observability out of the box. Every message auditable. Decoupled components enable independent scaling and deployment. | Dependency on Bellerophon platform. Additional infrastructure to operate. Message serialisation overhead (negligible compared to model inference, but non-zero). | | **Score-based judgment with rubrics** | Simple majority voting / single judge | Rubric-based evaluation is more reliable and domain-appropriate than unstructured voting. Dynamically generated rubrics outperform static ones. Scoring is auditable and explainable. | Rubric design requires domain expertise. Poor rubrics produce poor judgments. Rubric maintenance is an ongoing cost as domains evolve. | | **Discard-and-restart on verification failure** | Incremental refinement of failed response | Conservative: prevents contamination of second attempt by failed first attempt. Simpler to reason about correctness. | Wasteful: discards information about what went wrong. Redundant computation on re-entry. Improvement roadmap includes targeted refinement using failure signals. | ## Value Propositions ### Sovereign Deployability Annie runs entirely on customer-controlled infrastructure. Every component -- the classifier, all specialist models, the judgment and verification panels, the Cognition Stream, the Synapse context engine -- deploys on hardware the customer owns and operates. No data leaves the customer's jurisdiction. No external API call is required for any inference. This is not an abstract compliance checkbox. When Fable 5's service was suspended, organisations dependent on its API lost AI capability overnight with no recourse. Annie at L3-L4 sovereignty (see [SOTA Landscape](annie/sota-landscape.mdx)) means the customer can disconnect from the internet entirely and the system continues to function. For government, defence, healthcare, and financial services organisations operating under data residency requirements, this is a hard requirement, not a preference. The \$301.6 billion projected sovereign AI infrastructure market by 2040 consists primarily of hardware and cloud plays. Annie provides the model-layer product that makes sovereign infrastructure useful -- the part that turns GPUs into an AI coworker. ### Cost Efficiency The cost differential between Annie and frontier API access is not marginal; it is structural. | | Frontier API (Opus 4.8) | Frontier API (Gemini 3.1 Pro) | Annie Self-Hosted | | ------------------------- | ----------------------- | ----------------------------- | -------------------------------- | | **Input cost per MTok** | \$5.00 | \$2.00-4.00 | Electricity + amortised hardware | | **Output cost per MTok** | \$25.00 | \$12.00-18.00 | Electricity + amortised hardware | | **At 100M tokens/day** | \$1,500-5,000/day | \$700-2,200/day | \$50-200/day | | **Annual cost at scale** | \$550K-1.8M | \$255K-800K | \$18K-73K | | **Training a new domain** | Not possible | Not possible | \$2K-500K per specialist | | **Hardware investment** | None (API) | None (API) | \$5K-200K (one-time) | At 100M tokens per day, Annie's self-hosted cost is 10-100x lower than frontier API access. The break-even on hardware investment occurs within weeks to months depending on usage volume. And unlike API access, the marginal cost of additional inference approaches zero once hardware is provisioned -- there is no per-token charge. *Cost assumptions: 50/50 input/output token ratio, no prompt caching, standard context lengths (\<32K), no batch discounts for API pricing. Annie self-hosted estimates assume Australian electricity at A\$0.30/kWh and amortised hardware over 3 years. Actual costs will vary by workload profile. A detailed cost model with customer-specific assumptions should be built for each partner conversation.* Annie's consensus pipeline does multiply compute per complex query by 2-5x compared to single-pass inference. But 5x the cost of self-hosted inference on consumer GPUs is still dramatically cheaper than 1x the cost of frontier API access. ### Domain Extensibility Adding a new domain to Annie means training a new specialist model and registering it in the catalogue. The existing classifier is updated (or retrained -- at 250M parameters, this is cheap and fast) to recognise the new domain. No other component changes. The judgment and verification panels work with any domain because they evaluate against rubrics, not against hard-coded domain knowledge. This composability means Annie can serve radically different markets from the same architecture. An insurance deployment and a software engineering deployment share the pipeline, the Cognition Stream, the consensus mechanism, and the verification loop. They differ only in which specialists are loaded and which rubrics are configured. For customers, this means new capabilities are additive, not replacement. A deployment that starts with three specialists can grow to twelve or fifty as needs evolve, without rearchitecting anything. **Annie is a platform, not a model.** This distinction matters strategically. Annie ships with Evari's own sovereign models as proof of concept, but the twelve pipeline roles are open slots. A bank brings their own risk model. A defence contractor plugs in a classified domain specialist. A health service adds a clinical decision model. Each customer's models run through the same classification, consensus, and verification pipeline. The architecture is the product. The models are replaceable components. This means Annie's value grows with every customer who brings their own expertise to the platform, and every new model on the platform is additional compute demand on the infrastructure that hosts it. ### Quality Through Verification Annie's closed-loop verification is not a marketing feature; it is a structural property of the pipeline. Every complex response is evaluated by multiple models against rubrics before it reaches the user. Responses that fail verification are rejected and regenerated. The evidence for multi-model verification is strong: Chain-of-Verification reduces hallucination by 4-8% (Meta AI, 2023). VeriFY achieves 9.7-53.3% hallucination reduction (2026). Multi-agent consistency verification reduces Expected Calibration Error by 49-74% across medical benchmarks (2026). The Six Sigma Agent paper (2026) mathematically proves that consensus voting with n independent agents reduces error to O(p^(ceil(n/2))), enabling exponential reliability gains. For high-stakes domains where errors have financial, legal, or safety consequences, this is the difference between a useful tool and a liability. The 39% of AI-powered customer service bots that were pulled back or reworked in 2024 due to hallucination errors (ComputerTechReviews, 2025) represent exactly the failure mode that verification prevents. ### Hardware Accessibility Annie's specialist models run on consumer-grade GPUs. A quantised 7B model requires approximately 5-8GB of VRAM -- well within the capability of an NVIDIA RTX 4070 or equivalent. Even the largest specialists in the current catalogue (27B parameters) fit on a single RTX 4090 or equivalent with Q4 quantisation (\~20GB VRAM). A full Annie deployment serving a mid-sized organisation can run on hardware costing \$5K-\$50K. Compare this to the datacentre-scale infrastructure required for frontier models: DeepSeek V4-Pro's 1.6 trillion parameters require multi-node GPU clusters even with its efficient sparse activation. Fable 5's training infrastructure costs billions. Annie's operational infrastructure costs what a small business spends on office furniture. ### Continuous Improvement Every interaction makes Annie better at the specific work her users need. Sleep cycle training on accumulated interaction data is cheap (under \$5 per fine-tuning run for a 7B model), fast (hours, not weeks), and targeted (improving on actual user queries, not abstract benchmarks). This flywheel does not exist with frontier API access. OpenAI and Anthropic train their models on their schedule, optimising for their benchmarks, with no mechanism for a customer's domain-specific interactions to improve the model they are using. Annie's specialists get better at insurance work because they process insurance work. They get better at code review because they perform code reviews. The improvement is automatic, continuous, and domain-specific. ### Tool Capability Annie's specialist models are not text generators alone — they are agentic. During expert processing, specialists invoke two categories of tools. **Internal platform tools** (Synapse context queries, knowledge base searches, internal platform APIs) execute immediately as part of inference. **External tools** (querying customer databases, calling business APIs, interacting with ticketing systems like Jira, writing and executing code in external environments, retrieving documents from customer systems) are returned as tool call requests in the expert's response -- identical to how frontier models return tool\_use blocks. These external requests are resolved at judgment time and executed by the client. Tool inputs and outputs are captured in the Cognition Stream and flow through the same judgment and verification pipeline as generated responses, ensuring all external actions are evaluated for correctness before reaching the user. ### Relationship-Aware Interaction The Rapport Model transforms Annie from a stateless query processor into a colleague who understands working relationships. Research consistently shows that personalisation and adaptive tone improve satisfaction, trust, and engagement. Annie's approach is not a surface-level "add the user's name to responses" gimmick -- it is a model that understands relationship depth and calibrates communication style across a continuous spectrum. This matters for the coworker paradigm. Real coworkers develop rapport. They communicate differently with people they have worked with for years versus someone they just met. Annie's Rapport Model replicates this natural dynamic, making the "AI coworker" framing feel genuine rather than aspirational. ## The Specialist Model Landscape Annie's current specialists are five fine-tunes of the sovereign Workforce base model (sparse MoE, trained from scratch on 1.5T curated Evari tokens + 15T open-source tokens) plus Qwen 3.6:27B. These six physical models fill twelve logical pipeline roles, with some models serving multiple roles. Annie is currently domain-specialised for Evari's insurance technology and platform development work -- expanding to new domains requires training new specialists or customers injecting their own models. The table below shows the open-source landscape available for customer deployments and future specialist expansion. Because Annie is a platform with model-agnostic slots, any of these models can fill a role in the pipeline (see [SOTA Landscape](annie/sota-landscape.mdx) for detailed analysis). ```mermaid theme={null} graph TB subgraph "Annie's Specialist Candidate Pool" subgraph "Tier 1: Lightweight (250M-3B)" CL["Classifier
Custom 250M
or Gemma 4 E2B (~2B)"] RP["Rapport Model
Phi-4 Mini (3.8B)
MIT License"] end subgraph "Tier 2: Core Specialists (7B-14B)" CS["Coding Specialist
Qwen 3.6-35B-A3B
(3B active, Apache 2.0)"] GK["General Knowledge
Gemma 4 12B
Apache 2.0"] IS["Insurance Specialist
Phi-4 14B (fine-tuned)
MIT License"] JR["Jira/Workflow
Ministral 14B
Apache 2.0"] end subgraph "Tier 3: Heavy Specialists (14B-27B)" RN["Reasoning
Gemma 4 31B
Apache 2.0"] RC["Regulatory/Compliance
Qwen 3.6-27B (fine-tuned)
Apache 2.0"] end subgraph "Judgment & Verification" J1["Judge 1
Mistral Small 4
(6B active, Apache 2.0)"] J2["Judge 2
Gemma 4 12B
Apache 2.0"] V1["Verifier 1
Phi-4 14B
MIT License"] V2["Verifier 2
Different architecture
(diversity requirement)"] end end ``` Key candidates and their fit: | Role | Model Candidate | Parameters | VRAM (Q4) | License | Key Strength | | ------------------------------ | --------------------- | ---------------------- | --------- | ---------- | ------------------------------------------------------------ | | Classifier / Gating | Custom or Gemma 4 E2B | \~250M-2B | \~1.5GB | Apache 2.0 | Minimal latency, runs on anything | | Coding Specialist | Qwen 3.6-35B-A3B | 35B total / 3B active | \~21GB | Apache 2.0 | Beat Gemma 4 26B by 21 points on coding; 1M native context | | General Reasoning | Gemma 4 31B | 31B dense | \~20GB | Apache 2.0 | 89.2% AIME 2026, 80% LiveCodeBench | | Domain Specialist (fine-tuned) | Phi-4 14B | 14B dense | \~10GB | MIT | 93.7% GSM8K, strong base for fine-tuning | | Judgment Panel | Mistral Small 4 | 119B total / 6B active | \~25GB | Apache 2.0 | European sovereign champion; 128 experts/4 active internally | | Lightweight Tasks | Gemma 4 12B | 12B dense | \~8GB | Apache 2.0 | Strong all-rounder, multimodal variants available | | Reasoning Specialist | Ministral 14B | 14B dense | \~10GB | Apache 2.0 | 85% AIME 2025 | All candidates are available under permissive open-source licenses (Apache 2.0 or MIT). All run on consumer GPUs. The catalogue is extensible: as new models are released (the pace is accelerating), Annie can adopt them by training a new specialist or swapping an existing one -- with no changes to the pipeline. A critical note on the open-weight window: published open-weight models are currently exempt from US export controls (ECCN 4E091), but this could change. Annie's architecture is designed to work with whatever models are available -- if the export landscape shifts, the specialist slots are filled with whatever high-quality models remain accessible. The architecture is model-agnostic by design. ## Annie vs Frontier: An Honest Comparison Annie is not trying to be a general-purpose frontier model. She is trying to match or exceed frontier performance on defined domain tasks, with verification, at a fraction of the cost, on sovereign infrastructure. Here is where she wins, where she loses, and why the wins matter more for her target market. *Note: The following chart is a strategic positioning framework showing where Annie's architecture is best suited, not a benchmark. Formal performance measurements will be produced as part of pilot scoping.* ```mermaid theme={null} quadrantChart title Annie vs Frontier Models: Capability Map x-axis "Low Domain Specificity" --> "High Domain Specificity" y-axis "Low Verification Need" --> "High Verification Need" quadrant-1 "Annie's Sweet Spot" quadrant-2 "Annie Competitive" quadrant-3 "Frontier Wins" quadrant-4 "Annie Competitive" "Insurance underwriting": [0.85, 0.9] "Regulatory compliance": [0.8, 0.85] "Domain code generation": [0.75, 0.7] "Financial analysis": [0.7, 0.85] "Creative writing": [0.2, 0.15] "General chat": [0.15, 0.1] "Open-ended research": [0.3, 0.4] "Novel reasoning": [0.25, 0.5] "Customer support (domain)": [0.65, 0.6] "Document processing": [0.6, 0.55] ``` | Dimension | Annie | Frontier (e.g. Opus 4.8, GPT-5.5) | Verdict | | ----------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Domain task accuracy** | Consensus-verified, rubric-scored output from domain specialists. Evidence shows ensembles of specialists match or exceed large generalists on defined tasks. | Single-pass inference from a broadly capable model. Higher raw capability on novel tasks. | **Annie wins on defined domains.** The consensus pipeline and domain fine-tuning compensate for individual model size. Frontier wins on novel, undefined tasks. | | **General reasoning** | Limited by largest specialist (\~27B). Weaker on cross-domain synthesis requiring broad world knowledge. | State-of-the-art. Opus 4.8 at 88.6% SWE-Bench, GPT-5.5 at 88.7%. Hundreds of billions of active parameters. | **Frontier wins clearly.** Annie does not compete here and should not pretend to. | | **Hallucination rate** | Multi-stage verification reduces hallucination by 4-67% depending on method (CoVe, VeriFY, multi-modal verification). Cross-model consensus catches self-consistent errors that single-model approaches miss. | GPT-5.5 reported at 86% hallucination rate on factual recall tasks. Single-pass inference with no structural verification. | **Annie wins on verified output.** The pipeline exists specifically to catch and reject hallucinated responses. | | **Latency (simple queries)** | Sub-second via fast path (single specialist, no consensus). | Sub-second to seconds depending on provider and model. | **Comparable.** Fast path matches frontier speed. | | **Latency (complex queries)** | Seconds to minutes (expert processing + judgment + verification). | Seconds (single pass, even with chain-of-thought). | **Frontier wins on speed.** Annie trades latency for verified quality. This is the core tradeoff. | | **Cost at scale** | \$50-200/day at 100M tokens. Hardware investment \$5K-200K one-time. | \$700-5,000/day at 100M tokens. No hardware investment but perpetual per-token cost. | **Annie wins by 10-100x.** The cost advantage is structural and widens with volume. | | **Sovereignty** | Full L3-L4 sovereignty. Runs on customer infrastructure. No external dependencies. Annie's base model is trained from scratch with no dependency on external model weights; the fine-tuned domain specialists inherit this sovereign foundation -- the strongest sovereignty position possible. | L1-L2 at best. API dependency. Service can be suspended (demonstrated with Fable 5). Data leaves customer jurisdiction. | **Annie wins absolutely.** This is binary: either you control your AI infrastructure or you do not. Annie's from-scratch base model means no exposure to export controls on external model weights. | | **Domain extensibility** | Add a specialist (\$2K-500K), update classifier, deploy. No other changes. | Request a feature from the provider. Fine-tuning available for some models at provider-controlled cost. | **Annie wins.** Customer controls their own capability roadmap. | | **Continuous improvement** | Sleep cycle training on actual user interactions. Specialists improve on the work they do. | Provider trains on their schedule, their data, their priorities. | **Annie wins for domain specialisation.** Frontier wins for general capability advancement. | | **Explainability** | Every stage logged in the Cognition Stream. Classification, expert responses, judgment scores, verification decisions -- all auditable. | Black box. Some providers offer limited reasoning traces. | **Annie wins.** Full pipeline observability is architecturally guaranteed. | **The honest summary**: Annie loses to frontier models on general reasoning, creative breadth, and latency for complex queries. She wins on domain task accuracy (with verification), cost, sovereignty, extensibility, continuous domain improvement, and explainability. For organisations whose work is primarily in defined domains where accuracy matters more than speed and sovereignty is a requirement -- insurance, government, finance, healthcare, defence -- Annie's wins are the ones that matter. ### Current Scope and Honest Limitations Annie is currently 100% specialised for Evari's domain: insurance technology, platform development, and associated regulatory compliance work. All of Annie's current specialist models were trained on interaction data, domain knowledge, and use cases within this scope. Performance and capability outside this domain are untested and unverified. Annie does not include specialists for government intelligence, defence applications, healthcare, or other vertical markets. Deploying Annie into a new domain requires either: 1. **Training new specialists from scratch** for that domain, using domain-specific data, at a cost of \$2,000-\$500,000 per specialist and a timeline of weeks to months depending on data availability and model size. 2. **Customers providing their own specialist models** that are already trained for their use cases, which Annie can orchestrate through its existing pipeline. The architecture itself is proven and robust. The breadth of domain coverage is not yet proven. This is a pilot activity for a new market and geography (Australia). Partners and customers should expect domain-specific training, model development, and calibration of rubrics as part of any deployment. Initial deployments will improve Annie's capability in their specific domain over time through the sleep cycle training mechanism, but the first deployment into a new sector will require investment in specialist training or customer model contribution. For organisations evaluating Annie: if your use case falls outside insurance, fintech, and related regulated domains, honest scoping requires discussing what new specialists need to be trained or what models you can provide before committing to a deployment timeline. ## Sources ### Ensemble and Multi-Model Evidence 1. Kondratyuk, D., Tan, M., Brown, M., Gong, B. "When Ensembling Smaller Models is More Efficient than Single Large Models." Google AI, 2020. [arXiv:2005.00570](https://arxiv.org/abs/2005.00570). [https://arxiv.org/abs/2005.00570](https://arxiv.org/abs/2005.00570) 2. Chai Research. "Blending Is All You Need: Cheaper, Better Alternative to Trillion-Parameters LLM." 2024. [arXiv:2401.02994](https://arxiv.org/abs/2401.02994). [https://arxiv.org/abs/2401.02994](https://arxiv.org/abs/2401.02994) 3. Jiang, D., Ren, X., Lin, B.Y. "LLM-Blender: Ensembling Large Language Models with Pairwise Ranking and Generative Fusion." ACL 2023. [arXiv:2306.02561](https://arxiv.org/abs/2306.02561). [https://arxiv.org/abs/2306.02561](https://arxiv.org/abs/2306.02561) 4. Wang, C. et al. "SLM-MUX: Orchestrating Small Language Models for Reasoning." 2024. [arXiv:2510.05077](https://arxiv.org/abs/2510.05077). [https://arxiv.org/abs/2510.05077](https://arxiv.org/abs/2510.05077) 5. Li et al. "More Agents Is All You Need." 2024. [arXiv:2402.05120](https://arxiv.org/abs/2402.05120). [https://arxiv.org/abs/2402.05120](https://arxiv.org/abs/2402.05120) ### Cost-Optimised Routing 6. Chen, L. et al. "FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance." Stanford, TMLR 2024. [arXiv:2305.05176](https://arxiv.org/abs/2305.05176). [https://arxiv.org/abs/2305.05176](https://arxiv.org/abs/2305.05176) 7. UC Berkeley, Anyscale, Canva. "RouteLLM: Learning to Route LLMs with Preference Data." ICLR 2025. [arXiv:2406.18665](https://arxiv.org/abs/2406.18665). [https://arxiv.org/abs/2406.18665](https://arxiv.org/abs/2406.18665) ### Small Model Capability 8. Microsoft Research. "Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone." 2024. [https://www.microsoft.com/en-us/research/publication/phi-3-technical-report-a-highly-capable-language-model-locally-on-your-phone/](https://www.microsoft.com/en-us/research/publication/phi-3-technical-report-a-highly-capable-language-model-locally-on-your-phone/) 9. "Fine-Tune an SLM or Prompt an LLM? The Case of Generating Low-Code Workflows." 2025. [arXiv:2505.24189](https://arxiv.org/abs/2505.24189). [https://arxiv.org/html/2505.24189v1](https://arxiv.org/html/2505.24189v1) 10. Apple Machine Learning Research. "Introducing Apple Foundation Models." 2024-2025. [https://machinelearning.apple.com/research/introducing-apple-foundation-models](https://machinelearning.apple.com/research/introducing-apple-foundation-models) ### Mixture of Experts Architecture 11. "A Comprehensive Survey of Mixture-of-Experts: Algorithms, Theory, and Applications." 2025. [arXiv:2503.07137](https://arxiv.org/abs/2503.07137). [https://arxiv.org/html/2503.07137v1](https://arxiv.org/html/2503.07137v1) 12. Quirke, P., Oozeer, N., Bandi, C. et al. "Beyond Monoliths: Expert Orchestration for More Capable, Democratic, and Safe Language Models." Submitted to NeurIPS 2025. [arXiv:2506.00051](https://arxiv.org/abs/2506.00051). [https://arxiv.org/abs/2506.00051](https://arxiv.org/abs/2506.00051) 13. Chai, Z., Wang, G. et al. "An Expert is Worth One Token: Synergizing Multiple Expert LLMs as Generalist via Expert Token Routing." ACL 2024. [arXiv:2403.16854](https://arxiv.org/abs/2403.16854). [https://arxiv.org/abs/2403.16854](https://arxiv.org/abs/2403.16854) ### Consensus and Verification 14. Wang, X. et al. "Self-Consistency Improves Chain of Thought Reasoning in Language Models." ICLR 2023. [arXiv:2203.11171](https://arxiv.org/abs/2203.11171). [https://arxiv.org/abs/2203.11171](https://arxiv.org/abs/2203.11171) 15. Verga et al. "Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models (PoLL)." 2024. Referenced in [https://eugeneyan.com/writing/llm-evaluators/](https://eugeneyan.com/writing/llm-evaluators/) 16. Zhao, J., Plaza-del-Arco, F.M., Cercas Curry, A. "Language Model Council: Democratically Benchmarking Foundation Models on Highly Subjective Tasks." 2024. [arXiv:2406.08598](https://arxiv.org/abs/2406.08598). [https://arxiv.org/abs/2406.08598](https://arxiv.org/abs/2406.08598) 17. Meta AI. "Chain-of-Verification Reduces Hallucination in Large Language Models." 2023. [arXiv:2309.11495](https://arxiv.org/abs/2309.11495). [https://arxiv.org/abs/2309.11495](https://arxiv.org/abs/2309.11495) 18. "Do I Really Know? Learning Factual Self-Verification for Hallucination Reduction (VeriFY)." 2026. [arXiv:2602.02018](https://arxiv.org/abs/2602.02018). [https://arxiv.org/abs/2602.02018](https://arxiv.org/abs/2602.02018) 19. "Multi-Modal Fact-Verification Framework for Reducing Hallucinations in Large Language Models." 2025. [arXiv:2510.22751](https://arxiv.org/abs/2510.22751). [https://arxiv.org/html/2510.22751v1](https://arxiv.org/html/2510.22751v1) 20. "Sample, Scrutinize and Scale: Effective Inference-Time Search by Scaling Verification." 2025. [arXiv:2502.01839](https://arxiv.org/abs/2502.01839). [https://arxiv.org/pdf/2502.01839](https://arxiv.org/pdf/2502.01839) ### Rubric-Based Evaluation 21. "Beyond the Illusion of Consensus: From Surface Heuristics to Knowledge-Grounded Evaluation in LLM-as-a-Judge." 2026. [arXiv:2603.11027](https://arxiv.org/abs/2603.11027). [https://arxiv.org/pdf/2603.11027](https://arxiv.org/pdf/2603.11027) 22. "Rubric Is All You Need: Improving LLM-Based Code Evaluation." ACM ICER 2025. [arXiv:2503.23989](https://arxiv.org/abs/2503.23989). [https://arxiv.org/pdf/2503.23989](https://arxiv.org/pdf/2503.23989) ### Judge Independence and Limitations 23. "Nine Judges, Two Effective Votes: Correlated Errors Undermine LLM Evaluation Panels." 2025. [arXiv:2605.29800](https://arxiv.org/abs/2605.29800). [https://arxiv.org/abs/2605.29800](https://arxiv.org/abs/2605.29800) 24. Tan, H. et al. "Too Consistent to Detect: A Study of Self-Consistent Errors in LLMs." EMNLP 2025. [arXiv:2505.17656](https://arxiv.org/abs/2505.17656). [https://arxiv.org/abs/2505.17656](https://arxiv.org/abs/2505.17656) 25. "The Six Sigma Agent: Achieving Enterprise-Grade Reliability in LLM Systems Through Consensus-Driven Decomposed Execution." 2026. [arXiv:2601.22290](https://arxiv.org/abs/2601.22290). [https://arxiv.org/abs/2601.22290](https://arxiv.org/abs/2601.22290) 26. Maryanskyy, A. "When Agents Disagree: The Selection Bottleneck in Multi-Agent LLM Pipelines." 2025. [arXiv:2603.20324](https://arxiv.org/abs/2603.20324). [https://arxiv.org/abs/2603.20324](https://arxiv.org/abs/2603.20324) ### Market and Industry 27. BCG. "Agentic AI Blurs the Line Between Tool and Teammate." November 2025. [https://www.bcg.com/press/18november2025-agentic-ai-blurs-line-tool-teammate](https://www.bcg.com/press/18november2025-agentic-ai-blurs-line-tool-teammate) 28. Microsoft. "2026 Work Trend Index." 2026. [https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization](https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization) 29. Gartner. "40% of Enterprise Apps Will Feature AI Agents by 2026." 2025. [https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025](https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025) 30. CodeRabbit. "The Rise of Slow AI." 2026. [https://www.coderabbit.ai/blog/the-rise-of-slow-ai-why-devs-should-stop-speedrunning-stupid](https://www.coderabbit.ai/blog/the-rise-of-slow-ai-why-devs-should-stop-speedrunning-stupid) 31. TechTimes. "Cognition's \$26B Raise." 2026. [https://www.techtimes.com/articles/317354/20260529/ai-coding-agents-cognitions-26b-raise-bets-agent-first-architecture-beats-ide-tools.htm](https://www.techtimes.com/articles/317354/20260529/ai-coding-agents-cognitions-26b-raise-bets-agent-first-architecture-beats-ide-tools.htm) ### Personalisation and Rapport 32. Nature/Scientific Reports. "Building User Trust in AI Chatbots." 2026. [https://www.nature.com/articles/s41598-026-38179-2](https://www.nature.com/articles/s41598-026-38179-2) 33. ScienceDirect. "Emotional AI: Impact on Satisfaction." 2026. [https://www.sciencedirect.com/science/article/pii/S107158192600039X](https://www.sciencedirect.com/science/article/pii/S107158192600039X) ### Scaling Laws 34. Hoffmann, J. et al. "Chinchilla Scaling Laws." DeepMind, 2022. 35. Sutskever, I. NeurIPS 2024 keynote remarks on the end of the scaling era. ### Confidence and Calibration 36. Kadavath, S. et al. "Language Models (Mostly) Know What They Know." Anthropic, 2022. [arXiv:2207.05221](https://arxiv.org/abs/2207.05221). [https://arxiv.org/abs/2207.05221](https://arxiv.org/abs/2207.05221) 37. Du, Y., Li, S., Torralba, A., Tenenbaum, J.B., Mordatch, I. "Improving Factuality and Reasoning in Language Models through Multiagent Debate." ICML 2024. [arXiv:2305.14325](https://arxiv.org/abs/2305.14325). [https://arxiv.org/abs/2305.14325](https://arxiv.org/abs/2305.14325) # Competitive anaylsis Source: https://docs.quiva.ai/annie/competitive-anaylsis # Competitive Analysis: Annie vs The Field **Status:** Working Draft **Date:** 2026-06-22 **Depends on:** [SOTA Landscape](annie/sota-landscape.mdx), [Annie Architecture](annie/annie-architecture.mdx) *** ## Executive Summary Annie's competitive position is narrow but defensible. She wins where three conditions converge: the task is domain-specific, the output requires verified correctness, and the deployment must be sovereign. Outside that intersection, she loses to frontier models on raw capability and to raw open-weight models on simplicity. The honest picture: Annie will not match Opus 4.8 or GPT-5.5 on open-ended reasoning. She will not beat a single self-hosted Qwen 3.6-27B on deployment simplicity. She will not compete with ChatGPT on consumer chat. She should not try. What Annie offers is a verified, sovereign, domain-extensible pipeline that no single model -- frontier or open -- provides today. That pipeline costs 10-100x less than frontier APIs at scale and cannot be shut off by a foreign government. The real competitive threat is not frontier labs. It is the "good enough" single open model: a team that deploys Qwen 3.6-27B on a single GPU and decides the orchestration overhead is not worth it. Annie must prove the verification pipeline catches enough errors to justify the added complexity. The research says it does -- ensembles improve accuracy by 4-18% on domain tasks -- but that margin must be visible and measurable to buyers, not just claimed. *** ## Architectural Comparison Matrix Five distinct approaches to deploying language model capability exist in production today. Each makes a different tradeoff between capability, cost, control, and complexity. ### 1. Monolithic Dense (Historical, Still Used at Small Scale) **How it works:** A single dense transformer where every parameter activates for every token. All of GPT-2, GPT-3, and early GPT-4 were dense. Today, dense models persist at the small end: Phi-4 (3.8-15B), Gemma 4 12B/31B, Qwen 3.6-27B. **Strengths:** Simple to deploy. Predictable latency. Single file, single GPU (at small scale). Well-understood training dynamics. No routing failures possible. **Weaknesses:** Capability scales linearly with compute -- every token pays the full parameter cost. A 27B dense model is fundamentally limited by 27B parameters worth of knowledge and reasoning. No path to frontier performance without frontier compute. **Cost profile:** Training \$2K-\$500K depending on size. Inference is the lowest per-token cost at a given parameter count. A 27B model runs on a single consumer GPU (approximately 20GB VRAM at Q4). **Sovereign suitability:** Excellent. A laptop to a single GPU is all that is required. This is the simplest sovereign deployment possible. **Best use case:** Single-domain tasks where a fine-tuned specialist is sufficient and orchestration overhead is not justified. ### 2. Sparse MoE Single Model (Fable 5, GPT-5.x, Gemini, DeepSeek) **How it works:** A transformer with many expert sub-networks, of which only a fraction activate per token. A learned router selects which experts process each token. Gemini 2.5 Pro: 200B total parameters, 64 experts per block, 8 active per token (12.5% active). GPT-5.5: estimated 10-50+ trillion total, 2-5 trillion active. DeepSeek V4-Pro: 1.6T total, 49B active. **Strengths:** Near-dense quality at a fraction of compute. Frontier capability -- these are the models setting benchmarks. Fable 5 achieves 95% on SWE-Bench Verified (vendor-scaffold, contested). GPT-5.5 achieves 88.7%. The architecture scales to enormous total parameter counts while keeping inference cost manageable. **Weaknesses:** Requires datacenter-scale hardware at frontier sizes. Expert collapse is a real training pathology: the router learns to send most tokens to a few favored experts while others atrophy. A 2026 paper found that expert "specialization" reflects hidden-state geometry, not domain expertise -- what looks like intelligent routing is an emergent geometric property. Load-balancing auxiliary losses help but directly trade off against model performance. Token dropping under load degrades quality. **Cost profile:** Training: \$5M-\$500M+. Inference: \$2-\$50/MTok via API. Sovereign deployment of frontier-scale MoE models is impractical -- GPT-5.5's estimated parameter count requires hardware that only hyperscalers operate. **Sovereign suitability:** Poor at frontier scale. Moderate for smaller MoE models (DeepSeek V4-Flash at 284B total / 13B active could run on a modest GPU cluster). The irony: sparse MoE was designed for efficiency, but frontier providers have used the efficiency gains to scale up rather than scale down. **Best use case:** General-purpose capability where the broadest possible knowledge and reasoning are needed and API dependency is acceptable. ### 3. Reasoning Chains (o-series, DeepThink) **How it works:** An LLM generates an explicit chain-of-thought at inference time, spending more compute on harder problems. OpenAI's o3 and o4-mini, Google's Gemini 3.x with DeepThink (three-tier: Low/Medium/High). This is inference-time compute scaling -- the model "thinks longer" on harder problems. **Strengths:** State-of-the-art on hard problems. o3-pro achieves 98% on AIME 2025. Controllable compute: easy questions use less; hard questions use more. Can be applied on top of any base model architecture. **Weaknesses:** Overthinking on simple tasks: shorter reasoning chains are up to 34.5% more accurate than longer ones on the same easy questions. Chain-of-thought actively harms performance on implicit statistical learning (GPT-4o drops 23.1%, o1-preview drops 36.3% vs zero-shot). Cost explosion: average token usage nearly doubled across DeepSeek R1 versions (12K to 23K tokens per question). Reasoning loops where the model fails to recognise it has reached a correct answer. Difficulty miscalibration: disproportionate compute on simple problems, insufficient on complex ones. **Cost profile:** Highly variable. At OpenAI's o3 pricing (\$2/\$8 per MTok input/output), a complex reasoning chain generating 20K tokens costs approximately \$0.16 per query. o3-pro at \$20/\$80 per MTok makes extended reasoning chains expensive. The cost advantage over dense models on easy problems is offset by massive cost spikes on hard problems. **Sovereign suitability:** Same as the base model. If the base model is open-weight (DeepSeek R1 at 671B/37B active), reasoning chains can run on sovereign infrastructure. If the base model is API-only (o3-pro), it remains API-dependent. **Best use case:** Mathematical reasoning, formal logic, and complex multi-step problems where correctness matters more than latency or cost. Not suitable as a default mode for all queries. ### 4. Self-Hosted Open Models (Llama, Mistral, Qwen on Own Hardware) **How it works:** Download open-weight models (Apache 2.0, MIT, or similar license), quantise to fit available hardware, serve via vLLM, Ollama, or similar inference server. No orchestration layer -- single model, single endpoint. **Strengths:** Full sovereignty at Level 3-4 (see [SOTA Landscape](annie/sota-landscape.mdx)). No per-token API cost. The model quality of open weights in mid-2026 is genuinely impressive: Qwen 3.6-27B offers 1M native context and dense vision at approximately 20GB VRAM. Gemma 4 31B achieves 89.2% on AIME 2026. DeepSeek V4-Pro achieves 80.6% on SWE-Bench Verified. These are not toy models. **Weaknesses:** Single-model limitations: no verification, no consensus, no structural hallucination reduction. The "last mile" problem: going from a running model to a production system requires prompt engineering, guardrails, monitoring, error handling, domain tuning, and UX -- all of which must be built in-house. Small model hallucination rates are dramatically higher: global average across 1.7-3B models is 80.9%. Even at 8-32B, hallucination rates average 54.75% vs 11.91% for larger models. No structural protection against confidently wrong output. **Cost profile:** Hardware: \$5K-\$200K one-time. Electricity: \$105K-\$210K/year per rack at Australian rates. No per-token cost. Amortised daily cost at 100M tokens/day: \$50-\$200. **Sovereign suitability:** Excellent. This is the baseline sovereign AI deployment. Hardware in your rack, weights on your disk, no external dependencies. **Best use case:** Organisations that need sovereignty and have the engineering capability to build the "last mile" themselves, or where the task is simple enough that single-pass inference is acceptable. ### 5. Multi-Model Orchestration (Annie) **How it works:** Multiple specialist models (250M-27B) orchestrated through a message-based pipeline. Gateway receives work, Classifier routes to 1-6 domain specialists running in parallel, Judgment Panel scores responses using rubrics and consensus, Verification Panel checks output quality, Outbound Processing delivers results. Specialist models can make tool calls during processing (querying databases, calling APIs, writing and executing code), with outputs verified through the consensus pipeline. Fast path bypasses consensus for simple queries. Entire pipeline runs on Bellerophon BStream for durability and observability. **Strengths:** Verified output: multi-model consensus improves accuracy by 4-18% on domain tasks. Full pipeline observability -- every classification, expert response, judgment score, and verification decision is logged in the Cognition Stream. Sovereign at Level 3-4. Tool-capable experts: specialists invoke internal platform tools (Synapse, knowledge bases) during reasoning which execute immediately, and return external tool calls (querying databases, calling APIs, code execution) as requests for client execution at judgment time, with all inputs and outputs verified through the consensus pipeline. Each specialist can be independently trained, replaced, or upgraded. Domain extensibility: add a specialist for \$2K-\$500K, update the classifier, deploy. Cost at scale: \$50-\$200/day at 100M tokens vs \$700-\$5,000/day for frontier APIs. **Weaknesses:** Correlated errors: model pairs agree on wrong answers approximately 60% of the time vs 33% expected by chance, and more capable models exhibit higher error correlation. Classifier misrouting: leading routers achieve only 68-70% accuracy, and on queries where fewer than 3 models can answer correctly, accuracy drops to 23-25%. Latency: sequential pipeline stages add 800-2,000ms vs sub-500ms for single models. Coordination overhead: production measurements show 38.6% overhead from consensus mechanisms. System complexity: a 2025 study of 7 multi-agent systems found 41-86.7% failure rates across 14 distinct failure modes. The "good enough" problem: many tasks do not need verification, and the overhead is wasted. **Cost profile:** Hardware: \$5K-\$200K one-time (same as self-hosted single model -- Annie's specialists fit on the same hardware). Per-specialist training: \$2K-\$500K. Higher per-query compute than single model (2-5x for complex queries due to consensus pipeline). But dramatically lower than frontier APIs at any meaningful volume. **Sovereign suitability:** Excellent. Same hardware profile as self-hosted open models, with the orchestration layer (Bellerophon BStream) also self-hosted. **Best use case:** High-stakes domain tasks where verified correctness matters, sovereignty is required, and the organisation will invest in domain specialists. Insurance, regulatory compliance, financial analysis, medical review. ### Architectural Comparison Diagram ```mermaid theme={null} graph TB subgraph "Architectural Approaches: Capability vs Control" direction TB subgraph frontier["Frontier API (MoE / Reasoning)"] F1["Fable 5 / GPT-5.5 / Gemini 3.x"] F2["Capability: ★★★★★"] F3["Sovereignty: ★☆☆☆☆"] F4["Cost at Scale: ★☆☆☆☆"] F5["Verification: ★★☆☆☆"] F6["Complexity: ★☆☆☆☆"] end subgraph reasoning["Reasoning Chains"] R1["o3-pro / DeepThink High"] R2["Capability: ★★★★★"] R3["Sovereignty: ★☆☆☆☆ to ★★★★☆"] R4["Cost at Scale: ★★☆☆☆"] R5["Verification: ★★★☆☆"] R6["Complexity: ★★☆☆☆"] end subgraph selfhost["Self-Hosted Open Model"] S1["Qwen 3.6-27B / Gemma 4 31B"] S2["Capability: ★★★☆☆"] S3["Sovereignty: ★★★★★"] S4["Cost at Scale: ★★★★★"] S5["Verification: ★☆☆☆☆"] S6["Complexity: ★★☆☆☆"] end subgraph annie["Multi-Model Orchestration (Annie)"] A1["12 Specialists via BStream"] A2["Capability: ★★★★☆ domain / ★★☆☆☆ general"] A3["Sovereignty: ★★★★★"] A4["Cost at Scale: ★★★★☆"] A5["Verification: ★★★★★"] A6["Complexity: ★★★★☆"] end subgraph dense["Monolithic Dense (Small)"] D1["Phi-4 / Gemma 4 12B"] D2["Capability: ★★☆☆☆"] D3["Sovereignty: ★★★★★"] D4["Cost at Scale: ★★★★★"] D5["Verification: ★☆☆☆☆"] D6["Complexity: ★☆☆☆☆"] end end style frontier fill:#ff6b6b,color:#000 style reasoning fill:#ffa07a,color:#000 style selfhost fill:#87ceeb,color:#000 style annie fill:#90ee90,color:#000 style dense fill:#dda0dd,color:#000 ``` *** ## Head-to-Head Comparisons ### Annie vs Anthropic (Opus 4.8 / Fable 5) **The Contender:** Anthropic's model family spans Haiku 4.5 (\$1/\$5 per MTok) through Fable 5/Mythos 5 (\$10/\$50 per MTok). Fable 5 is widely believed to be sparse MoE, though Anthropic has not confirmed architecture details. Opus 4.8 achieves 88.6% on SWE-Bench Verified and 93.6% on GPQA Diamond. Fable 5 pushes to 95% SWE-Bench (vendor-scaffold, contested) and 128K max output tokens. Anthropic's projected 2026 losses are approximately \$29B against \$25-30B revenue, with committed compute partnerships exceeding \$330B. **Where Annie wins:** *Sovereignty.* Anthropic models are API-only. There are no self-hosted options. This is Level 1 sovereignty -- full dependency on a US provider. Service can be suspended, access revoked, or pricing changed unilaterally. For organisations in jurisdictions subject to export controls or geopolitical risk, this is not a technical limitation but an existential one. Annie's models are trained from scratch -- no dependency on external model weights that could become subject to export controls. This represents the strongest possible sovereignty position: full technical and legal independence from any external provider. *Cost at scale.* At 100M tokens/day, Opus 4.8 costs \$1,500-\$5,000/day (\$550K-\$1.8M/year). Fable 5 at \$10/\$50 per MTok would cost \$3,000-\$10,000/day. Annie self-hosted costs \$50-\$200/day (\$18K-\$73K/year) plus one-time hardware investment. The gap widens with volume and never closes. *Verification.* Anthropic offers no structural verification layer. Output is single-pass, with no consensus mechanism, no rubric-scored judgment, and no verification stage. Extended thinking provides reasoning traces but not independent verification. Annie's multi-stage pipeline catches errors that a single model -- even a frontier one -- will not catch by design. *Extensibility.* Adding a domain to Annie means training a specialist (\$2K-\$500K), updating the classifier, and deploying. Adding a domain to Anthropic's models means requesting a feature, waiting, and hoping. Fine-tuning is available but at provider-controlled pricing and with provider-controlled constraints. *Observability.* Every stage of Annie's pipeline is logged in the Cognition Stream on Bellerophon BStream. Classification decisions, expert responses, judgment scores, verification outcomes -- all auditable. Anthropic provides usage logs and, for some models, reasoning traces. There is no pipeline-level observability because there is no pipeline. **Where Anthropic wins:** *Raw capability.* Opus 4.8 at 88.6% SWE-Bench Verified and 93.6% GPQA Diamond represents hundreds of billions of active parameters trained on data that no specialist can match. Annie's largest specialist is 27B dense. For novel, cross-domain reasoning, open-ended creative work, or tasks that require broad world knowledge, Anthropic is categorically superior. Annie does not compete here and should not try. *General reasoning.* A 27B specialist will not match Opus 4.8 on a question it has never been trained for. Annie's consensus pipeline improves accuracy on defined domain tasks; it does not create knowledge that does not exist in the underlying specialists. *Ecosystem.* Claude Code, the Messages API, tool use, MCP integration, 1M context windows, batch processing -- Anthropic has a mature developer ecosystem. Annie is infrastructure, not an ecosystem. *Simplicity.* One API call vs a multi-stage pipeline. For many use cases, the simpler path wins. **The export control factor:** This is Annie's structural advantage that no amount of Anthropic engineering can address. Export controls are a political reality, not a technical one. If a jurisdiction faces US technology restrictions, Anthropic's models become unavailable regardless of their quality. Annie, built on open-weight models with Apache 2.0 and MIT licenses, deployed on customer hardware, is immune to this risk. This is not hypothetical -- service suspensions have been demonstrated. ### Annie vs OpenAI (GPT-5.5 / Codex) **The Contender:** OpenAI's GPT-5.5 (codenamed "Spud") is the first fully retrained base model since GPT-4.5 -- a ground-up rebuild estimated at 10-50+ trillion total parameters with 2-5 trillion active. It achieves 88.7% on SWE-Bench Verified, 92.4% on MMLU, 93.6% on GPQA Diamond, and 85% on ARC-AGI-2. The o-series reasoning models push further: o3-pro at 98% on AIME 2025. OpenAI's weekly users exceed 900M with annualised revenue around \$25B and projected 2026 losses of approximately \$14B. Training runs cost \$500M+ each. **Where Annie wins:** *The hallucination story.* GPT-5.5 achieved only 57% accuracy on the AA-Omniscience factual recall benchmark -- meaning 43% of its answers on factual questions were incorrect or hallucinated. This is not a cherry-picked number -- it is from an independent evaluation benchmark. Annie's verification pipeline exists specifically to catch hallucinated output through cross-model consensus and rubric-scored judgment. The research supports the approach: multi-stage verification reduces hallucination by 4-67% depending on method. A system that is wrong on 43% of factual recall and has no structural verification layer is fundamentally unreliable for high-stakes domain work, regardless of how impressive its reasoning benchmarks look. *Sovereignty.* Same analysis as Anthropic. GPT-5.5 is API-only. OpenAI has released open-weight models (gpt-oss-120b, gpt-oss-20b) but these are not GPT-5.5 -- they are smaller, less capable models. The frontier capability remains locked behind the API. *Cost at scale.* GPT-5.5 at \$5/\$30 per MTok input/output. At 100M tokens/day: \$1,750-\$6,000/day. GPT-5.5 Pro at \$30/\$180 per MTok is orders of magnitude more expensive. Annie: \$50-\$200/day. *The Azure sovereign cloud distinction.* Microsoft offers "Azure sovereign cloud" regions -- but this is Level 2 sovereignty (workloads in-jurisdiction, foreign entity operates). The customer does not hold the keys. The customer cannot disconnect without permission. Microsoft can revoke access. This is a marketing distinction, not a sovereignty one. Annie provides Level 3-4 sovereignty: the customer holds keys, makes decisions, and can disconnect without asking anyone. **Where OpenAI wins:** *Scale of capability.* GPT-5.5 with an estimated 2-5 trillion active parameters operates in a fundamentally different capability regime than Annie's 250M-27B specialists. The breadth of knowledge, the cross-domain transfer, the ability to handle novel queries that fall outside any specialist's training -- OpenAI wins decisively. *The reasoning stack.* The o-series (o3, o3-pro, o4-mini) represents a different paradigm: inference-time compute scaling. o3-pro at 98% AIME 2025 is solving problems that no 27B model can approach. Annie's consensus pipeline improves reliability; it does not create this level of raw problem-solving capability. *Ecosystem and reach.* 900M weekly users, ChatGPT, the Codex agent platform, Operator, deep Microsoft/Azure integration. OpenAI's ecosystem dwarfs anything Annie will build. *Codex comparison.* OpenAI Codex is a cloud-based coding agent that assigns tasks, works autonomously, and submits pull requests -- similar in concept to Annie's asynchronous coworker model. Codex has the advantage of GPT-5.5/o3 underlying capability, massive training data, and deep GitHub integration. Annie's advantage is sovereignty and verified output, but for pure coding capability, Codex with frontier models will outperform Annie's coding specialist (Qwen 3.6-35B-A3B) on complex, novel tasks. ### Annie vs Google (Gemini 3.x) **The Contender:** Google has the most credible sovereign deployment story of any frontier provider, and the only confirmed MoE architecture details in the field. Gemini 2.5 Pro: 200B total parameters, 80 layers, 16384 hidden dimension, 128 attention heads, 64 experts per block with 8 active per token (12.5% active, approximately 1.6x compute efficiency vs dense equivalent). Gemini 3.1 Pro achieves 80.6% SWE-Bench Verified, 94.3% GPQA Diamond, and 92.6% MMMLU. Google's 2026 capex guidance is \$175-185B, majority AI-directed. **Where Annie wins:** *True sovereignty vs GDC.* Google Distributed Cloud (GDC) is the best sovereign offering from a frontier provider. It puts Google hardware and software in customer-controlled or air-gapped facilities. But it is still Google's hardware running Google's software under Google's licensing. The customer cannot modify the models, cannot train their own specialists, and depends on Google for updates, patches, and continued operation. Annie at Level 3-4 sovereignty means the customer owns everything: hardware, model weights, training pipeline, orchestration layer. The distinction matters most when the relationship with the provider becomes adversarial -- whether through sanctions, contract disputes, or strategic divergence. *Domain extensibility.* Google offers Gemini fine-tuning, but at Google-controlled pricing and with Google-controlled constraints. Annie's add-a-specialist architecture means the customer controls the capability roadmap entirely. *Verification pipeline.* Same structural advantage as against all frontier providers: Google offers no multi-model consensus or verification layer. Gemini 3.x with DeepThink adds reasoning chains (three tiers: Low/Medium/High), which provides a form of self-verification but not independent cross-model verification. **Where Google wins:** *The TPU advantage.* Google operates custom silicon (TPU v5e, v6e Trillium, TPU 8t with 121 exaflops per superpod at 9,600 chips) that no one else can buy. This is a structural cost advantage in training and inference that translates to aggressive API pricing: Gemini 2.5 Flash at \$0.30/\$2.50 per MTok is cheaper than running Annie's specialists on commodity hardware at low volumes. *Confirmed MoE architecture.* Google is the only frontier provider that has published detailed architecture specifications. The 64-experts-per-block, 8-active-per-token design at 200B parameters is a known, well-characterised architecture. This matters for research and trust: customers know what they are buying. *GDC is genuinely the best provider sovereign story.* While it falls short of true sovereignty (Level 2-3 vs Annie's Level 3-4), for organisations that need frontier capability and some sovereign deployment, GDC is the least-bad option among frontier providers. It is meaningfully better than "API calls to us-east-1." *Context and multimodality.* Gemini 3.1 Pro offers 1M token context. The Gemma 4 family provides on-device multimodal capability. Google's native multimodal training (not bolted-on vision) is a genuine technical advantage that Annie's pipeline of text-focused specialists cannot match without significant investment. *Cost at lower volumes.* Gemini 3.1 Flash Lite at \$0.25/\$1.50 per MTok is extremely competitive. At 1M tokens/day, the API costs \$0.88/day. Annie's amortised hardware cost at the same volume is \$10-\$50/day. Google wins on cost until volume exceeds approximately 10M tokens/day. ### Annie vs Self-Hosted Open Models (Llama / Mistral / Qwen) **This is Annie's real competitive threat.** Not frontier APIs -- those serve a different market. The threat is an engineering team that downloads Qwen 3.6-27B, puts it on a single GPU, wraps it in a basic API, and decides they are done. **The case for "just deploy Qwen":** The quality of open models in mid-2026 is remarkable. Qwen 3.6-27B: 27B dense parameters, 1M native context, vision capability, Apache 2.0 license, approximately 20GB VRAM at Q4 quantisation. It runs on a single consumer GPU. Gemma 4 31B: 89.2% AIME 2026, 80% LiveCodeBench. DeepSeek V4-Pro: 80.6% SWE-Bench Verified, 1.6T total / 49B active, MIT license. These models are available today, free, with no orchestration complexity. Deployment is straightforward: download weights, quantise, serve via vLLM or Ollama, wrap in an API. One model, one endpoint, one thing to monitor. No classifier to train, no judgment panel to calibrate, no verification loops to debug. Full sovereignty. Low latency (no multi-stage pipeline). A competent team can have this running in production in a week. **Annie's cost advantage in context.** While a single open model is simpler, Annie's total initial investment -- a modest capital outlay covering base model pre-training and fine-tuning -- is a one-time cost that pays dividends through the life of the system. Annie's architecture consists of a single sovereign base model fine-tuned into five unique domain specialist variants, plus Qwen 3.6:27B, with post-training on Evari's internal servers. This produced a system that cannot be built by downloading a single open model. For organisations that need both sovereignty and verified correctness, the investment is recovered within months through improved accuracy and reduced hallucination on domain tasks. **Why Annie's orchestration layer matters:** *The hallucination problem is real and structural.* Small model hallucination rates are dramatically higher than large models. At 8-32B parameters, hallucination averages 54.75%. A single Qwen 3.6-27B will confidently produce wrong output with no mechanism to detect or correct it. For low-stakes applications (internal search, draft generation, chat), this is acceptable. For insurance coverage determination, regulatory compliance, or financial analysis, it is not. *Verification is the product.* Annie's value proposition is not "we have models" -- anyone can have models. The value is "we verify output through independent cross-model consensus with domain-specific rubrics." The research supports this: multi-model consensus improves accuracy by 4-18% on domain tasks. Fine-tuned Phi-3-mini beat GPT-4o on 6 of 7 financial NLP benchmarks. Ensembles of open-source LLMs scored 65.1% on AlpacaEval 2.0 vs GPT-4o's 57.5%. The verification layer is what transforms a collection of fallible small models into a system that catches its own mistakes. *Domain specialisation compounds.* A single generalist model has one set of weights trying to serve every domain. Annie's architecture allows each specialist to be fine-tuned independently for its domain. A fine-tuned Phi-4 14B for insurance can achieve 96% accuracy on domain tasks where a generalist achieves 80%. The specialist does not need to be good at everything -- it needs to be excellent at one thing. Then the judgment panel, using a different model architecture, independently evaluates whether the specialist's output is correct. *The "last mile" problem.* Going from a running model to a production system requires prompt engineering, guardrails, monitoring, error handling, domain tuning, output formatting, and UX. Most teams underestimate this. Annie provides the last mile as product: the classifier routes intelligently, the judgment panel scores with rubrics, the verification panel catches failures, and the Cognition Stream provides full observability. Building all of this around a single open model is possible but expensive in engineering time. **Honest assessment:** For simple, low-stakes tasks, a single self-hosted open model is the right answer. Annie's orchestration overhead is not justified when the output does not need verification. The "good enough" single model is Annie's most dangerous competitor because it is real, available today, and simple. Annie must compete on verifiable quality improvement, not on theoretical architectural elegance. ### Annie vs Agentic AI Platforms (Devin, Copilot, etc.) **Different category, but investors will ask.** Devin, GitHub Copilot, Cursor, and Google Antigravity are agent platforms -- they use LLMs to perform autonomous multi-step tasks with tool use. Annie is a verified inference pipeline. These solve different problems, but the market positioning overlaps. **How agentic platforms work:** *Devin (Cognition, valued at \$26B):* Multi-agent compound system with a Planner (high-reasoning model), Coder (code-specialised model), Critic (adversarial review), and Agent Router. Can spawn parallel sub-sessions. Devin Desktop (June 2026) opens orchestration to users. Excels at well-defined repetitive tasks (migrations, upgrades, tech debt). Struggles with novel, ambiguous problems. *Cursor:* Auto-routes queries across model tiers (GPT-5.4-mini for trivial, GPT-5.5/Opus for significant). In-house Composer model for low-latency agentic coding. Supports 8 parallel agents. Key insight from their architecture: "The features that actually determine whether Cursor works are not the headline models -- they are the indexing pipeline, the rules system, and MCP tool integration." *GitHub Copilot:* Multi-model (Claude Opus, GPT-5, Gemini 3 Pro). Copilot Coding Agent assigns GitHub issues directly to AI. Ecosystem play across GitHub, VS Code, Azure. *Google Antigravity:* Built agent-first with multi-agent orchestration as the defining feature. Multiple specialised AI agents working in parallel. **Where Annie differs:** Annie and agentic platforms share the multi-model orchestration pattern. But the purpose is different. Agentic platforms use multiple models to orchestrate and execute multi-step actions in the world (write code, file PRs, navigate UIs). Annie uses multiple models to verify the correctness of output. Agentic platforms return external tool calls for the client to execute autonomously in sequence. Annie's specialists return tool calls (internal platform tools execute during processing; external tool requests return for client execution at judgment time) with all outputs verified through consensus before reaching the user. Agentic platforms optimise for autonomous task completion. Annie optimises for verified output reliability. The philosophies diverge on trust and autonomy. Agentic platforms assume the model is probably right and focus on giving it more tools and multi-step autonomy. Annie assumes the model might be wrong and focuses on catching errors through verification before output reaches the user. **Reality check on agentic AI:** Gartner predicts 40%+ of agentic AI projects will be cancelled by end of 2027 due to escalating costs and unclear ROI. Scale AI's Remote Labor Index found the best AI agent (Manus) completed only 2.5% of real Upwork freelance tasks. Gartner coined "agent washing" in 2025, estimating only approximately 130 of thousands of self-proclaimed AI agent vendors are genuinely agentic. **Where agentic platforms win over Annie:** Autonomous task execution, IDE integration, developer workflow. Annie and agentic platforms both use tool calls, but differently: agentic platforms use tools for autonomous multi-step task execution and workflow automation (writing code, filing PRs, navigating UIs), while Annie's specialist models make tool calls during expert processing (querying databases, calling APIs, retrieving documents) within a verified consensus pipeline. Annie optimises for verified correctness; agentic platforms optimise for autonomous task completion. These are complementary, not competitive, capabilities. Annie could serve as the verified inference backend for an agentic platform. **The QuivaWorks Accelerator** QuivaWorks was developed by Evari to provide and agentic AI platform. This provides agent interaction and building UX, and provides a platform for Annie to offload to LLM APIs for performing tasks that those models excel in. -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ## Scenario Analysis How each approach performs across specific use cases, rated on a 5-point scale: Poor (1), Below Average (2), Average (3), Good (4), Excellent (5). ```mermaid theme={null} block-beta columns 7 block:header h0["Scenario"] h1["Frontier API"] h2["Reasoning Chains"] h3["Self-Hosted Open"] h4["Annie"] h5["Agentic Platform"] end block:insurance s1["Insurance Coverage Determination"] r1_1["3 - Capable but no verification, hallucination risk"] r1_2["4 - Good reasoning but costly, no domain tuning"] r1_3["2 - Single model, no verification, high hallucination"] r1_4["5 - Domain specialist + verification + sovereignty"] r1_5["2 - Wrong tool for the job"] end block:code s2["Code Review"] r2_1["5 - Frontier models excel here"] r2_2["4 - Good but overthinks simple issues"] r2_3["3 - Decent with Qwen/DeepSeek"] r2_4["3 - Coding specialist competitive but not frontier"] r2_5["5 - Built for this exact task"] end block:chat s3["General Q&A / Chat"] r3_1["5 - Designed for this"] r3_2["2 - Overkill, expensive"] r3_3["4 - Good with modern models"] r3_4["3 - Fast path works but overhead for simple tasks"] r3_5["2 - Wrong tool"] end block:compliance s4["Regulatory Compliance Check"] r4_1["3 - Capable but unverified, not sovereign"] r4_2["4 - Thorough analysis"] r4_3["2 - No verification layer"] r4_4["5 - Verification + sovereignty + domain specialist"] r4_5["2 - Wrong tool"] end block:creative s5["Creative Content"] r5_1["5 - Broad knowledge, style range"] r5_2["2 - Overthinks creative work"] r5_3["3 - Competent but limited style range"] r5_4["2 - Consensus dampens creative voice"] r5_5["1 - Not designed for this"] end ``` **Detailed scenario breakdown:** | Scenario | Best Approach | Runner-Up | Annie Rating | Notes | | -------------------------------- | ------------------------------- | ---------------- | ------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Insurance coverage determination | **Annie** | Reasoning Chains | 5/5 | Domain specialist + verification + sovereignty. Annie's sweet spot. | | Code review | Frontier API / Agentic Platform | Reasoning Chains | 3/5 | Annie's coding specialist (Qwen 3.6-35B-A3B) is competitive but does not match Opus 4.8 or GPT-5.5 on complex novel code. Agentic platforms have IDE integration Annie lacks. | | General Q\&A / Chat | Frontier API | Self-Hosted Open | 3/5 | Annie's fast path handles simple queries well, but the full pipeline adds unnecessary latency for casual conversation. Not Annie's market. | | Regulatory compliance check | **Annie** | Reasoning Chains | 5/5 | Verification-critical, domain-specific, sovereignty-sensitive. Annie's architecture is purpose-built for this. | | Creative content generation | Frontier API | Self-Hosted Open | 2/5 | Consensus-driven evaluation dampens creative voice. Multiple experts producing "average of opinions" output. Creative work benefits from a single strong voice, not committee review. | Annie excels in exactly two of five scenarios -- but those two (insurance, regulatory compliance) represent the highest-value, highest-risk work where buyers will pay for verified correctness. Annie should not pursue the other three as primary markets. *** ## Cost Modeling All figures in USD. Assumes 50/50 input/output token ratio unless noted. Annie costs assume hardware is amortised over 3 years. ### Low Volume: 1M tokens/day | Solution | Daily Cost | Annual Cost | Notes | | ------------------------ | ---------- | ------------ | ------------------------------------------------------------------------------ | | Annie self-hosted | \$10-\$50 | \$3.7K-\$18K | Dominated by amortised hardware. Overpaying for infrastructure at this volume. | | Opus 4.8 API | \$15 | \$5.5K | Simple, no infrastructure. | | GPT-5.5 API | \$17.50 | \$6.4K | Comparable to Opus. | | Gemini 3.1 Pro API | \$7-\$11 | \$2.6K-\$4K | Cheapest frontier option. | | Self-hosted Qwen 3.6-27B | \$10-\$50 | \$3.7K-\$18K | Same hardware cost as Annie, simpler system. | **Verdict at 1M tokens/day: API wins.** Annie's infrastructure cost is not justified at this volume. Gemini 3.1 Pro or GPT-5.4 API is the right choice unless sovereignty is a hard requirement. If sovereignty is required, a single self-hosted Qwen is simpler and cheaper than Annie's full pipeline. ### Medium Volume: 100M tokens/day | Solution | Daily Cost | Annual Cost | Notes | | ------------------------ | --------------- | ------------- | ---------------------------------------------------- | | Annie self-hosted | \$50-\$200 | \$18K-\$73K | Hardware amortised. Electricity is the primary cost. | | Opus 4.8 API | \$1,500-\$5,000 | \$550K-\$1.8M | 10-100x more expensive than Annie. | | GPT-5.5 API | \$1,750-\$6,000 | \$640K-\$2.2M | Similar to Opus. | | Gemini 3.1 Pro API | \$700-\$2,200 | \$255K-\$800K | Cheapest frontier, still 10x Annie. | | Self-hosted Qwen 3.6-27B | \$50-\$200 | \$18K-\$73K | Same cost as Annie but no verification. | **Verdict at 100M tokens/day: Self-hosted wins decisively.** The cost gap is 10-100x vs frontier APIs. The question is Annie vs raw Qwen -- same infrastructure cost, but Annie provides the verification pipeline. At this volume, the verification overhead (2-5x compute per complex query) is absorbed within the same hardware budget because the hardware is already paid for and has capacity. ### High Volume: 1B tokens/day | Solution | Daily Cost | Annual Cost | Notes | | ------------------------ | ----------------- | -------------- | ---------------------------------------------- | | Annie self-hosted | \$200-\$500 | \$73K-\$182K | Needs GPU scaling. Multiple inference servers. | | Opus 4.8 API | \$15,000-\$50,000 | \$5.5M-\$18.3M | Prohibitive for sustained use. | | GPT-5.5 API | \$17,500-\$60,000 | \$6.4M-\$21.9M | Batching reduces cost 50% but adds latency. | | Gemini 3.1 Pro API | \$7,000-\$22,000 | \$2.6M-\$8M | Best frontier price but still 35-100x Annie. | | Self-hosted Qwen 3.6-27B | \$200-\$500 | \$73K-\$182K | Needs same GPU scaling as Annie. | **Verdict at 1B tokens/day: Self-hosted is the only rational choice.** API costs are \$2.6M-\$21.9M/year vs \$73K-\$182K for self-hosted. The cost gap is so large that the entire hardware investment (including Annie's specialist training) pays for itself in weeks. At this volume, the question is never "should we self-host?" -- it is "which self-hosted approach?" ### Enterprise Fleet: 10 Deployments | Solution | Total Annual Cost | Notes | | ----------------------------------------------------- | ----------------- | --------------------------------------------------------------------------- | | Annie self-hosted (10 instances at 100M tok/day each) | \$180K-\$730K | Hardware: \$50K-\$2M one-time. Specialist training shared across instances. | | Opus 4.8 API (10 instances) | \$5.5M-\$18M | Volume discounts may apply but pricing is opaque. | | GPT-5.5 API (10 instances) | \$6.4M-\$22M | Same. | | Gemini 3.1 Pro API (10 instances) | \$2.6M-\$8M | GDC sovereign deployment adds significant cost. | | Self-hosted Qwen (10 instances) | \$180K-\$730K | Same infrastructure, no verification pipeline. | **Verdict for enterprise fleet:** Annie's per-deployment economics shine at fleet scale. Ten Annie deployments cost roughly the same as one year of a single enterprise frontier API contract. Specialist training cost (\$2K-\$500K per domain model) is paid once and deployed everywhere. The fleet economics are Annie's strongest cost story. ### Cost Trajectory Over Time ```mermaid theme={null} graph LR subgraph "Cost Crossover Points" A["1M tok/day
API wins
~$5-15/day"] -->|"Volume increases"| B["10M tok/day
Crossover zone
API: $50-150/day
Self-hosted: $20-80/day"] B -->|"Volume increases"| C["100M tok/day
Self-hosted wins 10-100x
API: $700-5,000/day
Annie: $50-200/day"] C -->|"Volume increases"| D["1B tok/day
Self-hosted wins 35-100x
API: $7K-50K/day
Annie: $200-500/day"] end style A fill:#ff6b6b,color:#000 style B fill:#ffa500,color:#000 style C fill:#90ee90,color:#000 style D fill:#006400,color:#fff ``` *** ## Failure Modes and Honest Risks This section is the most important in the document. A competitive analysis that hides weaknesses is worse than no analysis at all -- it creates false confidence that leads to bad decisions. Every architecture fails. The question is how it fails and whether you can recover. ### Failure Modes by Architecture **Frontier API (MoE / Reasoning) Failure Modes:** * **Hallucination at scale.** GPT-5.5 at 57% accuracy on factual recall (43% incorrect or hallucinated). Single-pass inference with no structural correction mechanism. The model is confident when wrong. * **Service dependency.** Outages, rate limiting, policy changes, export controls, price increases -- all outside customer control. Not theoretical: demonstrated in production. * **Expert collapse in training.** The router learns to favour a subset of experts. Dead experts waste parameters. Load-balancing auxiliary losses help but directly reduce model performance. This is a training pathology, not a deployment one, but it means the model you are paying for may not be using its full capacity. * **Reasoning chain failures.** Overthinking simple problems (up to 34.5% less accurate). CoT actively harmful on implicit statistical learning tasks. Reasoning loops. Cost explosion from uncontrolled token generation. These are systemic, not edge cases. **Self-Hosted Open Model Failure Modes:** * **Undetected hallucination.** Small model hallucination rates: 54.75% at 8-32B parameters. No verification mechanism. The model is wrong more than half the time on factual recall and has no way to know it. * **Quantisation degradation.** Running 27B models at Q4 quantisation to fit on consumer hardware loses precision. The impact varies by task and is difficult to characterise in advance. * **Context window degradation.** Models claiming 1M context often degrade significantly at long context lengths. Gemini 3.1 Pro's MRCR-v2 at 128K is 84.9% -- but what about at 500K? GPT-5.4's long-context MRCR-v2 drops to 36.6%. * **No guardrails by default.** Raw open models have no content filtering, no safety layer, no output validation. The "last mile" is entirely on the operator. **Annie-Specific Failure Modes:** These are Annie's known risks. They are real, some are structural, and they must be addressed in design and operation. 1. **Classifier misclassification cascade.** The classifier is the single point of failure at pipeline entry. If it routes an insurance query to the coding specialist, the entire downstream pipeline operates on the wrong expert output. The judgment panel may catch obviously wrong output, but subtle misclassification -- routing a regulatory question to general knowledge instead of the compliance specialist -- produces plausible but unverified output. Leading routers achieve only 68-70% accuracy in research benchmarks. On queries where fewer than 3 models can answer correctly, accuracy drops to 23-25%. This is precisely when routing matters most: ambiguous queries at domain boundaries. Annie's classifier will need to be significantly better than general-purpose routers, which means domain-specific training data and continuous evaluation. 2. **Correlated errors in consensus.** The fundamental assumption of Annie's judgment panel is that independent experts will make independent errors, and consensus catches mistakes. But LLM ensembles exhibit correlated errors at approximately 60% agreement on wrong answers vs 33% expected by chance. More capable models exhibit higher error correlation. If Annie's specialists are all trained on similar data distributions, they will converge on the same wrong answers, and the judgment panel will score the consensus as correct. This is the deepest architectural risk. Mitigation requires genuine diversity: different model architectures (Mistral, Qwen, Gemma, Phi), different training data, different fine-tuning approaches. Homogeneous panels provide minimal benefit -- the diversity is not optional, it is the mechanism. 3. **Verification loop.** If the verification panel repeatedly rejects output, the pipeline retries. Annie's design choice is discard-and-restart (not incremental refinement) to prevent contamination from the failed attempt. But what if the specialist consistently produces unverifiable output for a particular query type? The pipeline could loop indefinitely. This requires a circuit breaker: after N retries, escalate to a human or return a "low confidence" response rather than looping. The tradeoff: discard-and-restart is conservative and correct but wasteful. Redundant computation accumulates. 4. **Model loading latency.** Annie loads 1-6 specialists per query, not all 12 simultaneously. Cold start latency for loading a model from disk to GPU is significant (seconds to tens of seconds depending on model size and storage speed). Hot models (already in GPU memory) respond in milliseconds. Annie must maintain a warm pool of frequently-used specialists and predict which specialists will be needed. Misprediction means cold-start latency that breaks the user experience. At 250M-27B parameters, specialists range from 1.5GB to 20GB at Q4. A single 48GB GPU can hold 2-3 large specialists or 10+ small ones simultaneously. 5. **The "good enough" problem.** This is Annie's existential strategic risk, not a technical failure mode. For most use cases -- chat, general Q\&A, draft generation, brainstorming -- a single self-hosted open model is good enough. The verification pipeline adds complexity, latency, and cost. The 4-18% accuracy improvement from ensemble consensus is real but may not be visible or valued by buyers whose primary use case does not require verified correctness. Annie must find and serve the buyers who need verification (insurance, regulatory, financial, medical) rather than trying to convince the broader market that verification matters. The market for "verified AI output" is smaller than the market for "AI output." Annie needs the former, not the latter. 6. **Consensus failure from conformity.** A February 2026 paper found heterogeneous multi-agent teams consistently failed to match their best individual member, with performance losses up to 37.6%. The failure mechanism: consensus-seeking over expertise. Agents reinforce each other's errors rather than providing independent evaluation. This is conformity bias / monoculture collapse. If Annie's judgment panel weights consensus too heavily, it may select the average answer over the correct minority answer. Rubric-based scoring mitigates this (rubrics evaluate quality, not popularity), but the risk is real. 7. **Single poisoned model.** A single deceptive or compromised model in the pipeline can nullify ensemble gains. If a specialist is fine-tuned with adversarial data, its consistently wrong output could influence the judgment panel. Security surface area grows linearly with the number of models. Annie's specialist count (up to 12) means 12 potential attack surfaces. ### Risk Severity Matrix | Risk | Probability | Impact | Mitigation Difficulty | Priority | | ---------------------------- | --------------------------------------------------- | ----------------------------------------- | ----------------------------------------------- | -------- | | Classifier misclassification | High (68-70% router accuracy is the field baseline) | High (wrong expert, wrong output) | Medium (domain-specific training helps) | Critical | | Correlated expert errors | Medium (depends on diversity) | High (consensus validates wrong answer) | High (requires genuine architectural diversity) | Critical | | Verification loop | Low (requires consistent specialist failure) | Medium (wasted compute, delayed response) | Low (circuit breaker) | Medium | | Model loading latency | Medium (depends on query patterns) | Medium (poor UX on cold start) | Medium (predictive warm pool) | Medium | | "Good enough" single model | High (real market pressure) | High (strategic, not technical) | High (requires market education) | Critical | | Consensus conformity bias | Medium | Medium (selects average over correct) | Medium (rubric design) | High | | Single poisoned model | Low (requires supply chain compromise) | High (undermines entire pipeline) | Medium (model validation, weight checksums) | Medium | *** ## Strategic Positioning ### Where Annie Should Compete Annie's defensible market is the intersection of three requirements: 1. **High domain specificity.** The task requires domain knowledge that can be encoded in a specialist model. Generic tasks do not benefit from Annie's architecture -- a frontier model or a single open model handles them better. 2. **High verification need.** The output will be used for decisions with real consequences: financial, legal, medical, regulatory. The 4-18% accuracy improvement from ensemble consensus must translate to measurable reduction in costly errors. If being wrong 54% of the time (single model) vs 40% of the time (verified pipeline) does not materially affect outcomes, the pipeline is not worth the complexity. 3. **Sovereignty required.** The deploying organisation needs control over its AI infrastructure for regulatory, geopolitical, or strategic reasons. API dependency on a foreign provider is unacceptable. **Target sectors:** Insurance (coverage determination, claims assessment), financial services (regulatory compliance, risk assessment), government (sovereign AI requirement, classified environments), healthcare (clinical decision support in regulated markets), legal (document analysis in jurisdictions with data residency requirements). **Target geographies:** Australia, New Zealand, Southeast Asia, Middle East, parts of Europe -- jurisdictions where US/China technology dependency is a strategic concern, where frontier API access may be restricted or unreliable, and where sovereign AI capability is a national priority. ### Where Annie Should NOT Compete 1. **General consumer chat.** ChatGPT has 900M weekly users. Gemini is integrated into every Google product. Annie's verification pipeline adds latency and complexity for tasks that do not need verification. Do not enter this market. 2. **State-of-the-art reasoning.** o3-pro at 98% AIME 2025, Fable 5 at 95% SWE-Bench Verified. Annie's largest specialist is 27B parameters. The raw capability gap is unbridgeable at Annie's parameter budget. Do not claim frontier-competitive general reasoning. 3. **Creative content generation.** Consensus-driven evaluation actively harms creative output by selecting the average. A single model with a distinctive voice produces better creative content than a committee. Do not position Annie for content creation. 4. **Developer tooling.** Cursor, Copilot, and Devin have deep IDE integration, massive training data on code, and models specifically optimised for coding. Annie's coding specialist (Qwen 3.6-35B-A3B) is competitive on benchmarks but has no IDE integration, no repository indexing, and no agent framework. Building these from scratch is a multi-year effort. Do not compete on developer tooling as a primary market. Annie's coding specialist is an internal capability (Annie improving Annie), not a product. 5. **Low-volume deployments.** At under 10M tokens/day, frontier APIs are cheaper and simpler than Annie's self-hosted pipeline. Do not sell Annie to organisations that process less than this unless sovereignty is a non-negotiable requirement. ### The Target Quadrant ```mermaid theme={null} quadrantChart title "Annie's Competitive Positioning" x-axis "Low Domain Specificity" --> "High Domain Specificity" y-axis "Low Verification Need" --> "High Verification Need" quadrant-1 "Annie's Sweet Spot" quadrant-2 "Reasoning Chains" quadrant-3 "Frontier API / Single Open Model" quadrant-4 "Self-Hosted Specialist" "Insurance Coverage": [0.85, 0.90] "Regulatory Compliance": [0.80, 0.95] "Clinical Decision Support": [0.75, 0.85] "Financial Risk Assessment": [0.70, 0.80] "Code Review": [0.60, 0.50] "General Q&A": [0.15, 0.10] "Creative Writing": [0.20, 0.05] "Legal Document Analysis": [0.75, 0.75] "Customer Support": [0.40, 0.30] "Data Extraction": [0.50, 0.60] ``` The four targets in the upper-right quadrant (insurance coverage, regulatory compliance, clinical decision support, financial risk assessment) are where Annie wins. Everything else is better served by simpler, cheaper, or more capable alternatives. ### Strategic Implications 1. **Lead with verification, not sovereignty.** Sovereignty is a qualifying requirement (the buyer needs it), but verification is the differentiator (the buyer wants it). "Your AI output is verified through independent cross-model consensus with domain-specific rubrics" is a value proposition. "Your AI runs on your hardware" is a checkbox. 2. **Build depth before breadth.** Annie should be exceptional at insurance before she is mediocre at six domains. One domain with a fine-tuned specialist, calibrated rubrics, measured accuracy, and production track record is worth more than six domains with generic specialists and untested judgment criteria. 3. **Measure and publish error rates.** Annie's competitive advantage is verifiable correctness. That advantage only exists if it is measured and published. Run Annie and a single open model against the same domain-specific test set. Report the error rates. If Annie's verification pipeline catches 8% more errors on insurance coverage determination, publish that number. If it catches 2%, admit it and focus engineering on widening the gap. 4. **Do not fight the "good enough" market.** Accept that most AI use cases do not need Annie. Target the use cases that do. The addressable market is smaller but the willingness to pay is higher, the competitive moat is deeper, and the switching cost for buyers is substantial once domain specialists are trained. 5. **Position as infrastructure, not product.** Annie is not a chatbot. She is a verified inference pipeline that organisations embed into their decision-making systems. The buyer is the CTO or Head of AI, not the end user. The integration point is an API and a Cognition Stream, not a chat interface. *** ## Sources ### Academic Research * Mixture-of-Agents (MoA), ICLR 2025. [arXiv:2406.04692](https://arxiv.org/abs/2406.04692) * The Law of Multi-Model Collaboration, [arXiv:2512.23340](https://arxiv.org/abs/2512.23340) * LLM Ensemble Forecasting, ForecastBench, Wharton, ICLR 2025 * RouteLLM, Berkeley, [arXiv:2406.18665](https://arxiv.org/abs/2406.18665) * SWE-Protege, [arXiv:2602.22124](https://arxiv.org/abs/2602.22124) * Correlated Errors in Large Language Models, [arXiv:2506.07962](https://arxiv.org/abs/2506.07962) * Why Do Multi-Agent LLM Systems Fail?, [arXiv:2503.13657](https://arxiv.org/abs/2503.13657) * Don't Always Pick the Highest-Performing Model, [arXiv:2602.08003](https://arxiv.org/abs/2602.08003) * The Six Sigma Agent, [arXiv:2601.22290](https://arxiv.org/abs/2601.22290) * Can AI Agents Agree?, [arXiv:2603.01213](https://arxiv.org/abs/2603.01213) * The Myth of Expert Specialization in MoEs, [arXiv:2604.09780](https://arxiv.org/abs/2604.09780) * SD-MoE: Spectral Decomposition for Expert Specialization, [arXiv:2602.12556](https://arxiv.org/abs/2602.12556) * Understanding Agent Scaling via Diversity, [arXiv:2602.03794](https://arxiv.org/abs/2602.03794) * Don't Overthink It: Shorter Thinking Chains, [arXiv:2505.17813](https://arxiv.org/abs/2505.17813) * Mind Your Step: CoT Can Reduce Performance, [arXiv:2410.21333](https://arxiv.org/abs/2410.21333) * Towards Mechanistic Understanding of Large Reasoning Models, [arXiv:2601.19928](https://arxiv.org/abs/2601.19928) * A Geometric Analysis of Small-sized Language Model Hallucinations, [arXiv:2602.14778](https://arxiv.org/abs/2602.14778) * LLMRouterBench, [arXiv:2601.07206](https://arxiv.org/abs/2601.07206) * Risk Analysis for Multi-Agent LLM Systems, Gradient Institute * Towards a Science of Scaling Agent Systems, Google Research ### Industry Sources * CB Insights SLM Market Report, 2026 * Databricks Compound AI Systems Blog * Scale AI Remote Labor Index * Gartner Agentic AI Predictions (June 2025) * ACM Transactions on Management Information Systems (Fine-tuned SLM vs GPT-4o) * ZenML/Roots Insurance Case Study * CodeAnt SWE-bench Analysis * Microsoft Copilot Reorganisation (March 2026) ### Internal References * [SOTA Landscape](annie/sota-landscape.mdx) -- State of the Art Landscape * [Annie Architecture](annie/annie-architecture.mdx) -- Annie Architecture Decisions ### Benchmark Data Sources * SWE-Bench Verified (deprecated Feb 2026, contamination concerns) * GPQA Diamond * MMLU / MMMLU * ARC-AGI-2 * AIME 2025/2026 * FrontierMath * Terminal-Bench 2.0/2.1 * Humanity's Last Exam * AA Omniscience (GPT-5.5 hallucination rate) # Hypothesis Source: https://docs.quiva.ai/annie/hypothesis # The Annie Hypothesis: Why Small Expert Models Will Win ## Executive Summary The AI industry is converging on a conclusion that the frontier labs do not want to hear: the era of the monolithic model is ending. Not because large models are bad -- they are extraordinary feats of engineering -- but because they are economically unsustainable, operationally fragile, and architecturally mismatched to the work enterprises actually need done. OpenAI does not expect profitability until 2030. Anthropic burned \$9.7 billion in 2025 alone, \$6.8 billion of it on compute. GPT-5.5 costs three times what GPT-5 cost eight months earlier. The direction of travel is clear: frontier models are getting more expensive, not less, and the organisations depending on them are one export control away from losing access entirely. Annie -- the Agentic Neural Network Intelligence Engine -- is built on a different thesis. A system of small, domain-tuned specialist models (250M-27B parameters) with consensus verification, orchestrated through Evari's high-performance Bellerophon BStream messaging backbone, will outperform monolithic frontier models on domain-specific tasks at a fraction of the cost, while providing full sovereignty and continuous improvement. This is not a theoretical position. An 8B domain-specific model has been demonstrated to beat a 120B general-purpose model on financial question-answering by nearly five percentage points (Articul8, 2026). An ensemble of three models totalling 25B parameters outperformed ChatGPT at 175B+ in real-world user engagement over thirty days of A/B testing (Chai Research, 2024). Two small language models can outperform Qwen 2.5 72B on graduate-level reasoning benchmarks (Wang et al., 2024). The hypothesis is contrarian only in the sense that it contradicts the marketing of frontier labs. It is not contrarian in the sense that matters: the research supports it, the economics demand it, and the leading minds in the field -- Sutskever, LeCun, Karpathy -- are all saying, in different ways, that the age of scaling is over and the age of architecture has begun. *** ## The Five Pillars of the Hypothesis ### 1. The Meta-MoE Advantage **Hypothesis:** System-level Mixture of Experts with specialist models matches or exceeds frontier performance on domain tasks while activating a fraction of the total parameters. Modern frontier models already use Mixture of Experts internally. Mistral Small 4 has 119 billion total parameters but activates only 6 billion per token -- roughly 5% of the network. GPT-OSS-120B activates 5.1 billion of its 117 billion parameters. The insight that drives Annie is that this same principle applies at the system level: route entire prompts to the right specialist model, and you get the benefit of a massive combined parameter space while paying the inference cost of a single small model. This is the Meta-MoE architecture: inner MoE (within each specialist that uses it) composed with outer MoE (the system-level routing across specialists). The efficiency compounds multiplicatively. **The efficiency math.** A query arrives. The classifier (250M-2B parameters, sub-10ms latency) routes it to a coding specialist. That specialist -- Qwen 3.6-35B-A3B -- has 35 billion total parameters but activates only 3 billion per token. The system has access to twelve specialists totalling hundreds of billions of parameters, but this query touches roughly 3 billion active parameters. Compare this to sending the same query to a 175B+ dense model where every parameter fires on every token. **The training cost is real, not theoretical.** Annie's sovereign base model and five unique fine-tuned domain specialist variants were built at a modest initial capital investment -- orders of magnitude below frontier training costs -- covering cloud GPU compute for base model pre-training and internal server hardware for fine-tuning and continuous improvement. This is not a projection or estimate. This is actual expenditure that produced a working multi-specialist system. The key insight: one pre-training run plus five fine-tuning runs, not five separate pre-training runs. For comparison, training a single frontier model costs \$78-191 million minimum. Annie's cost advantage starts at the training phase, not just at inference. **The research base is substantial.** Kondratyuk et al. (Google AI, 2020) demonstrated that an ensemble of two EfficientNet-B5 models matches EfficientNet-B7 accuracy using approximately 50% fewer FLOPs, with the efficiency gap widening as models get larger. The "Beyond Monoliths" paper (Quirke et al., submitted NeurIPS 2025) found that expert orchestration delivers superior performance to monolithic models, with the additional benefit of clearer evaluation metrics and narrower input spaces for testing. Chai et al. (ACL 2024) showed that representing expert LLMs as tokens in a meta-LLM's vocabulary outperforms existing multi-LLM collaboration paradigms across six expert domains. Stanford's FrugalGPT achieved 85-98% cost reduction while maintaining quality through intelligent routing. RouteLLM (UC Berkeley/Anyscale/Canva, ICLR 2025) demonstrated that learning to route LLMs with preference data achieves comparable quality at a fraction of cost. These are not marginal findings. They indicate a structural advantage for architectures that route rather than brute-force. **The domain specialisation evidence is decisive.** Articul8's benchmarks showed an 8B domain-specific model outperforming a 120B general-purpose model on financial QA (80.63% vs 75.85%), on energy sector tasks by 25.6 percentage points (96.9% vs 71.3%), and on Verilog compilation (89.2% vs 72.6%). Microsoft's Phi-3 at 3.8B parameters rivals Mixtral 8x7B and GPT-3.5 on standard benchmarks. Apple deploys a roughly 3B on-device model with runtime-swappable LoRA adapters for task specialisation at billion-device scale. The pattern is consistent: for defined work within a known domain, a well-trained small model beats a general-purpose large model. Not sometimes. Repeatedly, across independent research groups, on different tasks, with different architectures. ```mermaid theme={null} graph TB subgraph "The Meta-MoE Efficiency Cascade" Q[Incoming Query] --> C[Classifier
250M-2B params
sub-10ms] C -->|coding| S1[Coding Specialist
Qwen 3.6-35B-A3B
35B total / 3B active] C -->|reasoning| S2[Reasoning Specialist
Gemma 4 31B
31B dense] C -->|domain| S3[Domain Specialist
Phi-4 14B fine-tuned
14B dense] C -->|simple| S4[Lightweight
Gemma 4 12B
12B dense] subgraph "Inner MoE (per specialist)" S1 --> IM1["3B of 35B active
~8.6% activation"] end subgraph "Outer MoE (system routing)" OM["1 of 12 specialists selected
~8.3% of system"] end end subgraph "Effective Utilisation" E["Combined: ~0.7% of total system
parameters active per query
vs Frontier: 100% of 175B+
parameters active per token"] end IM1 --> E OM --> E style E fill:#2d5016,stroke:#4a8529,color:#ffffff style C fill:#1a3a5c,stroke:#2980b9,color:#ffffff style Q fill:#4a1a5c,stroke:#7b29b9,color:#ffffff ``` ### 2. Verification Beats Speed **Hypothesis:** For consequential work, consensus-verified answers are more valuable than fast answers. The industry's optimisation for latency is the wrong target for knowledge work. The AI industry has been optimising for the wrong metric. Time-to-first-token, tokens-per-second, sub-second responses -- these matter for chat. They do not matter for an insurance underwriting decision, a regulatory compliance assessment, or a legal document review. For work that has consequences, the question is not "how fast did you answer?" but "how confident should I be that the answer is correct?" The hallucination problem is not a bug being fixed; it is a structural property of single-pass autoregressive generation. GPT-5.5 achieved only 57% accuracy on the AA-Omniscience factual recall benchmark -- meaning 43% of its answers on factual questions were incorrect or hallucinated. Thirty-nine percent of AI-powered customer service bots were pulled back or reworked in 2024 due to hallucination. These are not edge cases. These are the central tendency of how large language models behave when asked factual questions without retrieval augmentation. **Consensus mechanisms provide a structural solution.** Wang et al. (ICLR 2023) demonstrated that sampling multiple reasoning paths and selecting the most consistent answer improves accuracy by 17.9% on GSM8K, 11.0% on SVAMP, and 12.2% on AQuA. The "Six Sigma Agent" paper (2026) mathematically proved that consensus voting with n independent agents reduces error to O(p^(ceil(n/2))), enabling exponential reliability gains. Multi-agent consistency verification (2026) reduces Expected Calibration Error by 49-74% across medical benchmarks. The VeriFY system (2026) demonstrated 9.7-53.3% hallucination reduction through consistency-based self-verification. **The Panel of LLM Evaluators (PoLL) finding is particularly relevant to Annie's architecture.** A panel of three small, diverse models from different providers outperformed a single large judge across six datasets while reducing intra-model bias at significantly lower cost. This directly validates the multi-model judgment panel design. **But naive consensus is insufficient.** The "Nine Judges, Two Effective Votes" paper (2025) showed that nine frontier LLMs from seven model families provide only about two independent votes' worth of information due to correlated errors. Simple majority voting is vulnerable to "agreeableness bias" where models follow the herd into a dominant but wrong consensus. The 2026 reasoning-tree auditing approach demonstrated that auditing branch evidence reliably selects correct minority answers over popular wrong ones. And "Beyond the Illusion of Consensus" (2026) found that dynamically generated rubrics with domain-knowledge grounding increase agreement by 22-27%. Annie's architecture addresses these limitations structurally. Its specialists are genuinely different models -- different architectures, different training data, different parameter counts -- providing real independence for consensus. This is supported by Tan et al. (EMNLP 2025) finding that cross-model probes significantly enhance error detection where within-model consistency fails. The judgment panel uses rubric-based scoring rather than simple majority voting, and the verification panel provides a separate stage of validation. This is not marketing; it is a pipeline property, and every stage is logged in the Cognition Stream via Bellerophon BStream for full auditability. **The async coworker paradigm.** The shift from "AI tool" to "AI coworker" is well underway. BCG reported that 76% of executives frame agentic AI as "AI coworker." Google launched Chrome Enterprise with agentic "AI coworker" features. But the metaphor matters: HBR research (May 2026) found that anthropomorphising AI agents as "employees" reduced accountability, while framing them as scoped-work participants with clear escalation paths improved outcomes. Annie operates as an asynchronous coworker: you give it consequential work, it researches (invoking internal platform tools and returning external tool requests for the client to execute), and produces a verified answer through multi-model consensus. Internal platform tools (Synapse searches, knowledge base queries, internal APIs) execute immediately during expert processing. External tool calls (querying customer systems, calling customer APIs, code execution in external environments) are returned as requests for the client to execute at judgment time, with all inputs and outputs verified through the consensus pipeline before returning the result. The output includes full provenance of the reasoning and verification chain. The correct analogy is not a search engine that returns in milliseconds. It is a competent colleague who takes an hour to research something properly, consult relevant systems and data, and come back with a reliable, verified answer. For knowledge work -- insurance underwriting, regulatory compliance, contract analysis -- that is what organisations actually need. ```mermaid theme={null} graph LR subgraph "Single-Pass Inference (Frontier)" FQ[Query] --> FM[Single Large Model
175B+ params] --> FA[Answer
Fast but unverified
57% accuracy
on factual recall] end subgraph "Annie Consensus Pipeline" AQ[Query] --> AC[Classifier] AC --> AE1[Expert 1] AC --> AE2[Expert 2] AE1 --> AJ[Judgment Panel
Rubric-scored evaluation
3 diverse models] AE2 --> AJ AJ --> AV[Verification Panel
Cross-architecture check
Different model families] AV -->|Pass| AA[Answer
Slower but verified
49-74% lower
calibration error] AV -->|Fail| AR[Retry with
different approach] end style FA fill:#8b1a1a,stroke:#cc3333,color:#ffffff style AA fill:#2d5016,stroke:#4a8529,color:#ffffff style AJ fill:#1a3a5c,stroke:#2980b9,color:#ffffff style AV fill:#1a3a5c,stroke:#2980b9,color:#ffffff ``` ### 3. Sovereignty Is Non-Negotiable **Hypothesis:** The Fable 5 precedent permanently changed the risk calculus for AI-dependent organisations. Any architecture that depends on a foreign API is an architecture with a single point of geopolitical failure. On June 12, 2026, the US Commerce Department ordered Anthropic to suspend its Fable 5 and Mythos 5 models globally. By June 13, both models were disabled for all users worldwide. No exemption for Five Eyes partners. No exemption for EU allies. No exemption for any country. The suspension remains in effect as of June 22 with no restoration date announced. This was not a hypothetical risk scenario from a consulting firm's slide deck. This was a real event that affected real organisations with real dependencies on a single AI provider. Nationality-based filtering was technically infeasible, so the response was a total global kill switch. Every organisation that had built workflows, products, or services on Fable 5 lost access simultaneously and without warning. **The international reaction was immediate and unambiguous.** France's Bruno Retailleau called it a "wake-up call" and accelerated support for Mistral AI. The UK's Al Carns said: "This isn't an AI story. It's the story of every industry we used to lead." The Netherlands' Wilders called for accelerating domestic AI model development. The EU's Cloud and AI Development Act, proposed June 3 (before the suspension), saw its political support dramatically accelerate. At the G7 summit on June 17, France announced a coordinated AI cooperation platform to be established within one month. SoftBank committed 75 billion euros to French AI investment. Mistral received 2.1 billion euros in state investment and a French Ministry of Armed Forces framework agreement for 2026-2030. In Australia, Kate Carruthers of UNSW stated that "sovereign AI just got real." SmartCompany reported that AI access now depends on "export controls, nationality, and geopolitical considerations." The pre-existing tension was already visible: Trump had directed all federal agencies to cease using Anthropic in February 2026, months before the Fable 5 suspension. **The market data confirms the structural shift.** Eighty-one percent of enterprises now run three or more AI model families, up from 13% a year ago. Sixty-plus nations have published AI strategies, with thirty-plus committing funding. The sovereign AI infrastructure market is projected to reach \$301.6 billion by 2040. These are not reactions to Fable 5 alone; they are the culmination of years of growing awareness that API dependency is operational risk. **Australia's position is particularly exposed.** The National AI Plan (March 2026) established no standalone AI Act, relying instead on sector regulators. The \$1.2 billion defence budget for sovereign AI and autonomous systems, the ASD-AWS "Top Secret Cloud" at approximately AUD \$2 billion over a decade, CDC's 200MW AI campus near Perth at AUD \$415 million, Macquarie's IC3 Super West 47MW AI data centre at AUD \$350 million, and NEXTDC's S7 site at 650MW partnered with OpenAI -- all of this infrastructure is building toward Level 2-3 sovereignty. But infrastructure without sovereign application-layer capability is a data centre with American software running in it. Five Eyes membership provides no exemption from US export controls. That was demonstrated, not theorised. Annie provides Level 3-4 sovereignty: full deployment on customer-controlled infrastructure, no external API dependency, ability to disconnect from the internet and continue functioning. Every specialist model uses an Apache 2.0 or MIT licence. Every component runs on consumer-grade hardware. The total hardware investment for a full deployment is \$5,000-\$200,000, compared to the data-centre-scale infrastructure required for frontier models. This is not sovereignty for governments with billion-dollar budgets. This is sovereignty for any organisation that needs it. ```mermaid theme={null} graph TB subgraph "Before Fable 5 Suspension (Pre-June 2026)" direction TB B1[Enterprise AI Strategy] --> B2[Choose Best Frontier API] B2 --> B3[Build on Single Provider] B3 --> B4[Assume Continuous Access] B4 --> B5["Risk Assessment:
'Provider lock-in'
categorised as medium risk"] end subgraph "After Fable 5 Suspension (June 12, 2026)" direction TB A1[Enterprise AI Strategy] --> A2[Multi-Model / Multi-Provider
81% now run 3+ families] A2 --> A3[Sovereign Deployability
Required, not optional] A3 --> A4[No Single Points of
Geopolitical Failure] A4 --> A5["Risk Assessment:
'API dependency on foreign provider'
categorised as critical risk"] end subgraph "The Fable 5 Precedent" F1["June 12: US Commerce Dept
orders global suspension"] --> F2["June 13: All users worldwide
lose access — no exceptions"] F2 --> F3["No Five Eyes exemption
No EU exemption
No restoration date"] end B5 -.->|"Permanent shift"| A1 F3 -.->|"Caused by"| A1 style B5 fill:#8b6914,stroke:#cc9a1d,color:#ffffff style A5 fill:#8b1a1a,stroke:#cc3333,color:#ffffff style F3 fill:#4a1a5c,stroke:#7b29b9,color:#ffffff ``` ### 4. The Cost Inversion **Hypothesis:** Frontier model costs are increasing while small model capabilities are increasing, creating a crossover point that has already arrived for domain-specific work. The conventional wisdom is that AI models will get cheaper over time. The data says the opposite for frontier models. GPT-5.5 costs three times what GPT-5 cost eight months earlier. Gemini 3.5 Flash tripled versus its predecessor. Fable 5 output pricing sits at \$50 per million tokens; Opus 4.8 at \$25; GPT-5.5 Pro at \$180. The direction is up, not down. Gartner analyst Will Sommer explained the mechanism: "Yes, token costs are coming down, that is going to unlock relatively low-value capabilities," but higher-value applications "are going to be more expensive, not less." Agentic AI requires 5-30 times more tokens per query than generative AI, which means the effective cost per unit of work is increasing even as the per-token price decreases for basic inference. **The training cost divergence is even more stark.** A single frontier training run now costs \$78-100 million (GPT-4) to \$191 million (Gemini Ultra), with projections exceeding \$1 billion by 2027. Anthropic spent \$6.8 billion on compute alone in 2025. Google has guided \$175-185 billion in capital expenditure. OpenAI does not expect profitability until at least 2030 and projects \$150 billion on inference costs alone through 2030. Annie's entire initial investment was a modest capital outlay -- orders of magnitude below frontier training costs -- covering cloud GPU compute for specialist training and internal server hardware for post-training, continuous improvement, and ongoing operations. This is not a theoretical cost projection -- it is actual expenditure. Adding an additional specialist costs \$2K-\$500K per model. Fine-tuning a 7B model costs under \$5 per run. The entire Annie hardware investment for a full deployment is \$5K-\$200K, one time. These are not comparable numbers to frontier training. They are different categories of expenditure. **At scale, the economics are decisive.** At 100 million tokens per day: | Metric | Frontier API (Opus 4.8) | Frontier API (Gemini 3.1 Pro) | Annie Self-Hosted | | ------------------- | ----------------------- | ----------------------------- | -------------------- | | Daily cost | \$1,500-5,000 | \$700-2,200 | \$50-200 | | Annual cost | \$550K-1.8M | \$255K-800K | \$18K-73K | | Hardware investment | None | None | \$5K-200K (one-time) | Annie is 10-100 times cheaper at scale. Even multiplying Annie's per-query compute by 2-5 times for the consensus pipeline, 5 times self-hosted is still dramatically cheaper than 1 times frontier API. **The sustainability problem is structural, not cyclical.** Anthropic projects approximately \$29 billion in losses against \$25-30 billion in revenue for 2026. OpenAI projects \$14 billion in losses. The frontier labs sustain themselves through what Jacobin described as "circular financing rather than genuine profitability." ChatGPT costs approximately \$17 billion per year to run with 800-900 million weekly users, but only 35 million are paying subscribers. As Professor Andy Wu of Harvard Business School observed: "The pool of people willing to pay \$20 a month for generative AI is smaller than that willing to pay \$20 a month for Netflix." This is not a business model with a path to equilibrium. It is a capital-burning exercise sustained by the assumption that scale will eventually produce returns. Meanwhile, the "Small is Sufficient" paper ([arXiv:2510.01889](https://arxiv.org/abs/2510.01889)) demonstrated that switching to appropriately-sized models is 65.8% more energy efficient at the cost of only 3.9% utility loss, with task-specific energy reductions reaching 92.8% for time series forecasting and 80.6% for speech recognition. Globally, model selection could save 31.9 TWh in 2025 alone -- equivalent to the annual output of five nuclear reactors. The crossover has already happened. For defined domain work, small specialist models are cheaper, more accurate, more energy-efficient, and more operationally resilient than frontier APIs. The remaining question is not whether organisations will adopt this architecture, but how quickly. ```mermaid theme={null} graph TB subgraph "Frontier Cost Trajectory (Rising)" FC1["2023: GPT-4 Training ~$100M"] --> FC2["2024: Gemini Ultra ~$191M"] FC2 --> FC3["2025: Anthropic $6.8B compute"] FC3 --> FC4["2026: Projected >$1B per run"] FC4 --> FC5["2030: OpenAI projects $150B
inference costs alone"] FP1["GPT-5: $X/MTok"] --> FP2["GPT-5.5: 3X/MTok
(8 months later)"] FP2 --> FP3["Fable 5: $50/MTok output"] end subgraph "Specialist Cost Trajectory (Falling)" SC1["New specialist: $2K-$500K"] --> SC2["Fine-tuning run: <$5"] SC2 --> SC3["Full deployment hardware:
$5K-$200K one-time"] SC3 --> SC4["Annual operation at 100M tok/day:
$18K-$73K"] end subgraph "The Crossover" X["At 100M tokens/day:
Frontier: $255K-$1.8M/year
Annie: $18K-$73K/year
10-100x cost advantage
Already here for domain work"] end FC5 --> X SC4 --> X style X fill:#2d5016,stroke:#4a8529,color:#ffffff style FC5 fill:#8b1a1a,stroke:#cc3333,color:#ffffff style SC4 fill:#1a3a5c,stroke:#2980b9,color:#ffffff ``` ### 5. Continuous Domain Improvement **Hypothesis:** Small models trained on actual user interactions will outperform large models trained on internet data for specific domain work. The flywheel effect creates compounding advantage over time. Frontier models are trained on internet-scale data. This gives them broad knowledge but shallow domain expertise. A model that has seen billions of web pages knows something about insurance underwriting, but it does not know how a specific company underwrites specific products for specific markets. The data it would need to be genuinely expert in that domain does not exist on the internet. It exists in the company's systems, in the interactions between underwriters and the tools they use, in the decisions made and the reasoning behind them. Annie's sleep cycle training captures exactly this data. Every interaction flows through the Cognition Stream via Bellerophon BStream. Classification decisions, expert responses, judgment scores, verification outcomes -- all logged, all structured, all available as training signal. During off-peak hours, specialists fine-tune on accumulated interaction data at a cost of under \$5 per run. The specialist that handles insurance pricing today is better than the one that handled it yesterday, because it has seen one more day of real insurance pricing work. **The research supports this mechanism.** "Fine-Tune an SLM or Prompt an LLM?" ([arXiv:2505.24189](https://arxiv.org/abs/2505.24189), 2025) found that fine-tuning a small language model can outperform prompting a frontier LLM on domain-specific tasks. Industry analyses suggest domain-focused models lower hallucination rates by 70-85% compared to general-purpose systems. The "Small is Sufficient" paper demonstrated that appropriately-sized, well-trained models sacrifice only 3.9% utility while achieving 65.8% energy savings -- and that utility gap narrows further with domain-specific fine-tuning. **The flywheel compounds.** Each interaction generates training data. Each training cycle improves specialist performance. Improved performance leads to more usage, which generates more training data. The specialist becomes more expert in the specific domain of the specific organisation over time. A frontier model cannot do this. It is too expensive to retrain (DeepSeek V3's \$5.6 million training cost shocked the industry as 10-20 times lower than assumed, and it is still orders of magnitude more than Annie's per-specialist cost). It is too general to benefit from narrow domain data without extensive prompt engineering. And its training data is static between major releases. Apple understood this at device scale: a roughly 3B on-device model with runtime-swappable LoRA adapters for task specialisation, deployed across billions of devices, each adapting to its user's patterns. Annie applies the same principle at the enterprise scale: small models, domain-adapted, continuously improving on the work they actually do. **The economic asymmetry is important.** Retraining a frontier model costs \$500 million or more. Adding or improving an Annie specialist costs \$2,000-\$500,000. Fine-tuning an existing specialist on new interaction data costs under \$5. This means Annie can iterate daily where frontier labs iterate quarterly or annually. Over the course of a year, that is not a marginal advantage. It is a structural one. ```mermaid theme={null} graph TB subgraph "The Continuous Improvement Flywheel" U[User Interactions] -->|"Logged to
Cognition Stream"| CS[Bellerophon BStream
Full observability] CS -->|"Classification, responses,
judgments, verifications"| TD[Training Data
Accumulated daily] TD -->|"Off-peak hours
<$5 per fine-tuning run"| SC[Sleep Cycle Training
Specialist fine-tuning] SC -->|"Updated weights
deployed next cycle"| SP[Improved Specialists
Better domain performance] SP -->|"Higher quality
More usage"| U end subgraph "Frontier Comparison" FR["Frontier Retraining
$500M+ per cycle
Quarterly at best
No org-specific data"] end subgraph "The Compounding Gap" CG["Day 1: Annie matches frontier on domain tasks
Day 30: Annie leads by fine-tuning margin
Day 365: Annie has seen 365 cycles of real domain work
Frontier has seen 0-2 retraining cycles"] end SP --> CG FR --> CG style CG fill:#2d5016,stroke:#4a8529,color:#ffffff style SC fill:#1a3a5c,stroke:#2980b9,color:#ffffff style FR fill:#8b6914,stroke:#cc9a1d,color:#ffffff ``` *** ## The Contrarian Position The Annie hypothesis runs counter to the dominant narrative in Silicon Valley, which holds that scaling will continue to produce breakthroughs, that a single model will eventually do everything well enough, and that users want instant answers above all else. We disagree. But intellectual honesty requires engaging with the strongest version of the opposing arguments, not the weakest. ### "Scaling will continue to work" **The argument:** Every time someone has predicted the end of scaling, they have been wrong. GPT-4 was better than GPT-3. GPT-5 was better than GPT-4. The curve has not flattened yet, and there is no theoretical reason it must. **Why we think it is wrong, but not obviously wrong.** Scaling may continue to produce marginal improvements. The question is not whether larger models are better in absolute terms, but whether the improvement per dollar is sustainable. Ilya Sutskever, who co-led the scaling revolution at OpenAI, said at NeurIPS 2024: "Pretraining as we know it will end. The 2010s were the age of scaling, now we're back in the age of wonder and discovery." Yann LeCun left Meta and raised \$1.03 billion for AMI Labs to build world models, arguing that LLMs "cannot, on their own, reach human-level intelligence." MIT research warns that "in the next five to ten years, things are very likely to start narrowing" on returns from the biggest models. The data wall is real: Chinchilla-optimal training for a 1T parameter model requires roughly 20T tokens, while high-quality internet text is estimated at 10-50T tokens total. We are approaching the limits of what internet data can teach. And the economic wall is real too: when training runs cost \$1 billion and the organisations funding them are losing \$14-29 billion per year, the question is not "can we build a bigger model?" but "should we?" Our position is not that scaling is dead. It is that scaling alone is insufficient, and that for domain-specific work, the marginal return on scale is already below the marginal return on specialisation. ### "One model to rule them all" **The argument:** Convenience wins. Developers do not want to manage twelve models. They want one API call. The history of technology is the history of consolidation: one operating system, one cloud provider, one search engine. **Why we think it is wrong for consequential work.** The history of technology is also the history of specialisation at the application layer. Enterprises do not use one database. They do not use one programming language. They do not use one security tool. They use the right tool for each job, integrated through middleware and orchestration. Annie's innovation is not requiring organisations to manage twelve models; it is managing them so the organisation does not have to. The classifier routes, the pipeline orchestrates, and the user interacts with a single interface. More fundamentally, "one model to rule them all" is the wrong framing for regulated industries. An insurer does not want one model that is pretty good at everything. They want provably correct answers on the specific tasks that matter to their business, with an audit trail that satisfies their regulator. A general-purpose model that is 80% accurate on insurance pricing is worthless if a specialist can be 96% accurate. The stakes are too high for "good enough." ### "Users want instant answers" **The argument:** ChatGPT won because it was fast. Users have been trained to expect sub-second responses. Anything that takes minutes will feel broken. **Why we think it is wrong for work, but right for chat.** The distinction between chat and work is the key insight. For a question like "what is the capital of France?" -- yes, speed is the correct optimisation target. For a question like "should we underwrite this policy at \$4.2 million?" -- no reasonable professional would prefer a fast wrong answer to a slow correct one. The analogy is email versus instant messaging. Both are communication tools. Both are valuable. But nobody argues that email should be replaced by instant messaging because instant messaging is faster. They serve different purposes. Annie serves the "email" purpose: consequential work that benefits from deliberation, verification, and auditability. Chatbots serve the "instant messaging" purpose. We are not competing with chatbots. We are competing with the alternative to chatbots that most organisations have not yet built. The error-propagation problem makes this distinction critical. A 10% error rate is acceptable for chatbots -- users can re-ask. It is catastrophic for autonomous agents executing business logic, where one failed step corrupts downstream state. For agentic work, reliability is worth latency. ### "Frontier models will get cheaper" **The argument:** Token prices have fallen dramatically. GPT-3.5 cost far more per token than GPT-4-mini. Moore's Law applies to inference. Give it time. **Why the data says otherwise.** Per-token prices for low-capability inference are falling. Per-token prices for high-capability inference are rising. GPT-5.5 costs three times what GPT-5 cost eight months earlier. Gemini 3.5 Flash tripled versus its predecessor. Fable 5 sits at \$50 per million output tokens. The cheap tokens are the ones that do simple work; the expensive tokens are the ones that do hard work. And the total cost of ownership is what matters, not the per-token price. Agentic AI requires 5-30 times more tokens per query than single-pass chat. An agentic workflow that makes ten API calls, each requiring reasoning-mode inference, consumes orders of magnitude more tokens than a chat response. The effective cost per unit of business value is increasing even as the base per-token price decreases. Uber spent \$3.4 billion on AI in 2025 and exhausted its entire 2026 AI budget by April. Per-developer consumption increased 5-20 times with no matching documented output value increase. The "it will get cheaper" argument assumes that consumption will remain constant while prices fall. In practice, consumption is exploding while prices for capable inference are rising. *** ## White Paper Roadmap The following white papers are designed to convert Annie's architectural thesis into credible, published evidence. Each paper targets a specific audience, fills a documented gap in the existing literature, and builds the evidence base needed for investor and enterprise conversations. ### 1. "The Meta-MoE: Hierarchical Mixture of Experts for Sovereign AI" **Target audience:** Technical leaders, AI architects, academic researchers, investors with technical due diligence requirements. **Key thesis:** System-level Mixture of Experts -- routing entire queries to domain specialists rather than tokens to expert sub-networks -- achieves frontier-class performance on domain tasks at a fraction of the compute cost, while enabling sovereign deployment on commodity hardware. Inner MoE (within specialists) and outer MoE (across the system) compound multiplicatively to produce extreme parameter efficiency. **Evidence base:** Kondratyuk et al. (2020) on ensemble efficiency; Quirke et al. (NeurIPS 2025 submission) on expert orchestration outperforming monoliths; Chai et al. (ACL 2024) on expert-as-token representation; FrugalGPT and RouteLLM on cost-quality routing; Microsoft Phi-3 and Apple Foundation Models on small model capability; Articul8 domain benchmarks on specialisation advantages. **Publication gap:** No existing paper bridges multi-model orchestration with sovereign AI goals. The routing and ensemble literature is purely about cost-quality optimisation. Nobody has published on how multi-model architectures serve sovereignty (vendor independence, jurisdictional data control, resilience against provider lock-in). The dynamic model routing survey (arXiv 2026) identifies generalisation to new models and domains, and underexplored multi-stage cascades, as specific open research gaps. **Estimated scope:** Full white paper (15-20 pages). This is the core architecture document and must be comprehensive. **Priority:** HIGH. This is the foundational document for all technical conversations with investors, data centre partners, and enterprise customers. *** ### 2. "Verified AI: Why Consensus Pipelines Beat Single-Pass Inference for Enterprise Decisions" **Target audience:** Enterprise buyers in regulated industries (insurance, finance, healthcare, government), compliance officers, risk managers. **Key thesis:** For decisions with material consequences -- underwriting, compliance assessment, legal review -- a multi-model consensus pipeline with rubric-based evaluation and cross-architecture verification produces structurally more reliable outputs than any single model, regardless of that model's size. The latency cost of verification is a worthwhile trade for the reliability gain. **Evidence base:** Wang et al. (ICLR 2023) on self-consistency; "Six Sigma Agent" (2026) on mathematical proof of consensus error reduction; PoLL on small diverse panels outperforming single large judges; "Nine Judges, Two Effective Votes" on the importance of genuine model diversity; Chain-of-Verification (Meta, 2023); VeriFY (2026); multi-agent consistency verification on medical benchmarks; "Beyond the Illusion of Consensus" on rubric-grounded evaluation. **Publication gap:** No publication addresses lightweight, practical consensus mechanisms for enterprise multi-model systems that do not rely on blockchain. The literature is split between simple majority voting (shown to be flawed) and complex cryptographic approaches (impractical for real-time enterprise use). Consensus mechanisms that deliberately use different model architectures to increase verification robustness are absent from the literature. **Estimated scope:** Full white paper (12-15 pages), with potential for a shorter arXiv preprint if empirical results from insurance domain are included. **Priority:** HIGH. Verification is Annie's most defensible differentiator against both frontier APIs and other multi-model approaches. This paper makes the case that matters most to enterprise buyers. *** ### 3. "Sovereign AI Without Sovereign Budgets: Application-Layer Architecture as a Sovereignty Strategy" **Target audience:** Government decision-makers, defence procurement, enterprise CISOs, Australian data centre operators (CDC, Macquarie, NEXTDC), sovereign AI investors. **Key thesis:** Sovereign AI does not require building a national frontier lab. Application-layer architecture choices -- multi-model, vendor-diverse, locally-deployable, open-weight -- achieve functional sovereignty at a cost accessible to mid-market enterprises and smaller nations, not just superpowers. The Fable 5 suspension demonstrated that API dependency is geopolitical risk, and Five Eyes membership provides no exemption. **Evidence base:** Fable 5 suspension timeline and international reaction; sovereign AI market projections (\$301.6B by 2040); 81% multi-model enterprise adoption; Australian National AI Plan and defence budget; CDC, Macquarie, NEXTDC investment data; Mistral state investment and French military framework; EU Cloud and AI Development Act; NVIDIA sovereign AI white paper (as infrastructure-only framing to contrast against). **Publication gap:** This is a Tier 1 gap. Every major publication treats sovereign AI as an infrastructure problem -- buy GPUs, build data centres, deploy cloud. Almost nothing exists on sovereign AI at the application and orchestration layer. The insurance and regulated-industry-specific sovereign AI literature is essentially nonexistent as rigorous technical publication. **Estimated scope:** Full white paper (15-20 pages). This is the primary document for the Australian data centre and government audience. **Priority:** HIGH. Directly supports the Evari fundraising narrative and data centre partnership conversations. Should be published before investor meetings. *** ### 4. "The Asynchronous AI Coworker: From Chatbot Paradigm to Knowledge Work Architecture" **Target audience:** Enterprise product leaders, CIOs, agentic AI buyers, industry analysts. **Key thesis:** The chatbot paradigm -- synchronous, single-turn, optimised for speed -- is architecturally mismatched to knowledge work. An event-driven, asynchronous architecture where AI operates as a scoped-work participant (not an instant-answer machine) produces better outcomes for consequential tasks, with clearer governance, full auditability, and natural integration into existing business workflows. **Evidence base:** Gartner 40% agent penetration prediction; \$201.9B agentic AI spending (2026); HBR research on AI agent framing (scoped-work vs employee metaphor); 39% chatbot pullback rate; error propagation in multi-step agent workflows; Temporal for durable agent execution; AutoGen v0.4 async-first architecture; CIO article on three non-negotiable agentic infrastructure pillars (event-driven messaging, observability, governance). **Publication gap:** The architectural paradigm shift from synchronous to asynchronous AI is discussed in industry blogs and Gartner reports but lacks rigorous academic treatment. No paper formally analyses the reliability, latency, cost, and user-experience tradeoffs of async agentic architectures versus synchronous chatbots in production. The error-propagation problem in multi-step agent workflows is mentioned everywhere but formally modelled nowhere. **Estimated scope:** Short paper or long-form blog series (8-12 pages). More accessible than the architecture papers, designed for broad enterprise readership. **Priority:** MEDIUM-HIGH. Positions Annie's product thesis against the dominant chatbot paradigm. Strong for enterprise sales conversations. *** ### 5. "Continuous Learning Through Sleep Cycles: Domain Adaptation Economics for Enterprise AI" **Target audience:** AI/ML engineers, enterprise AI platform teams, data science leaders, investors evaluating defensibility. **Key thesis:** Continuous fine-tuning of specialist models on accumulated interaction data -- at under \$5 per training run -- creates a compounding performance advantage that no frontier model can replicate. The flywheel of use-train-improve-use turns an operational cost into an appreciating asset, and the economics favour the small-model approach by orders of magnitude. **Evidence base:** "Fine-Tune an SLM or Prompt an LLM?" ([arXiv:2505.24189](https://arxiv.org/abs/2505.24189)); Apple Foundation Models LoRA adapter approach; Phi-3 data-curation results; DeepSeek V3 training cost shock (\$5.6M); frontier retraining costs (\$500M+); domain-specific hallucination reduction (70-85%); "Small is Sufficient" energy efficiency findings. **Publication gap:** Rigorous, peer-reviewed empirical studies on small model economics in production are scarce. The 75% cost reduction claims come from blog posts and vendor marketing. A controlled study showing, for a real enterprise workload, the actual cost-quality-latency tradeoffs of a continuously-improving SLM-first architecture versus an LLM-only approach with real production data would be genuinely novel and highly citable. **Estimated scope:** Technical paper (10-12 pages). Best published with real production data from an Annie deployment, making it dependent on pilot timing. **Priority:** MEDIUM. High long-term value as a defensibility argument, but requires production data to be credible. Sequence after initial pilots are running. *** ### 6. "The Cognition Stream: Event-Driven AI Orchestration Through Bellerophon BStream" **Target audience:** AI infrastructure engineers, platform architects, DevOps and MLOps teams, technical investors. **Key thesis:** AI orchestration requires a purpose-built event-driven backbone that provides durable message delivery, full observability, and replay capability -- not the request-response patterns inherited from web APIs. Bellerophon BStream provides the Cognition Stream that makes Annie's multi-stage pipeline auditable, debuggable, and resilient, treating AI reasoning as a first-class event stream rather than a black-box function call. **Evidence base:** CIO article on three non-negotiable agentic infrastructure pillars; AutoGen v0.4's adoption of event-driven architecture; Temporal's emergence as standard for durable agent execution; ACM CAIS 2026 conference framing compound AI systems as "the norm"; the observability gap in current AI systems; event sourcing patterns from financial systems applied to AI reasoning. **Publication gap:** The "how to build reliable AI infrastructure" space is dominated by cloud provider marketing (Azure AI, Google Vertex, AWS Bedrock). Independent technical publications on event-driven AI orchestration architectures that are not vendor-specific are rare. The connection between event sourcing (well-understood in financial systems) and AI reasoning auditability is essentially unexplored in the published literature. **Estimated scope:** Technical report (10-15 pages) with architecture diagrams and implementation patterns. Could be accompanied by a blog series introducing the concepts incrementally. **Priority:** MEDIUM. Important for technical credibility but less urgent for investor or enterprise buyer conversations than the sovereignty and verification papers. *** ### 7. "The Rapport Model: Adaptive AI Communication for Long-Term Enterprise Relationships" **Target audience:** UX researchers, enterprise product teams, HR and change management leaders, CHI (human-computer interaction) community. **Key thesis:** Enterprise AI effectiveness depends not just on answer quality but on the quality of the human-AI working relationship over time. A lightweight rapport model that learns communication preferences, adapts to expertise level, remembers context across sessions, and calibrates tone along a continuous spectrum produces measurably better adoption, trust, and outcomes than stateless API interactions. **Evidence base:** PersonaMem-v2 (arXiv 2025) on implicit user personas; CloneMem and "Beyond Dialogue Time" on temporal semantic memory; State of AI Agent Memory 2026 on open problems (temporal abstraction, cross-session evolution, privacy architecture); HBR research on AI agent framing; Deloitte finding that only 6% achieve significant enterprise-wide AI impact (suggesting adoption, not capability, is the bottleneck). **Publication gap:** This is a Tier 1 gap. AI rapport -- the quality of the human-AI working relationship over time -- is almost entirely absent from the literature. Personalisation research focuses on recommendation systems and content delivery. Nobody is publishing on how an enterprise AI assistant builds and maintains a productive working relationship with a specific user over weeks and months. This is a genuinely underserved area with real commercial value. **Estimated scope:** Position paper (8-10 pages) initially, expanding to a full research paper with longitudinal data from Annie deployments. **Priority:** MEDIUM-LOW for investor readiness (the concept is harder to quantify), but HIGH for product differentiation and long-term category creation. Consider submitting to CHI 2027. *** ### 8. "Small Models, Big Decisions: Empirical Cost-Quality Analysis for Insurance AI" **Target audience:** Insurance industry executives, actuaries, insurtech investors, regulatory bodies. **Key thesis:** For insurance-specific tasks -- pricing, underwriting, claims assessment, regulatory compliance -- a system of fine-tuned small models demonstrably outperforms frontier APIs on accuracy, cost, latency, and auditability, with empirical data from production deployments. The insurance industry's unique requirements (explainability, audit trails, regulatory compliance, data sovereignty) make it the ideal proving ground for specialist AI architecture. **Evidence base:** Articul8 domain benchmarks (8B vs 120B on financial QA); 39% chatbot pullback rate; GPT-5.5 57% accuracy / 43% error rate on factual recall; Annie cost comparisons at scale; consensus pipeline verification rates; production data from Annie insurance pilots (when available). **Publication gap:** Insurance and fintech-specific sovereign AI literature is essentially nonexistent as rigorous technical publication. Current coverage is infrastructure vendors selling cloud services. A controlled study of SLM-first architecture performance on real insurance workloads would be the first of its kind. **Estimated scope:** Full white paper (12-15 pages), potentially co-authored with an insurance industry body or university research group for credibility. **Priority:** HIGH for enterprise sales in the insurance vertical. Dependent on pilot data availability. Should be the first paper published with production evidence. *** ### 9. "Model Diversity as a Verification Feature: Why Architectural Heterogeneity Matters for AI Reliability" **Target audience:** AI safety researchers, ML engineers building evaluation systems, enterprise AI governance teams. **Key thesis:** The reliability of multi-model verification depends critically on the genuine independence of the models involved. Architecturally homogeneous panels (same family, similar training data) provide far fewer effective independent votes than their headcount suggests. Deliberate architectural heterogeneity -- different model families, different parameter counts, different training approaches -- is a design requirement for reliable consensus, not an implementation detail. **Evidence base:** "Nine Judges, Two Effective Votes" (2025) on correlated errors in homogeneous panels; Tan et al. (EMNLP 2025) on cross-model probes enhancing error detection; PoLL on diverse panels outperforming homogeneous ones; "Beyond the Illusion of Consensus" on fragile sample-level agreement; "Six Sigma Agent" on independence requirements for consensus error reduction. **Publication gap:** Using architecturally different models deliberately to increase consensus robustness is an emerging finding with no dedicated paper. The "Nine Judges" paper identifies the problem; no paper proposes the solution as an architectural principle. This is a gap Annie is well-positioned to fill given its deliberate use of diverse specialist architectures. **Estimated scope:** Short paper (6-8 pages). Suitable for arXiv preprint or workshop paper at a safety/evaluation venue. **Priority:** MEDIUM. Strengthens the verification narrative and addresses the strongest technical counterargument to consensus approaches. *** ## Publication Strategy and Sequencing ### Phase 1: Investor Readiness (Months 1-2) Publish papers 1, 2, and 3 as company white papers (ungated). These form the core narrative for data centre partner and investor conversations: here is the architecture, here is why it is more reliable, here is why sovereignty matters and how we deliver it affordably. No production data required; these are architecture and evidence-synthesis papers. ### Phase 2: Market Positioning (Months 3-4) Publish papers 4 and 6 as blog series leading to formal publications. Paper 4 (async coworker) positions Annie against the chatbot paradigm for enterprise buyers. Paper 6 (Cognition Stream) establishes technical credibility with platform engineers who will evaluate Annie for integration. ### Phase 3: Production Evidence (Months 5-8) Publish papers 5, 8, and 9 once pilot data is available. Paper 8 (insurance empirical) is the highest-impact publication in this phase -- the first rigorous, production-data-backed comparison of specialist versus frontier performance on real insurance workloads. Paper 5 (sleep cycles) demonstrates the flywheel in action. Paper 9 (model diversity) provides the academic anchor for the verification thesis. ### Phase 4: Category Creation (Months 6-12) Publish paper 7 (rapport model) with longitudinal data from Annie deployments. Target CHI 2027 or a human-computer interaction venue. This is a longer-term investment in defining a new category. ### Publishing Venues | Venue | Best For | Timeline | | ---------------------------------------------------------- | ---------------------------------------------------------- | -------------------------- | | Company blog / technical report (ungated) | Broad reach, GenAI discovery, authentic operator voice | 2-4 weeks per piece | | arXiv preprint | Academic credibility, citation, technical audience | Submit with empirical data | | NeurIPS 2027 Industry Track | Mixed academic/industry, high prestige | Check submission deadlines | | AAAI-27 Emerging Trends | Architecture and systems focus | Submission likely H1 2027 | | Insurance / fintech conferences (ACORD, Insurtech Connect) | Buyer audience, underserved by serious AI content | 2026-2027 calendar | | CHI 2027 | Human-computer interaction, rapport model | Submission likely H2 2026 | | Partnership publications | Co-author with university or industry body for credibility | Ongoing | *** ## Evidence Index ### Pillar 1: The Meta-MoE Advantage **Ensemble efficiency:** * Kondratyuk et al. (Google AI, 2020). "When Ensembling Smaller Models is More Efficient than Single Large Models." [arXiv:2005.00570](https://arxiv.org/abs/2005.00570). Two EfficientNet-B5 models match B7 at 50% fewer FLOPs. * Chai Research (2024). "Blending Is All You Need." [arXiv:2401.02994](https://arxiv.org/abs/2401.02994). Three models at 25B total outperform ChatGPT 175B+ in user retention. * Wang et al. (2024). "SLM-MUX: Orchestrating Small Language Models for Reasoning." [arXiv:2510.05077](https://arxiv.org/abs/2510.05077). Two SLMs outperform Qwen 2.5 72B on GPQA and GSM8K. * Li et al. (2024). "More Agents Is All You Need." [arXiv:2402.05120](https://arxiv.org/abs/2402.05120). Scaling agent count improves outcomes. * Jiang et al. (ACL 2023). "LLM-Blender." [arXiv:2306.02561](https://arxiv.org/abs/2306.02561). Pairwise ranking and generative fusion for ensemble LLMs. * Quirke et al. (submitted NeurIPS 2025). "Beyond Monoliths." [arXiv:2506.00051](https://arxiv.org/abs/2506.00051). Expert orchestration outperforms monolithic models. * Chai et al. (ACL 2024). "An Expert is Worth One Token." [arXiv:2403.16854](https://arxiv.org/abs/2403.16854). Expert LLMs as tokens in meta-LLM vocabulary. **Small model capability:** * Microsoft Phi-3 Technical Report (2024). 3.8B model rivals Mixtral 8x7B and GPT-3.5. * [arXiv:2505.24189](https://arxiv.org/abs/2505.24189) (2025). "Fine-Tune an SLM or Prompt an LLM?" Fine-tuned SLM outperforms prompted frontier LLM on domain tasks. * Apple Foundation Models (2024-2025). \~3B on-device model with LoRA adapters at billion-device scale. * Articul8 benchmarks (2026). 8B domain model beats 120B general model on financial QA (80.63% vs 75.85%), energy sector (96.9% vs 71.3%). **Cost-optimised routing:** * Chen et al. (Stanford, TMLR 2024). "FrugalGPT." [arXiv:2305.05176](https://arxiv.org/abs/2305.05176). 85-98% cost reduction while maintaining quality. * UC Berkeley/Anyscale/Canva (ICLR 2025). "RouteLLM." [arXiv:2406.18665](https://arxiv.org/abs/2406.18665). Comparable quality at fraction of cost via learned routing. * [arXiv:2605.06116](https://arxiv.org/abs/2605.06116) (2026). "Policy-Guided Stepwise Model Routing for Cost-Effective Reasoning." **Scaling limitations:** * Ilya Sutskever, NeurIPS 2024. "Pretraining as we know it will end." * Yann LeCun / AMI Labs (2026). \$1.03B raised to build world models; LLMs "cannot reach human-level intelligence." * Falcon 180B (2023) outperformed by Llama 3 8B (2024). ### Pillar 2: Verification Beats Speed **Consensus mechanisms:** * Wang et al. (ICLR 2023). "Self-Consistency Improves Chain of Thought Reasoning." [arXiv:2203.11171](https://arxiv.org/abs/2203.11171). +17.9% accuracy on GSM8K. * Verga et al. (2024). "Panel of LLM Evaluators (PoLL)." Three small diverse models outperform single large judge. * Zhao et al. (2024). "Language Model Council." [arXiv:2406.08598](https://arxiv.org/abs/2406.08598). Council rankings more robust than individual judge. * "The Six Sigma Agent" (2026). [arXiv:2601.22290](https://arxiv.org/abs/2601.22290). Consensus error reduces to O(p^(ceil(n/2))). * "Nine Judges, Two Effective Votes" (2025). [arXiv:2605.29800](https://arxiv.org/abs/2605.29800). Correlated errors limit effective votes in homogeneous panels. * Tan et al. (EMNLP 2025). [arXiv:2505.17656](https://arxiv.org/abs/2505.17656). Cross-model probes enhance error detection. **Verification and hallucination reduction:** * Meta AI (2023). "Chain-of-Verification (CoVe)." [arXiv:2309.11495](https://arxiv.org/abs/2309.11495). 4-8% factual accuracy improvement. * VeriFY (2026). [arXiv:2602.02018](https://arxiv.org/abs/2602.02018). 9.7-53.3% hallucination reduction. * Multi-Modal Fact-Verification Framework (2025). [arXiv:2510.22751](https://arxiv.org/abs/2510.22751). 67% hallucination reduction. * "Sample, Scrutinize and Scale" (2025). [arXiv:2502.01839](https://arxiv.org/abs/2502.01839). Gemini v1.5 surpassed o1-Preview via inference-time verification scaling. * Multi-agent consistency verification (2026). 49-74% reduction in Expected Calibration Error on medical benchmarks. **Rubric-based evaluation:** * "Beyond the Illusion of Consensus" (2026). [arXiv:2603.11027](https://arxiv.org/abs/2603.11027). Domain-grounded rubrics increase agreement by 22-27%. * "Rubric Is All You Need" (ACM ICER 2025). [arXiv:2503.23989](https://arxiv.org/abs/2503.23989). Question-specific rubrics outperform generic criteria. **Hallucination prevalence:** * GPT-5.5 accuracy: 57% on factual recall benchmarks (AA-Omniscience) -- 43% of answers incorrect or hallucinated. * 39% of AI chatbots pulled back or reworked in 2024 due to hallucination (ComputerTechReviews, 2025). ### Pillar 3: Sovereignty Is Non-Negotiable **The Fable 5 precedent:** * US Commerce Department order, June 12, 2026. Global suspension of Fable 5 and Mythos 5. * Triggered by NSA red-team exercise; Mythos broke into nearly all NSA classified systems. * No Five Eyes, EU, or allied exemptions. UK exemption collapsed. * Trump directed federal agencies to cease using Anthropic, February 2026. **International reaction:** * France: Bruno Retailleau "wake-up call"; accelerated Mistral support. * UK: Al Carns: "This isn't an AI story. It's the story of every industry we used to lead." * Netherlands: Wilders called for domestic AI model development. * EU: Cloud and AI Development Act political support accelerated post-suspension. * G7 summit, June 17: France announced coordinated AI cooperation platform. * Australia: Kate Carruthers (UNSW): "sovereign AI just got real." **Market data:** * 81% of enterprises run 3+ AI model families (up from 13% a year ago). * Sovereign AI infrastructure market: \$301.6B by 2040 (Roots Analysis). * 60+ nations published AI strategies; 30+ committed funding. * SoftBank EUR 75B French AI investment; Mistral EUR 2.1B state investment. **Australian context:** * National AI Plan (March 2026); \$1.2B defence sovereign AI budget. * ASD-AWS "Top Secret Cloud" \~AUD \$2B/decade. * CDC 200MW AI campus near Perth (AUD \$415M). * Macquarie IC3 Super West 47MW (AUD \$350M). * NEXTDC S7 650MW (AUD \$7B+, OpenAI partnership, H2 2027). ### Pillar 4: The Cost Inversion **Frontier pricing (output per million tokens):** * Fable 5: \$50. Opus 4.8: \$25. GPT-5.5: \$30. GPT-5.5 Pro: \$180. Gemini 3.1 Pro: \$12-18. **Frontier training and operating costs:** * GPT-4 training: \$78-100M (2023). Gemini Ultra: \~\$191M (2023). * Anthropic 2025 spending: \$9.7B (\$6.8B compute). Committed \$50B infrastructure. * OpenAI: \$25B+ training spend projected 2026. \$150B inference costs projected through 2030. * Google: \$175-185B capex guided. * ChatGPT: \~\$17B/year to run; 800-900M weekly users, 35M paying subscribers. * Anthropic projected 2026: \~\$29B loss against \$25-30B revenue. * OpenAI projected 2026: \~\$14B loss. **Annie economics:** * New specialist: \$2K-\$500K. Fine-tuning run: \<\$5. Full deployment hardware: \$5K-\$200K one-time. * At 100M tokens/day: \$18K-73K/year vs \$255K-\$1.8M/year frontier. * 10-100x cost advantage at scale. **Industry analysis:** * Gartner (Will Sommer): Higher-value applications "are going to be more expensive, not less." * Jacobin: "No evidence exists that current spending models achieve cost-benefit equilibrium." * Professor Andy Wu (HBS): Paying subscriber pool smaller than Netflix. * Uber: \$3.4B AI spend in 2025; exhausted 2026 budget by April. **Energy efficiency:** * "Small is Sufficient" ([arXiv:2510.01889](https://arxiv.org/abs/2510.01889)). 65.8% energy savings at 3.9% utility loss. * Task-specific: time series forecasting 92.8% reduction, speech recognition 80.6%. * Global model selection could save 31.9 TWh (2025), 106 TWh (2028). * AI data centre electricity: 460-490 TWh (2025); projected 945 TWh (2030), 1,200 TWh (2035). * Inference now 63% of lifecycle energy (passed training Q3 2025). ### Pillar 5: Continuous Domain Improvement **Domain adaptation:** * [arXiv:2505.24189](https://arxiv.org/abs/2505.24189) (2025). Fine-tuned SLM outperforms prompted frontier LLM on domain tasks. * Articul8 domain benchmarks. 70-85% hallucination reduction from domain-focused models. * Microsoft Phi-3. 3.8B model trained on curated data rivals much larger models. * Apple Foundation Models. LoRA adapters for runtime task specialisation. **Training economics:** * DeepSeek V3: \~\$5.6M training cost (10-20x lower than industry assumed). * Annie fine-tuning: \<\$5 per run. * Frontier retraining: \$500M+. **The flywheel:** * Every interaction generates training signal via Cognition Stream. * Off-peak fine-tuning at negligible cost. * Daily improvement cycles vs quarterly (at best) frontier retraining. * Specialists improve on actual user work, not abstract benchmarks. ### Cross-Cutting: Market Direction **Agentic AI market:** * Gartner: 40% of enterprise apps to include AI agents by end of 2026 (up from \<5% in 2025). * Agentic AI spending: \$201.9B (2026), overtaking chatbot spending by 2027. * IDC: Agentic AI to handle 40% of Global 2000 jobs by end of 2026. * 76% of executives frame agentic AI as "AI coworker" (BCG, Nov 2025). * Gartner: 1,445% surge in multi-agent system inquiries Q1 2024 to Q2 2025. **Compound systems:** * AlphaCode 2: Outperformed 85% of human competitors via generation and filtering. * AlphaGeometry 2: Gold-medal level geometry via neuro-symbolic hybrid. * FactSet: 55% to 85% accuracy via modularised compound architecture. * ACM CAIS 2026: "Compound AI systems have become the norm." **Analyst consensus:** * TechCrunch: "2026 will be the year the tech gets practical." * AT\&T Andy Markus: "Fine-tuned SLMs will be the big trend." * Andrej Karpathy: Useful autonomous agents are "a decade out." * Deloitte: Only 6% of enterprises achieve significant enterprise-wide AI impact. * Gartner: >40% of agentic AI projects could be cancelled by 2027 due to escalating costs. # Partner propositions Source: https://docs.quiva.ai/annie/partner-propositions # Partner Propositions: CDC, Macquarie Data Centres, NEXTDC ## Executive Summary Evari is approaching three of Australia's leading data centre operators with a proposition that goes beyond a traditional tenancy arrangement. The model is three-sided: Evari brings Annie, a sovereign AI platform that creates a new category of workload. Data centre partners provide certified sovereign infrastructure and compute capacity. Enterprise customers -- government agencies, banks, insurers, defence organisations -- pay to run Annie on that infrastructure. Every Annie deployment is a tenant on the partner's facility. Every new customer Evari signs is revenue the partner did not have to sell. This is a workload-generation partnership, not a capital raise. Data centre operators are in an unprecedented build-out phase. CDC is constructing 800MW of new capacity. Macquarie is opening a 47MW AI-optimised facility in September 2026. NEXTDC has 550MW to fill at its S7 campus. Their shared challenge is the same: filling that capacity with creditworthy, long-term tenants running real production workloads. Annie addresses that challenge directly. The three partners are complementary, not competing. CDC provides the deepest government certification and the greenfield Perth opportunity. Macquarie provides the Dell/NVIDIA sovereign AI factory hardware stack and the warmest introduction path. NEXTDC provides national scale, public company credibility, and a diversification story against OpenAI concentration risk. ```mermaid theme={null} graph TB subgraph "Three-Sided Value Model" direction TB E["Evari / Annie
Sovereign AI platform
Enterprise workloads
Domain specialists"] DC["Data Centre Partner
CDC / Macquarie / NEXTDC
Certified sovereign infrastructure
Compute capacity"] EC["Enterprise Customers
Government agencies
Banks & insurers
Defence & health"] E -->|"Deploys platform on
partner infrastructure"| DC E -->|"Brings paying
customers"| DC DC -->|"Provides certified
sovereign compute"| E DC -->|"Provides credibility
& introductions"| E EC -->|"Pays for Annie
on partner facilities"| DC EC -->|"Pays subscription
for Annie"| E E -->|"Delivers sovereign AI
coworker capability"| EC end style E fill:#2563eb,stroke:#1e40af,color:#fff style DC fill:#059669,stroke:#047857,color:#fff style EC fill:#d97706,stroke:#b45309,color:#fff ``` **What Evari brings to every partner:** * A new category of sovereign AI workload that does not exist today in Australian data centres * Enterprise customers in government, defence, financial services, healthcare, and insurance who will pay for Annie and consume infrastructure capacity * Modest hardware requirements (standard inference GPUs with 32GB+ VRAM, not training-grade clusters) that deliver high revenue per rack at low power density * Differentiation narrative: "the only data centre in Australia running an Australian-built sovereign AI coworker platform" * The Fable 5 precedent as a compelling demand driver -- every organisation that lost AI access on June 13 is now evaluating sovereign alternatives **Why this scales beyond Evari:** Annie is a platform, not a model. She ships with Evari's own sovereign specialists, but her pipeline roles are open slots. Every customer who deploys Annie can bring their own domain models into the architecture. A bank adds their risk model. A defence contractor plugs in a classified specialist. Each new customer model is additional compute running on the partner's infrastructure. This means Annie is not a single tenant -- she is a workload multiplier. The more customers on the platform, the more models in the pipeline, the more racks consumed. **What Evari asks from every partner:** * Compute capacity without massive upfront capital investment * Access to the partner's customer relationships and sales channels * Credibility by association with certified sovereign infrastructure * Potential co-investment to fund Annie's go-to-market *** ## The Market Opportunity ### The Sovereign AI Imperative On June 12, 2026, the US Commerce Department ordered Anthropic to suspend global access to Fable 5 and Mythos 5. Because nationality-based filtering proved technically infeasible, Anthropic disabled both models for all users worldwide -- including paying enterprise customers. As of June 22, they remain suspended with no restoration date. No exemption exists for Five Eyes partners, EU allies, or any other country. This is not a theoretical risk. It has been demonstrated. Every Australian government agency, bank, insurer, and healthcare provider running critical workloads on US-hosted frontier models now faces the same question: what happens when access is severed without warning? Annie addresses this by providing sovereign AI that was built with sovereign infrastructure. Annie's architecture consists of one sovereign base model (sparse MoE, pre-trained from scratch) fine-tuned into five unique domain specialist variants, plus Qwen 3.6:27B. These six physical models fill twelve logical pipeline roles -- classifier, domain experts, judgment models, verification models, rapport model, etc. -- with some models serving multiple roles. The sovereign base model was trained from scratch on 1.5 trillion tokens of Evari domain data plus 15 trillion tokens of open-source permissible datasets. The five unique fine-tuned variants specialize in domains like coding, insurance, general knowledge, regulatory compliance, etc. The total initial training investment was a modest capital outlay -- orders of magnitude below frontier training costs. This is not an adaptation or fine-tune of an external model. This is capability built entirely within Australia's control. Qwen 3.6:27B has extensive additional training and serves as the largest specialist for complex reasoning. All post-training occurs on Evari's internal servers. The result: a sovereign AI system that cannot be suspended by an external government and has no dependency on external model weights that could become subject to export controls. The European Commission proposed the Cloud and AI Development Act on June 3, aiming to triple European data centre capacity over five to seven years. France, the Netherlands, and the UK have all announced accelerated sovereign AI programmes. Australia has no equivalent response yet, but the political and institutional demand is building. ### The Australian Opportunity The Australian sovereign AI market sits at the intersection of four forces: 1. **Government mandate.** The DTA's AI procurement guidance (December 2025) requires AI Impact Assessments for all new AI use cases, with mandatory compliance by December 2026. GovAI Chat trials began in April 2026. The government is building the procurement framework; it now needs the platforms. 2. **Defence investment.** The 2025-26 budget allocated \$1.2 billion for sovereign AI and autonomous systems. The Advanced Strategic Capabilities Accelerator (ASCA) committed \$3.4 billion over its first decade, with roughly 40% directed at AI. Project Redspice is a \$9.9 billion cyber and intelligence programme with significant AI components. 3. **Regulatory pressure.** Financial services (APRA), healthcare (TGA), and insurance (ASIC) all face regulatory requirements around data sovereignty, model transparency, and audit trails that US-hosted AI services cannot currently satisfy. 4. **Critical infrastructure designation.** Data centres are now classified as critical infrastructure under the Security of Critical Infrastructure Act 2018. AI platforms processing government and financial data will face increasing scrutiny around where computation happens and who controls it. ### Why Now The Fable 5 suspension created a window. Before June 12, sovereign AI was a policy aspiration. After June 12, it is an operational requirement with executive-level urgency. Organisations that were evaluating sovereign alternatives over an 18-month horizon are now compressing that timeline to months. Annie is ready for pilot deployment. The data centre partners are in a build-out phase with capacity to fill. The enterprise customers have budget and urgency. The three-sided model works because all three sides are motivated simultaneously. ### Target Sectors | Sector | Why They Care | Estimated Market Size | | ---------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------- | | **Federal Government** | Fable 5 demonstrated the risk. Hosting Certification Framework mandates sovereign infrastructure. DTA procurement guidance requires AI Impact Assessments. | 98 federal agencies, \$1.8B annual IT spend on external providers | | **Defence** | \$1.2B allocated for sovereign AI in 2025-26. Five Eyes interoperability requires Australian-controlled AI for classified work. | \$3.4B ASCA commitment over decade | | **Financial Services** | APRA prudential requirements on data sovereignty. Board-level concern about operational resilience after June 12. | Four major banks, \~20 significant insurers | | **Healthcare** | Patient data sovereignty requirements. Medicare and PBS claims processing at massive scale. | \$8B+ annual health IT spend | | **Insurance** | Domain-specific AI requirements that generalist models handle poorly. Regulatory audit trail obligations. | \$70B+ annual premium market | *** ## Partner Proposition: CDC Data Centres ### About CDC CDC is Australia's largest sovereign, privately owned data centre operator. Founded in 2007 by Greg Boorer in Canberra, it now operates 18 data centres with five more under construction, delivering over 800MW across Canberra, Sydney, Melbourne, Auckland, and Perth. The ownership structure tells the story: Infratil (49.75%), Australian Future Fund (34.55%), Commonwealth Superannuation Corporation (12.04%), and management (3.66%). Two of the three institutional shareholders are Australian government-affiliated entities. No foreign hyperscaler or Chinese capital involvement. The February 2025 share transaction implied an AUD \$17 billion enterprise valuation. CDC holds Certified Strategic status -- the highest level under the Australian Government Hosting Certification Framework -- across all of its facilities. It was one of the first three providers to achieve this certification. Government contracts have surpassed \$1 billion, anchored by a \$226 million Services Australia extension and \$91.5 million in Defence contracts. CDC's AI strategy centres on the Firmus/NVIDIA partnership (Project Southgate), targeting 1.6GW of AI factories through 2028 using 18,500 NVIDIA GB300 GPUs. It also hosts the Monash MAVERIC supercomputer (\$60M, NVIDIA GB200 NVL72) and CSIRO's Virga HPC cluster (NVIDIA H100s). **Key decision-makers:** Greg Boorer (CEO, ultimate signoff), Dr Jack Dan (CSO, strategic partnerships), Andrew Kirker (MD Hyperscale), Simon Black (CCO, commercial). ### What Annie Brings to CDC **A new sovereign AI workload category that fills capacity.** CDC is building hundreds of megawatts of new capacity across Marsden Park (504MW, \$3.1B), Melbourne (800MW+ target), and Perth (200MW, \$415M Stage 1). Their biggest risk is filling it. Every Annie customer is a tenant on CDC infrastructure -- government agencies, banks, insurers -- exactly the customer profile CDC already serves. Annie does not compete with CDC's existing tenants; it brings new ones. **Differentiation from the Firmus relationship.** Firmus is focused on massive-scale AI training factories (GB300 GPUs, 1.6GW). Annie occupies a different tier entirely: enterprise-grade, inference-focused, domain-specific AI coworker services. A 7B specialist model running insurance coverage determinations is not competing with a GB300 training cluster. Annie complements Firmus by adding a sovereign AI application layer on top of CDC's infrastructure story. **Modest hardware, high margins.** Annie's twelve specialist models range from 250M to 27B parameters and run on standard inference GPUs with 32GB+ VRAM (e.g., RTX 5090, L40S, A6000). A full Annie deployment requires a fraction of the power and cooling that a training-grade cluster demands. For CDC, this means high revenue per rack at standard power density -- no need for the 100kW+ liquid-cooled racks that Firmus workloads require. Annie workloads can fill standard-density capacity that might otherwise sit empty while high-density AI racks are under construction. **An anchor workload for the Perth campus.** The Maddington campus (200MW, operational 2026) is greenfield with no legacy tenants. It is explicitly positioned for "AI and advanced technology deployments in Western Australia." Annie as an early-stage platform partner -- committed before the facility opens -- is strategically valuable to CDC as a signal to the WA market. Perth's proximity to the mining/resources sector (growing AI/ML needs), the defence sector (naval base, Five Eyes), and renewable energy from the WA grid aligns with Annie's target customer base. **The sovereign narrative amplified.** CDC's entire brand is sovereign. Annie amplifies that story: not just sovereign infrastructure, but a sovereign AI platform running on sovereign infrastructure. Australian-built AI, Australian-owned data centres, Australian government-affiliated shareholders. No foreign AI provider involved. No US CLOUD Act exposure. This is the pitch CDC's sales team makes to government customers, made stronger. ### What CDC Brings to Annie **Government certification at the highest level.** CDC's Certified Strategic status across all facilities means Annie can immediately serve PROTECTED-level government workloads without building or certifying infrastructure. This is years of accreditation work that Annie inherits by deploying on CDC. **Government customer relationships.** CDC serves Services Australia, the Department of Defence, and dozens of federal and state agencies. These are exactly the organisations most likely to pilot sovereign AI. CDC's sales team already has the relationships; Annie provides them with a new product to sell. **Compute capacity without upfront capital.** Annie does not need to build data centres or buy servers. CDC provides the physical environment, power, and cooling. Evari deploys software and lightweight GPU hardware. The capital requirement drops from hundreds of millions to hundreds of thousands. **Credibility by association.** "Annie runs on CDC's Certified Strategic infrastructure, the same facilities that host Services Australia and the Department of Defence." This is a sentence that opens doors with government procurement teams. **The Perth greenfield timing.** Getting into Perth early -- before the facility opens and tenancy commitments are locked -- gives Annie favourable terms and a strategic position as the anchor AI platform in Western Australia. ### Proposed Engagement Model **Phase 1: Pilot (Months 1-6).** Deploy Annie on CDC infrastructure at one Canberra facility (Hume or Fyshwick). Three to five pilot customers drawn from CDC's existing government relationships. Minimal hardware: four to eight inference GPUs with 32GB+ VRAM for the specialist model pool. CDC provides rack space and power at reduced pilot rates. Evari provides the platform, specialist models, and customer onboarding. Pilot scope includes domain-specific tuning: Annie is currently optimised for insurance and fintech domains. Government use cases will require training or fine-tuning of specialists for government-specific workflows, policy interpretation, and regulatory domains. This is factored into pilot timelines and deliverables. **Phase 2: Expansion (Months 6-18).** Scale based on pilot outcomes. Add Sydney (Marsden Park) and Perth (Maddington) deployments. Expand to financial services and insurance customers. Graduate to commercial colocation rates with committed power draw. Discuss co-investment if pilot metrics justify it. Specialist model expansion continues as new use cases are discovered. **Phase 3: Strategic Partnership (Months 18-36).** Formal partnership agreement positioning Annie as CDC's sovereign AI platform offering. Joint go-to-market for government and enterprise customers. CDC sales team trained to sell Annie alongside their core colocation services. Potential equity investment from CDC or its shareholders. Continuous specialist development for new domains identified through customer engagement. **The commercial structure:** Colocation lease for Annie's infrastructure at CDC facilities, with a revenue-sharing component on Annie subscription fees generated from customers introduced through CDC's sales channels. This aligns incentives: CDC earns more when Annie succeeds, and Annie succeeds when CDC's customers adopt it. Specialist training and domain customisation are handled through Evari's professional services, separate from colocation licensing. ### Alignment and Risks **Where interests align perfectly:** * Both need government and enterprise tenants * Both benefit from the sovereign narrative * Annie fills standard-density capacity while Firmus fills high-density * Annie's growth directly increases CDC's recurring revenue * Perth timing works for both parties **Where friction could arise:** * CDC may view Annie as too small or too early for a partnership at CDC's current scale (\$17B valuation) * The Firmus/NVIDIA relationship could create exclusivity concerns -- does Firmus see Annie as a competitor? * CDC's 555MW US hyperscaler deal suggests a strategic direction toward hyperscale, not boutique AI platforms * If Annie's pilot customers are too few or too small, CDC may not allocate sales attention **Mitigation:** Position Annie explicitly as complementary to Firmus (application layer vs. training infrastructure). Start small with a pilot that requires minimal CDC commitment. Demonstrate demand through signed pilot customers before asking for a broader partnership. Target the Perth campus specifically, where CDC has capacity to fill and fewer competing priorities. *** ## Partner Proposition: Macquarie Data Centres ### About Macquarie Macquarie Data Centres is a division of Macquarie Technology Group (ASX: MAQ, market cap \~\$1.82B), a 100% Australian-owned company founded in 1992 by David and Aidan Tudehope. It operates five data centres across Sydney and Canberra, all holding Certified Strategic status. The flagship asset is IC3 Super West: a 47MW AI-optimised facility at the Macquarie Park campus, opening September 2026 with Phase 1 delivering 6MW. It supports direct-to-chip liquid cooling for high-density GPU workloads and is being positioned as "the only data centre to add new AI capacity to Sydney's north zone in 2026." Macquarie's AI strategy centres on the Dell Sovereign AI Factory partnership (announced August 2025), combining Dell PowerEdge XE9680 servers, NVIDIA GPUs, and NVIDIA Spectrum-X networking inside IC3 Super West. The target sectors are healthcare, financial services, education, research, government, and critical infrastructure -- sectors requiring strict data residency and regulatory compliance. In March 2026, the Australian Government's National Reconstruction Fund committed \$200 million to Macquarie for sovereign cloud expansion, AI-enabled cybersecurity capability, and a new sovereign data facility. This is the federal government putting taxpayer capital directly behind Macquarie's sovereign infrastructure vision. Macquarie serves 42% of Australian Federal Government agencies, employs 200+ NV1+ cleared engineers, and operates a 24/7 Security Operations Centre with cleared staff. IC5 "The Bunker" in Canberra is Zone 5 Top Secret capable. **Key decision-makers:** David Hirst (Group Executive, Macquarie Data Centres -- primary for AI infrastructure partnerships), Aidan Tudehope (MD Hosting Group -- government/sovereign cloud angle), David Tudehope (Group CEO). **Evari's existing connection:** A warm introduction path exists, which materially changes the engagement approach. ### What Annie Brings to Macquarie **The software layer for the Dell/NVIDIA hardware stack.** Macquarie has the physical infrastructure (IC3 Super West) and the hardware (Dell AI Factory with NVIDIA). What they likely need are the software platforms, AI applications, and managed services that run on top. Annie is that missing layer. The Dell partnership gives Macquarie a sovereign AI factory; Annie gives the factory a product to manufacture. Without a platform layer, Macquarie has infrastructure looking for applications. Annie provides the applications looking for infrastructure. **Tenants for IC3 Super West.** IC3 Super West opens in September 2026 with 6MW in Phase 1, expanding to 19MW and then 47MW. Macquarie needs anchor tenants and compelling use cases to fill that capacity. Annie's deployment plus the enterprise customers Annie brings directly addresses Macquarie's most pressing business need in Q3 2026. **A competitive answer to NEXTDC's OpenAI partnership.** NEXTDC's OpenAI deal (announced December 2025) put pressure on Macquarie to demonstrate a comparable sovereign AI ecosystem. The Dell partnership is infrastructure; Annie provides the application layer that makes the ecosystem tangible. "Macquarie hosts Australia's sovereign AI coworker platform" is a headline that competes with "NEXTDC hosts OpenAI." Annie arrives in 2026, a full year before NEXTDC's S7 campus delivers its first phase in H2 2027. **Enterprise customers in Macquarie's target sectors.** Macquarie's customer base is 42% of Federal Government agencies, Fortune 500 enterprises, and critical infrastructure providers. Annie's target customers are government, defence, financial services, healthcare, and insurance. The overlap is nearly total. Annie brings workloads from customers Macquarie already knows or wants to know. **Alignment with the NRF investment mandate.** The \$200M NRF investment is earmarked partly for "AI-enabled cybersecurity capability" and sovereign cloud expansion. A sovereign AI platform that runs on Macquarie infrastructure, serving government customers, directly addresses the outcomes the NRF investment was designed to achieve. ### What Macquarie Brings to Annie **AI-optimised infrastructure purpose-built for GPU workloads.** IC3 Super West was specifically designed for high-density AI compute with liquid cooling support. While Annie's specialist models run on standard inference GPUs (32GB+ VRAM) rather than training-grade clusters, having access to purpose-built AI infrastructure provides headroom for scaling and for customers who want to run larger fine-tuning workloads alongside Annie's inference pipeline. **The Dell/NVIDIA partnership ecosystem.** Access to the Dell AI Factory hardware stack, NVIDIA DGX-Ready certification, and the engineering expertise that comes with a 15-year Dell relationship. If Annie's hardware requirements evolve (larger specialists, more concurrent models, customer-specific fine-tuning), the Dell/NVIDIA ecosystem provides a scaling path. **Deep government relationships and cleared personnel.** 200+ NV1+ cleared engineers is a capability that takes years and significant investment to build. For Annie deployments serving classified government workloads, Macquarie provides the cleared operational staff that Evari does not yet have. IC5 The Bunker offers Zone 5 Top Secret capability for the most sensitive deployments. **The warm introduction.** Evari's existing connection to Macquarie reduces the cold-start problem that plagues data centre partnerships. The conversation can start at a strategic level rather than fighting through procurement gatekeepers. **The Macquarie University partnership.** Macquarie's April 2026 strategic partnership with Macquarie University for joint research in cloud computing, cybersecurity, and data centre engineering provides a potential research collaboration channel for Annie's specialist model development. ### Proposed Engagement Model **Phase 0: Introduction (Weeks 1-4).** Leverage the existing connection for a strategic conversation with David Hirst (data centre partnerships) and Aidan Tudehope (sovereign cloud). Frame Annie as the software platform that completes their Dell AI Factory stack. Provide a technical brief demonstrating Annie's architecture and how it maps onto IC3 Super West's capabilities. **Phase 1: IC3 Super West Launch Partner (Months 2-8).** Position Annie as one of the anchor tenants for IC3 Super West's September 2026 opening. Deploy the Annie platform on a modest initial allocation (two to four racks, standard inference GPUs with 32GB+ VRAM). Run three to five pilot customers from Macquarie's government and enterprise base. This gives Macquarie a tangible "sovereign AI platform" story for the IC3 Super West launch -- a headline beyond "we have racks and cooling." Pilot scope includes domain-specific tuning: Annie is currently optimised for insurance and fintech. Government and enterprise customers will require specialist training for their target use cases. Specialist development is factored into pilot timelines. **Phase 2: Joint Go-to-Market (Months 8-18).** Develop a joint value proposition: "Macquarie Sovereign AI, powered by Annie." Macquarie's sales team offers Annie as a managed service running on their certified infrastructure. Annie handles the AI platform; Macquarie handles the infrastructure, security, and government compliance. Revenue share on jointly acquired customers. Specialist model expansion for government and enterprise domains continues through partnership. **Phase 3: Integrated Offering (Months 18+).** Annie becomes part of Macquarie's sovereign cloud services portfolio. Potential co-investment or equity arrangement. Expansion to the planned 150MW+ campus. Integration with Macquarie Government's cyber security services for an AI-enabled security offering (aligned with NRF investment outcomes). Specialist development roadmap covers new domains as customer base grows. ### Alignment and Risks **Where interests align perfectly:** * IC3 Super West needs tenants for its September 2026 opening -- Annie arrives at exactly the right time * The Dell AI Factory needs a software platform -- Annie fills the gap * Macquarie needs a competitive response to NEXTDC's OpenAI deal -- Annie provides one * Both target the same customer segments (government, defence, finance, healthcare) * The NRF investment creates government expectations that Macquarie will deliver sovereign AI outcomes * The warm introduction reduces partnership friction **Where friction could arise:** * Macquarie is a smaller operator (\$1.82B market cap vs. CDC at \$17B and NEXTDC at \$9.2B) -- does this limit their reach? * IC3 Super West's Phase 1 is only 6MW -- initial capacity is constrained * Macquarie may prefer a larger, more established AI platform partner (a recognisable global brand) * If Annie is perceived as too early-stage, Macquarie's risk-averse government customers may hesitate * Dell and NVIDIA have their own AI platform preferences that may not include Annie **Mitigation:** The warm introduction is the single biggest advantage. Use it to get an honest assessment of Macquarie's appetite before investing significant effort. Position Annie not as a replacement for Dell/NVIDIA software but as a complementary Australian sovereign layer. Start with a small footprint that demonstrates value without requiring Macquarie to make a significant commitment. *** ## Partner Proposition: NEXTDC ### About NEXTDC NEXTDC (ASX: NXT, market cap \~\$9.17B) is Australia's largest listed data centre operator. It runs 17 data centres across Australia and one in Malaysia, with a development pipeline exceeding 1.5GW. The company was founded in 2010 and has grown to become the second-largest data centre operator in Oceania behind Equinix. The defining recent event is the OpenAI partnership. In December 2025, NEXTDC signed a Memorandum of Understanding under OpenAI's "OpenAI for Countries" programme -- the first in Asia-Pacific. The deal covers a next-generation 550MW AI campus at S7 Eastern Creek in western Sydney, valued at \$7 billion. First phase delivery is expected H2 2027. NEXTDC is DTA Certified Strategic across its entire national network. It holds NVIDIA DGX-Ready certification, supports rack densities up to 150kW (scalable to 600kW), and offers direct-to-chip and immersion cooling. It is carrier-neutral and cloud-neutral, hosting AWS, Azure, GCP, Oracle, and IBM across its facilities. The company is not yet profitable (H1 FY2026 net loss of \$39.4M) but is growing aggressively, with contracted utilisation up 60% to 667MW and a forward order book up 83% to 544MW. FY26 capex guidance is \$2.7B-\$3.0B. Pro forma liquidity stands at \$8.4 billion. The board includes Steve Smith, former CEO of Equinix, who grew that company from \$2B to \$34B market value. His appointment signals NEXTDC's global ambitions. **Key decision-makers:** Craig Scroggie (CEO), David Dzienciol (Chief Customer and Commercial Officer), Simon Cooper (Group Chief Development Officer). ### What Annie Brings to NEXTDC **Diversification from OpenAI dependency.** The OpenAI MoU is non-binding and covers a single US company whose models are now subject to demonstrated export control risk. If US export controls tighten further -- or if the Fable 5 precedent is applied to GPT models -- NEXTDC's flagship AI story is exposed. Annie provides a hedge: an Australian-built, Australian-controlled sovereign AI platform that cannot be suspended by a US Commerce Department directive. Having multiple AI tenants, including a sovereign one, is a stronger story than dependence on a single US provider. This is not speculative. The Fable 5 suspension proved that access to US-built AI can be revoked without warning, without geographic exemption, and without recourse. NEXTDC's board and shareholders understand this risk. Annie is the insurance policy. **Access to sectors OpenAI may not reach.** Classified government workloads, defence applications, and highly regulated sectors (banking, insurance) may require AI platforms that operate entirely under Australian jurisdiction with no foreign dependencies. OpenAI, as a US company subject to US law, faces structural barriers in these sectors. Annie, as an Australian platform on Australian infrastructure, does not. NEXTDC can offer Annie for the sovereign-sensitive workloads and OpenAI for everything else. **Enterprise pilot customers in NEXTDC's target sectors.** NEXTDC is DTA Certified Strategic and explicitly targets government, financial services, defence, education, healthcare, and research for S7. Annie's target customer base is identical. Every Annie pilot customer consumes NEXTDC infrastructure. **Complementary positioning, not competitive.** OpenAI provides frontier general-purpose AI. Annie provides sovereign domain-specific AI. These serve different use cases and different procurement requirements. A government agency might use OpenAI (via NEXTDC) for general productivity and Annie (also via NEXTDC) for classified domain work. NEXTDC becomes the venue where both coexist, offering customers sovereign choice rather than a single provider. **Modest hardware requirements that fill capacity immediately.** S7's first phase is H2 2027. But NEXTDC has existing facilities (S1-S6, C1, P1-P2, M1-M3) with available capacity today. Annie's standard GPU requirements (32GB+ VRAM) can deploy into existing NEXTDC facilities immediately, generating revenue while S7 is under construction. Annie does not need to wait for the 550MW campus; it can start in a single rack at any Certified Strategic NEXTDC site. ### What NEXTDC Brings to Annie **National scale.** 17 data centres across Sydney, Melbourne, Brisbane, Perth, Canberra, Adelaide, Darwin, and more. Annie can deploy nationally on a single partner's infrastructure, serving customers in every major Australian market without managing multiple data centre relationships. **Public company credibility.** NEXTDC is an ASX-listed company with a \$9.17B market cap, institutional backing from La Caisse de depot et placement du Quebec, and Steve Smith (ex-Equinix CEO) on the board. "Annie runs on NEXTDC infrastructure" carries weight with enterprise procurement teams that "Annie runs on a startup's servers" does not. **AI infrastructure expertise from the OpenAI partnership.** NEXTDC's engagement with OpenAI has forced them to build deep expertise in AI infrastructure: liquid cooling, high-density power distribution, GPU cluster networking. This expertise benefits all AI tenants, including Annie. Even though Annie's requirements are modest by comparison, the operational maturity that comes from hosting frontier AI workloads provides a better platform for every AI customer. **Enterprise customer relationships.** NEXTDC's 770+ ecosystem partners, carrier-neutral interconnection fabric, and existing enterprise customer base provide a sales channel that Annie could not build independently. NEXTDC's CCO (David Dzienciol) and VP Customer (Adam Scully) manage relationships with exactly the organisations Annie needs to reach. **The JV co-investment structure.** NEXTDC is actively seeking third-party capital partners for its S4/S7 joint venture platform, targeting \$15B over 10+ years for 850MW of new Sydney capacity. While Evari is not a capital partner at that scale, the JV structure signals that NEXTDC is open to partnership models beyond simple tenancy. ### Proposed Engagement Model **Phase 1: Existing Facility Pilot (Months 1-6).** Deploy Annie at one or two existing NEXTDC facilities (C1 Canberra for government, S-series Sydney for enterprise). Consumer-GPU hardware in two to four racks. Three to five pilot customers. Position Annie as a sovereign AI option alongside NEXTDC's cloud-neutral ecosystem. No dependency on S7 timeline. Pilot scope includes domain-specific tuning for government and enterprise sectors identified with customers. Annie is currently optimised for insurance and fintech; expanding to new domains requires specialist training or customer model injection. **Phase 2: National Expansion (Months 6-18).** Scale Annie across NEXTDC's national footprint: Melbourne (M-series), Perth (P-series), Brisbane (B-series). Target state government customers in each geography. Develop the "sovereign AI choice" narrative: government and enterprise customers can access OpenAI for general AI and Annie for sovereign domain-specific AI, both on NEXTDC infrastructure. Specialist model development continues for new sectors identified through customer engagement. **Phase 3: S7 Integration (Months 18-36).** As S7's first phase comes online (H2 2027), Annie is positioned as a confirmed additional tenant alongside OpenAI. Annie's track record from Phase 1 and Phase 2 demonstrates demand and creditworthiness. Negotiate a long-term capacity commitment at S7 for Annie's growing customer base. Explore co-investment or revenue-sharing arrangements. Specialist roadmap for new domains identified through expanded customer base is in place. **The commercial structure:** Standard colocation lease initially, with a strategic partnership overlay that includes co-marketing, joint sales enablement, and customer referral arrangements. NEXTDC's carrier-neutral, cloud-neutral DNA means they are comfortable hosting multiple AI providers -- Annie does not need to be the only or the biggest, just creditworthy and growing. Specialist training is handled through Evari's professional services, separate from colocation licensing. ### Alignment and Risks **Where interests align perfectly:** * NEXTDC needs tenant diversification beyond OpenAI -- Annie provides it * Both target the same government and enterprise sectors * Annie can deploy immediately in existing facilities, generating revenue before S7 opens * NEXTDC's carrier-neutral model naturally accommodates multiple AI platforms * The sovereign narrative is stronger with both frontier (OpenAI) and sovereign (Annie) AI on the same infrastructure * NEXTDC's not-yet-profitable status means they are motivated by near-term revenue from real tenants **Where friction could arise:** * NEXTDC may view Annie as too small to be strategically significant at their scale (667MW contracted, 544MW forward order book vs. Annie's initial requirement of perhaps 0.5MW) * OpenAI could object to NEXTDC actively promoting a competing AI platform, even if the MoU is non-exclusive * NEXTDC's board may question Evari's credit profile and funding runway -- the "bad AI tenant" concern * NEXTDC might prefer to wait for larger, more established sovereign AI platforms (Mistral, DeepSeek) to approach them * The S7 timeline (H2 2027) may not align with Annie's urgency to deploy now **Mitigation:** Start in existing facilities, not S7. This removes the timeline mismatch and allows Annie to demonstrate demand before asking for a strategic commitment. Position Annie explicitly as complementary to OpenAI -- "we serve the sectors OpenAI structurally cannot." Address the credit concern head-on by presenting signed pilot customer commitments and a funded runway. Accept that Annie will be a small tenant initially and build the relationship through demonstrated growth rather than day-one scale. *** ## Pilot Customer Strategy The three-sided model only works if all three sides are present. Data centre partners provide infrastructure. Annie provides the platform. But enterprise customers provide the revenue that makes the model financially real. Without signed pilot customers, the data centre partnership is a pitch. With them, it is a business. *Current status: No pilot customers are signed or named yet. Identifying and securing at least one named pilot customer is the highest-priority pre-requisite before data centre partner conversations advance beyond the initial introduction stage.* ```mermaid theme={null} graph LR subgraph "Pilot-to-Scale Pipeline" direction LR ID["Identify
Target organisations
with sovereign AI need
and budget authority"] QU["Qualify
Confirm data sovereignty
requirement, budget,
decision timeline"] PI["Pilot
3-6 month deployment
Limited users, defined
use case, measurable KPIs"] EX["Expand
Additional use cases
More users
Department-wide"] SC["Scale
Enterprise-wide
Multi-year commitment
Additional workloads"] ID --> QU --> PI --> EX --> SC end subgraph "Revenue Impact" direction TB R1["Pilot: $50K-$150K"] R2["Expand: $200K-$500K"] R3["Scale: $500K-$2M+"] end PI --- R1 EX --- R2 SC --- R3 style PI fill:#2563eb,stroke:#1e40af,color:#fff style EX fill:#059669,stroke:#047857,color:#fff style SC fill:#d97706,stroke:#b45309,color:#fff ``` ### Government Agencies **Why they care:** The Fable 5 suspension was a concrete demonstration of dependency risk. DTA's AI procurement guidance (December 2025) creates the framework. GovAI Chat (trials from April 2026) establishes the baseline. Agencies need specialised AI capabilities beyond what GovAI provides -- domain-specific reasoning, policy interpretation, compliance checking -- that a sovereign platform can deliver without US dependency. **Priority targets:** * **Services Australia** -- Already the most advanced agency in operational AI deployment (chatbot "Sam," myGov virtual assistant). Existing CDC customer (\$226M contract). The logical first pilot because they have proven AI appetite and a relationship with CDC. * **Australian Taxation Office (ATO)** -- Massive-scale data processing, fraud detection, compliance automation. Existing Macquarie customer. Budget and procurement sophistication to run a pilot. * **Department of Home Affairs** -- Immigration, border security, national security use cases. Recently consolidated a \$55M 10-year CDC contract. * **Digital Transformation Agency (DTA)** -- Owns the policy framework and GovAI. Positioning Annie as a complementary specialised platform to GovAI could secure the gatekeeper's endorsement. **Estimated pilot scope:** 20-50 users within a single business unit. Defined use case (e.g., policy interpretation, claims processing assistance, compliance checking). Three to six months. \$50K-\$150K pilot fee. **How they procure:** Digital Sourcing Framework. AI Impact Assessment required (mandatory from June 2026, full compliance by December 2026). Can use existing panel arrangements (Data Centre Facilities 2 Whole of Government Panel) or direct approach for innovative technology. Multidisciplinary procurement teams required. Contract clauses must address data sovereignty, model transparency, and audit rights. ### Defence **Why they care:** \$1.2B allocated for sovereign AI in 2025-26. ASCA's \$3.4B over a decade with \~40% directed at AI. Four defined sovereign AI focus areas: signals intelligence, autonomous maritime surveillance, electronic warfare, logistics optimisation. Australia currently lacks a dedicated defence AI testing range. The clearance and sovereignty requirements mean US-hosted AI is structurally unsuitable for classified work. **Priority targets:** * **Defence Innovation Hub** -- 80+ AI-related projects funded since 2024 expansion. Grant range from \$500K (early feasibility) to \$45M (advanced prototype). This is the entry point for defence work. * **Advanced Strategic Capabilities Accelerator (ASCA)** -- Headquartered at Lot Fourteen, Adelaide. Emerging and Disruptive Technologies Decision Advantage Program: \$40M, 14 new contracts. * **Defence Science and Technology Group (DSTG)** -- Currently employs \~400 AI/data science specialists, targeting 600 by 2028. Potential research collaboration. **Estimated pilot scope:** Classified domain-specific AI assistant for a defined intelligence or logistics function. Requires DISP membership, IRAP assessment, and deployment on Certified Strategic infrastructure. Six to twelve months. \$200K-\$500K through Defence Innovation Hub funding. **How they procure:** Defence Innovation Hub for early-stage; ASCA for larger programmes. DISP membership required. IRAP assessment for IT systems handling classified information. Longer procurement cycles (12-18 months typical) but larger contract values. ### Financial Services **Why they care:** APRA prudential requirements on operational resilience and data sovereignty. Board-level urgency after Fable 5 -- major banks were using Claude for code review, document analysis, and customer service automation. Compliance requirements around model transparency and audit trails that US-hosted AI cannot fully satisfy. **Priority targets:** * **Major banks (CBA, NAB, ANZ, Westpac)** -- All four have AI programmes. CBA is a named OpenAI Australia customer. The pitch to the others: "CBA is locked into a US provider. You can choose sovereign AI with full data control and audit trail." * **Major insurers** -- Domain-specific AI for claims processing, underwriting, compliance. Annie's insurance specialist models are a natural fit. **Estimated pilot scope:** AI assistant for a specific business function (compliance checking, document analysis, claims triage). 50-100 users. Three to six months. \$100K-\$300K pilot fee. **How they procure:** Enterprise procurement with APRA compliance review. Vendor risk assessment, data sovereignty certification, model governance documentation. Three to six month procurement cycle for pilots. ### Healthcare **Why they care:** Patient data sovereignty under the Privacy Act and My Health Records Act. Medicare and PBS claims processing at massive scale (Services Australia handles this). Clinical decision support requires auditability and explainability. **Priority targets:** * **Department of Health and Aged Care** -- Policy interpretation, Medicare claims processing support. * **State health departments** -- Clinical coding, administrative AI assistance. * **Private health insurers** -- Claims processing, provider management. **Estimated pilot scope:** AI assistant for clinical coding or claims processing. 20-50 users. Three to six months. \$50K-\$150K pilot fee. **How they procure:** Similar to government procurement frameworks. My Health Records Act compliance requirements. Ethics committee review for clinical applications. *** ## Compliance and Certification Roadmap Annie does not currently hold formal security or compliance certifications. This is expected for an early-stage platform and is not unusual for pilot-phase engagements. However, any path from pilot to production in government, defence, financial services, or healthcare will require certifications. The following are anticipated requirements: * **IRAP assessment** for government deployments (required for Protected-level workloads) * **ISO 27001** as baseline information security certification * **SOC 2 Type II** for enterprise and financial services customers * **DISP membership** if defence deployments are pursued * **APRA CPS 230/234 alignment** for financial services and insurance customers Certification scoping and timeline will be determined during the pilot phase, informed by which customer sectors are engaged first. The partner's existing certification framework provides the infrastructure baseline; Annie's certification covers the application and data handling layer. *** ## The Ask This section specifies what Evari wants from each partner. Vague asks get vague responses. Concrete asks get concrete answers. ### From CDC | Item | Specifics | | -------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Compute allocation** | Two racks at a Canberra facility (Hume or Fyshwick) for pilot. Eight inference GPUs with 32GB+ VRAM (e.g., RTX 5090, L40S, A6000). Standard power density (\~10kW per rack). | | **Commercial terms** | Reduced colocation rate for a 6-month pilot period. Transition to standard commercial rates upon expansion. | | **Partnership duration** | 6-month pilot with option to extend to 3-year strategic partnership. | | **Customer introductions** | Introduction to Services Australia, ATO, and Department of Home Affairs procurement teams. Joint presentation at one government customer event. | | **Perth campus** | Letter of intent for early tenancy at Maddington when operational. Preferred terms as an anchor AI platform partner. | | **Investment** | Not requested in Phase 1. Discussed in Phase 2 if pilot metrics justify it. Potential \$2M-\$5M strategic investment. | | **What Evari commits** | Three signed pilot customers within 6 months. Technical integration with CDC's operational monitoring. Joint case study upon pilot completion. Sovereign deployment with no external dependencies. Annie is currently domain-specialised for insurance and fintech. Pilot scope includes training new specialists (or fine-tuning existing ones) for government-specific use cases identified with customers. Timeline and investment for specialist development are scoped as part of pilot agreements. | ### From Macquarie | Item | Specifics | | -------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Compute allocation** | Two racks at the Macquarie Park campus (IC3 or IC3 Super West upon opening). Eight inference GPUs with 32GB+ VRAM. Liquid cooling not required for Phase 1. | | **Commercial terms** | Launch partner pricing for IC3 Super West. Revenue sharing on customers jointly acquired. | | **Partnership duration** | 12-month strategic pilot with path to a 3-year partnership. | | **Customer introductions** | Introduction to three to five Federal Government agencies from Macquarie's 42% base. Joint briefing to Macquarie Government's cyber security customers. | | **Dell/NVIDIA ecosystem** | Technical engagement with Dell AI Factory engineering team for integration planning. | | **NRF alignment** | Joint proposal positioning Annie as an outcome of the NRF investment in AI-enabled capability. | | **Investment** | Strategic investment of \$1M-\$3M as part of IC3 Super West launch partnership. | | **What Evari commits** | Platform ready for IC3 Super West opening (September 2026). Three signed pilot customers. Integration with Macquarie's sovereign cloud services. Joint marketing for IC3 Super West launch. Annie is currently domain-specialised for insurance and fintech. Pilot scope includes training new specialists for government and enterprise customer target sectors. Specialist development and domain tuning are included as part of the launch partnership. | ### From NEXTDC | Item | Specifics | | ---------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | **Compute allocation** | Two racks at C1 (Canberra) and two racks at one S-series facility (Sydney). Eight inference GPUs with 32GB+ VRAM per site. | | **Commercial terms** | Standard colocation rates with a 6-month pilot discount. No long-term commitment required initially. | | **Partnership duration** | 6-month pilot, expanding to national deployment over 18 months. | | **Customer introductions** | Introduction to NEXTDC's ecosystem partners in government and financial services. Joint presentation at one NEXTDC customer event. | | **Sovereign AI positioning** | Inclusion in NEXTDC's sovereign AI narrative as the Australian-built complement to OpenAI. Joint marketing that positions NEXTDC as offering both frontier and sovereign AI options. | | **S7 consideration** | Early discussion on Annie as a confirmed tenant for S7 Phase 1 (H2 2027), conditional on pilot success. | | **Investment** | Not requested. NEXTDC's capital is committed to infrastructure build-out. Relationship is commercial, not equity. | | **What Evari commits** | Five signed pilot customers within 6 months across two NEXTDC sites. Growth to ten or more customers within 12 months. Committed power draw increase as customer base scales. Joint case study and marketing upon pilot completion. Annie is currently domain-specialised for insurance and fintech. Pilot scope includes training new specialists for government and enterprise customer sectors. Specialist development timelines and investment are scoped with each customer. Annie does not automatically apply to all domains; new domains require specialist training or customer model contribution. | *** ## Timeline ```mermaid theme={null} gantt title Annie Data Centre Partnership Timeline dateFormat YYYY-MM-DD axisFormat %b %Y section Macquarie (Warm Intro) Initial conversation via connection :mac1, 2026-07-01, 2026-07-15 Strategic briefing with David Hirst :mac2, 2026-07-15, 2026-08-01 Technical integration planning :mac3, 2026-08-01, 2026-09-01 IC3 Super West launch deployment :mac4, 2026-09-01, 2026-09-30 Pilot customers onboarded :mac5, 2026-10-01, 2026-12-31 Joint go-to-market :mac6, 2027-01-01, 2027-06-30 section CDC Approach Dr Jack Dan (CSO) :cdc1, 2026-07-15, 2026-08-01 Partnership proposal submitted :cdc2, 2026-08-01, 2026-09-01 Pilot agreement signed :cdc3, 2026-09-01, 2026-10-01 Canberra deployment live :cdc4, 2026-10-01, 2026-11-01 Pilot customers onboarded :cdc5, 2026-11-01, 2027-02-28 Perth campus LOI :cdc6, 2027-01-01, 2027-03-31 Perth deployment :cdc7, 2027-04-01, 2027-09-30 section NEXTDC Approach Craig Scroggie (CEO) :nxt1, 2026-08-01, 2026-09-01 Pilot proposal submitted :nxt2, 2026-09-01, 2026-10-01 C1 Canberra deployment :nxt3, 2026-10-01, 2026-11-15 Sydney deployment :nxt4, 2026-11-15, 2027-01-01 Pilot customers onboarded :nxt5, 2027-01-01, 2027-04-30 National expansion discussion :nxt6, 2027-04-01, 2027-06-30 S7 tenancy negotiation :nxt7, 2027-06-01, 2027-09-30 section Enterprise Pilots First 3 government pilot customers signed :ent1, 2026-10-01, 2026-12-31 First financial services pilot :ent2, 2027-01-01, 2027-03-31 Defence Innovation Hub application :ent3, 2027-01-01, 2027-06-30 10 paying customers milestone :milestone, ent4, 2027-06-30, 0d ``` ### Sequencing Rationale **Macquarie first (July 2026).** The warm introduction makes this the lowest-friction starting point. IC3 Super West opens in September, creating natural urgency. Success here provides proof points for CDC and NEXTDC conversations. **CDC second (mid-July 2026).** CDC's Canberra presence and government customer base make it the highest-value partnership for government pilots. Approaching two to three weeks after Macquarie allows the initial conversation to reference the Macquarie engagement (without oversharing) as evidence of market interest. **NEXTDC third (August 2026).** NEXTDC is the longest sales cycle (larger company, more stakeholders, public company governance). Starting later allows Annie to bring early traction data from Macquarie and CDC conversations. NEXTDC's S7 timeline (H2 2027) means there is less urgency but more strategic value in establishing the relationship now. **Enterprise pilots in parallel (October 2026 onward).** Pilot customer conversations start in July but deployments begin once data centre infrastructure is in place. The goal is three signed pilot customers by end of 2026, ten by mid-2027. ```mermaid theme={null} graph TB subgraph "Partnership Sequencing Logic" direction TB M["Macquarie
FIRST: Warm intro
IC3 Super West urgency
July 2026"] C["CDC
SECOND: Highest gov value
Perth greenfield
Mid-July 2026"] N["NEXTDC
THIRD: Longest cycle
Bring traction data
August 2026"] M -->|"Proof points
from early
engagement"| C M -->|"Traction
evidence"| N C -->|"Gov pilot
results"| N end style M fill:#2563eb,stroke:#1e40af,color:#fff style C fill:#059669,stroke:#047857,color:#fff style N fill:#d97706,stroke:#b45309,color:#fff ``` ### Key Milestones | Date | Milestone | Success Criteria | | ------------------ | ----------------------------------- | ---------------------------------------------------------------------- | | **July 2026** | Macquarie strategic conversation | Meeting with David Hirst secured. Interest confirmed. | | **August 2026** | CDC proposal submitted | Dr Jack Dan engaged. Proposal under review. | | **September 2026** | Macquarie IC3 Super West deployment | Annie live on IC3 Super West at launch. | | **October 2026** | CDC Canberra deployment | Annie live on CDC Canberra facility. | | **December 2026** | Three pilot customers signed | Three paying enterprise customers across partner sites. | | **March 2027** | NEXTDC pilot live | Annie live on two NEXTDC sites. | | **June 2027** | Ten pilot customers | Ten paying customers, \$500K+ ARR. | | **September 2027** | Strategic partnerships formalised | At least one partnership graduated from pilot to multi-year agreement. | *** ## Sources ### Strategy Documents * [SOTA Landscape](annie/sota-landscape.mdx) -- Fable 5 precedent, sovereign AI imperative, open-weight ecosystem viability * [Annie Architecture](annie/annie-architecture.mdx) -- Hierarchical MoE design, Bellerophon BStream backbone, hardware requirements, consensus verification pipeline ### CDC Data Centres * [CDC secures AUD \$17B valuation](https://telconews.com.au/story/cdc-secures-aud-17-billion-valuation-in-ownership-deal) * [Infratil CDC valuation March 2026](https://infratil.com/news/cdc-independent-valuation-31-march-2026/) * [CDC Government sector](https://cdc.com/sectors/government/) * [First three Certified Strategic providers](https://www.itnews.com.au/news/govt-certifies-first-three-strategic-data-centre-providers-565631) * [Services Australia \$226M extension](https://www.datacenterdynamics.com/en/news/services-australia-signs-au226m-extension-with-canberra-data-centres/) * [CDC AI sector](https://cdc.com/sectors/artificial-intelligence/) * [Firmus/CDC/NVIDIA partnership](https://cdc.com/resources/news/driving-australia-s-ai-future-with-cdc-firmus-and-nvidia/) * [Perth AI campus](https://cdc.com/resources/news/cdc-announces-landmark-plan-to-drive-ai-and-advanced-technology-deployments-in-western-australia/) * [CDC Marsden Park NSW approval](https://www.nsw.gov.au/ministerial-releases/southern-hemispheres-biggest-data-centre-gets-green-light) * [CDC 555MW contract](https://cdc.com/resources/news/cdc-signs-555mw-data-centre-contract-with-us-customer/) ### Macquarie Data Centres * [Macquarie Data Centres portfolio](https://www.macquariedatacentres.com/data-centres/) * [IC3 Super West](https://www.macquariedatacentres.com/data-centres/sydney/macquarie-park-campus/ic3-super-west/) * [IC3 Super West topping-out](https://www.macquariedatacentres.com/blog/treasurer-tops-out-macquarie-data-centres-newest-47mw-ai-and-cloud-data-centre/) * [Dell Sovereign AI Factory partnership](https://www.macquariedatacentres.com/blog/mdc-and-dell-technologies-bring-sovereign-ai-factories-to-australia/) * [Macquarie certifications](https://macquariedatacentres.com/why-us/compliance-certifications/) * [NRFC \$200M investment](https://www.nrf.gov.au/news-and-media-releases/national-reconstruction-fund-invests-macquarie-technology-group-strengthen-australias-sovereign-cloud-and-cybersecurity-capabilities) * [Macquarie Technology Group Wikipedia](https://en.wikipedia.org/wiki/Macquarie_Technology_Group) * [Macquarie University partnership](https://itbrief.com.au/story/macquarie-university-tech-group-deepen-digital-ties) ### NEXTDC * [NEXTDC-OpenAI sovereign AI infrastructure](https://www.nextdc.com/news/building-the-next-generation-of-sovereign-ai-infrastructure-in-australia) * [OpenAI for Australia](https://openai.com/global-affairs/openai-for-australia/) * [NEXTDC S7 Sydney](https://www.nextdc.com/data-centres/sydney-data-centres/s7-sydney) * [NEXTDC 1H26 results](https://www.nextdc.com/news/asx-release-1h26-record-results) * [NEXTDC \$8.4B liquidity](https://www.nextdc.com/news/nextdc-bolsters-liquidity-to-a8.4-billion-to-accelerate-ai-infrastructure-rollout) * [NEXTDC leadership](https://www.nextdc.com/about-us/our-leadership) * [NEXTDC NVIDIA DGX-Ready certification](https://www.nextdc.com/news/nextdc-secures-certification-in-the-nvidia-dgx-ready-data-center-program) * [NEXTDC seeks JV partner for 850MW](https://www.datacenterdynamics.com/en/news/nextdc-seeks-partner-for-850mw-data-center-jv-in-sydney-australia/) ### Government Procurement and AI Policy * [DTA AI procurement guidance](https://www.dta.gov.au/media-releases/ai-policy-overhauled-new-impact-assessment-tool-and-procurement-guidance) * [GovAI](https://www.govai.gov.au/about) * [APS AI Plan](https://www.finance.gov.au/about-us/news/2025/introducing-aps-ai-plan) * [ASCA](https://www.asca.gov.au/) * [Defence sovereign AI capability](https://fbi.org.au/blog/2026-03-11-australia-defence-tech-sovereign-ai-capability/) * [AI Accelerator funding](https://www.industry.gov.au/news/ai-accelerator-initiative-kicks-funding-industry-led-research) * [Hosting Certification Framework](https://www.hostingcertification.gov.au/framework) ### Industry and Market Analysis * [Data Centre Cost Guide](https://encoradvisors.com/data-center-cost/) * [Colocation pricing 2026](https://datacenterhawk.com/resources/fundamentals/colocation-data-center-pricing-a-2026-beginner-s-guide) * [AI data centre market report](https://www.marketsandmarkets.com/Market-Reports/ai-data-center-market-267395404.html) * [Data centre JV structures](https://www.bclplaw.com/en-US/events-insights-news/the-rise-of-strategic-joint-ventures-and-alternative-structures-in-data-centre-investment.html) * [Data Centres Australia peak body](https://datacentres.org.au/introducing-data-centres-australia/) * [Australia colocation portfolio analysis 2025](https://www.businesswire.com/news/home/20251205007956/en/) # Sota landscape Source: https://docs.quiva.ai/annie/sota-landscape # SOTA AI Model Landscape -- June 2026 ## Executive Summary The AI model landscape in June 2026 is defined by three converging forces: frontier models have reached genuinely dangerous capability levels, the United States has demonstrated willingness to unilaterally disable access to those models worldwide, and the open-weight ecosystem has matured to the point where sovereign alternatives are technically viable. On June 12, 2026, the US Commerce Department ordered Anthropic to suspend global access to its most capable models -- Fable 5 and Mythos 5 -- three days after launch. Because nationality-based filtering proved technically infeasible, Anthropic disabled both models for all users worldwide, including paying enterprise customers. As of June 22, 2026, they remain suspended with no restoration date. This is the first time a commercially deployed frontier AI model has been forcibly recalled by government order. The incident transformed "sovereign AI" from a policy talking point into an operational imperative. Every organisation running critical workloads on US-hosted frontier models now faces a demonstrated risk: a single government directive can sever access without warning, without recourse, and without geographic exemption. Meanwhile, the open-weight ecosystem offers a credible alternative. Models like DeepSeek V4-Pro (MIT license, 80.6% SWE-Bench Verified), Qwen 3.6 (Apache 2.0, runs on a single consumer GPU), and Mistral Large 3 (Apache 2.0, European sovereign infrastructure) deliver performance that would have been frontier-class twelve months ago, under licenses that permit unrestricted sovereign deployment. Published open-weight models are currently exempt from US export controls. The architectural landscape has also diversified. Monolithic scaling continues at the frontier, but Mixture-of-Experts architectures now dominate (used by Anthropic, OpenAI, Google, DeepSeek, Mistral, Qwen, and Meta), reasoning chains add inference-time compute for hard problems, and multi-model ensemble systems offer a path to frontier-competitive performance at a fraction of the cost. For organisations willing to invest in orchestration rather than raw scale, the gap between "what you can build yourself" and "what the frontier offers" is narrower than it has ever been. *** ## The Export Control Watershed ### Timeline | Date | Event | | ----------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **June 9, 2026** | Anthropic launches Fable 5 and Mythos 5 globally. Fable 5 is the commercial product; Mythos 5 is the unrestricted variant for approved cybersecurity and government partners. Same weights, different safety layers. | | **June 11** | After criticism from cybersecurity researchers that silent rerouting to Opus 4.8 was blocking legitimate defensive work, Anthropic makes the safety fallback visible. | | **June 12, 5:21 PM ET** | US Commerce Department's Bureau of Industry and Security (BIS), under Secretary Howard Lutnick, issues directive to suspend all access for foreign nationals. | | **June 13** | Anthropic disables both models for ALL users worldwide. Services removed from AWS Bedrock, Google Cloud, Microsoft Foundry, Snowflake, Box, and direct APIs. | | **June 17** | G7 summit in Evian-les-Bains. AI executives meet with G7 heads of state. France announces Western democracies will establish a coordinated AI cooperation platform within one month. | | **June 18** | Proposed UK exemption collapses. US House members demand answers from the administration. | | **June 22** | Both models remain suspended for all users worldwide. No restoration date published. | ### The Stated Trigger The Commerce Department cited a jailbreak technique that could cause Fable 5 to exhibit Mythos 5's cybersecurity analysis capabilities -- the kind of vulnerability discovery reasoning that could accelerate offensive cyber operations. Anthropic maintained the vulnerabilities were "known in advance" and "relatively minor in severity," and that similar capabilities exist in GPT-5.5. ### The Broader Context This did not emerge from nothing. In February 2026, President Trump directed all federal agencies to cease using Anthropic after the company refused to waive contractual restrictions on Claude's use for mass domestic surveillance and fully autonomous weapons. Defense Secretary Hegseth designated Anthropic a "supply chain risk" -- the first time this designation was applied to an American company. ### Global Reaction **France**: Bruno Retailleau called it a "wake-up call." Benjamin Haddad characterised it as "an accelerator of the geopolitical battle over AI." Jordan Bardella urged accelerated government support for Mistral AI. **United Kingdom**: Al Carns stated "This isn't an AI story. It's the story of every industry we used to lead." The UK's proposed exemption from the directive collapsed. **Netherlands**: Geert Wilders called for accelerating domestic AI model development: "AI is more and more national sovereignty." **EU**: The European Commission had already proposed the Cloud and AI Development Act on June 3 (pre-suspension), with goals to triple European data centre capacity over 5-7 years. The suspension dramatically accelerated political support. **Australia**: No formal government statement, but the incident has strengthened the sovereign AI debate domestically. Kate Carruthers (UNSW) wrote that the incident "makes sovereign AI real." SmartCompany reported that access to advanced AI capabilities now depends on "export controls, nationality, and geopolitical considerations rather than just commercial decisions." ### What This Means The Fable 5 suspension establishes three precedents: 1. **The US government will act unilaterally** against specific model deployments when it perceives a national security basis. 2. **The practical effect is global**, regardless of the targeted users' nationality or location, because providers cannot technically segregate access in real time. 3. **No exemption exists** for Five Eyes partners, EU allies, or any other country. The UK exemption proposal collapsed. For any non-US organisation running critical workloads on US-hosted frontier models, the risk is no longer theoretical. It has been demonstrated. *** ## Model Landscape ### Anthropic (Fable 5, Mythos 5, Opus 4.8, Sonnet 4.6) #### Architecture Anthropic has not officially disclosed architecture type or parameter counts for any of its models. Third-party analysis strongly suggests Fable 5 / Mythos 5 use a **sparse Mixture-of-Experts (MoE)** architecture optimised for RAG and massive codebases. Anthropic has not confirmed or denied this. Fable 5 and Mythos 5 share **identical weights** -- same training, same base model, same capability ceiling. The only difference is the safety layer: Fable 5 uses a multi-layer content classifier that reroutes high-risk queries to Opus 4.8; Mythos 5 is unrestricted. Mythos 5 is limited to Project Glasswing cybersecurity partners and select US government collaborators. #### Capabilities and Benchmarks | Benchmark | Fable 5 | Opus 4.8 | Sonnet 4.6 | Haiku 4.5 | | --------------------- | -------------------------------- | -------- | ---------- | --------- | | SWE-Bench Verified | 95.0%\* | 88.6% | 72.7-79.6% | 73.3% | | SWE-Bench Pro | 80.3%\* | 69.2% | -- | 39.5% | | FrontierCode Diamond | 29.3% | 13.4% | -- | -- | | Terminal-Bench 2.1 | 88.0% | 82.7% | -- | -- | | GPQA Diamond | -- | 93.6% | \~83% | -- | | MMLU | -- | -- | 91.8% | -- | | Humanity's Last Exam | 59.0% (no tools) / 64.5% (tools) | -- | -- | -- | | ExploitBench (Mythos) | 78.0% | 40.0% | -- | -- | \*The 80.3% SWE-Bench Pro score was produced using Anthropic's own scaffolding, not a neutral evaluation harness. Independent evaluators have contested this figure. Vendor-scaffold numbers consistently run 10-30 points above Scale's standardised leaderboard. Anthropic did not publish MMLU or HumanEval scores for Fable 5. #### Context and Output | Model | Context | Max Output | | ------------------ | ---------------- | ----------- | | Fable 5 / Mythos 5 | 1M tokens | 128K tokens | | Opus 4.8 | 1M tokens | 128K tokens | | Sonnet 4.6 | 1M tokens (beta) | 64K tokens | | Haiku 4.5 | 200K tokens | 64K tokens | #### Pricing | Model | Input/MTok | Output/MTok | Cache Hit | Batch (In/Out) | | ------------------ | ---------- | ----------- | --------- | ---------------- | | Fable 5 / Mythos 5 | \$10.00 | \$50.00 | \$1.00 | \$5.00 / \$25.00 | | Opus 4.8 | \$5.00 | \$25.00 | \$0.50 | \$2.50 / \$12.50 | | Sonnet 4.6 | \$3.00 | \$15.00 | \$0.30 | \$1.50 / \$7.50 | | Haiku 4.5 | \$1.00 | \$5.00 | \$0.10 | \$0.50 / \$2.50 | #### Deployment Model and Sovereign Limitations Anthropic operates **API-only** through its own infrastructure, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. There is **no on-premise or self-hosted option**. All data flows through US-based infrastructure. The June 12 export control incident demonstrated the operational consequence: the US government effectively exercised a kill switch over global access, and Anthropic had no technical means to maintain service for non-US customers even if it wanted to. #### Compute and Financials Anthropic is described as a "highly capital-intensive, quasi-infrastructure entity" rather than an asset-light SaaS business. Committed compute partnerships exceed \$330 billion (Amazon >\$100B over 10 years, Google \~\$200B over 5 years, Microsoft \$30B). Projected 2026 losses: approximately \$29 billion against \$25-30 billion in revenue, with 65-80% consumed by compute costs. Peak training spend estimated at \~\$30 billion in the 2028 timeframe. *** ### OpenAI (Codex, GPT Series, o-Series) #### Architecture OpenAI uses a **Mixture-of-Experts (MoE)** architecture for the GPT-5.x family. Exact parameter counts are not disclosed. Estimates suggest active parameters in the 2-5 trillion range with a total expert pool potentially 10-50+ trillion. The widely circulated 52.5 trillion figure represents total parameter capacity, not active parameters per inference. GPT-5.5 (codenamed "Spud," released April 23, 2026) is the first fully retrained base model since GPT-4.5. Every model from GPT-5.0 through GPT-5.4 was an incremental post-training iteration on the same foundation; 5.5 is a ground-up rebuild. #### Codex Platform Codex is now OpenAI's agentic coding platform, not a standalone model. It runs across four surfaces: the Codex app (desktop), Codex CLI (terminal agent), IDE extensions, and Codex Cloud (web). The underlying models are GPT-5.x variants. Current capabilities include computer use, Record and Replay workflow automation, PR review, multi-file terminal view, in-app browser, and SSH to remote devboxes. #### Key Benchmarks | Benchmark | GPT-5.5 | GPT-5.4 | Notes | | ------------------------------ | ------- | ------- | ----------------- | | SWE-Bench Verified | 88.7% | 74.9% | | | SWE-Bench Pro | 58.6% | 57.7% | | | Terminal-Bench 2.0 | 82.7% | 75.1% | | | MMLU | 92.4% | -- | | | GPQA Diamond | 93.6% | 92.8% | | | ARC-AGI-2 | 85.0% | 73.3% | | | FrontierMath T1-3 | 51.7% | 47.6% | | | Long-context 512K-1M (MRCR v2) | 74.0% | 36.6% | Major improvement | **Hallucination caveat**: GPT-5.5 scores highest on factual recall (57% accuracy on AA-Omniscience) but has an **86% hallucination rate** on that benchmark vs. Claude Opus 4.7's 36%. It confabulates more aggressively at knowledge boundaries. #### Reasoning Models (o-Series) | Model | Input/MTok | Output/MTok | Context | Key Score | | ------- | ---------- | ----------- | ------- | ---------------------------------- | | o4-mini | \$1.10 | \$4.40 | 200K | AIME 2025: 92.7%, SWE-Bench: 68.1% | | o3 | \$2.00 | \$8.00 | 200K | Codeforces SOTA, MMMU leader | | o3-pro | \$20.00 | \$80.00 | 200K | AIME 2025: 98%, GPQA Diamond: 86% | The o-series models add explicit reasoning chains (inference-time compute) for harder problems. o3-pro targets the hardest 5% of problems: PhD-level science, competitive maths, complex formal reasoning. #### Pricing | Model | Input/MTok | Output/MTok | Cached Input | Batch (In/Out) | | ------------ | ---------- | ----------- | ------------ | ----------------- | | GPT-5.5 | \$5.00 | \$30.00 | \$0.50 | \$2.50 / \$15.00 | | GPT-5.5 Pro | \$30.00 | \$180.00 | -- | \$15.00 / \$90.00 | | GPT-5.4 | \$2.50 | \$15.00 | \$0.25 | \$1.25 / \$7.50 | | GPT-5.4 mini | \$0.75 | \$4.50 | \$0.075 | \$0.375 / \$2.25 | | GPT-5.4 nano | \$0.20 | \$1.25 | \$0.02 | \$0.10 / \$0.625 | | GPT-4.1 | \$2.00 | \$8.00 | \$0.50 | \$1.00 / \$4.00 | | GPT-4.1 mini | \$0.40 | \$1.60 | \$0.10 | \$0.20 / \$0.80 | | GPT-4.1 nano | \$0.10 | \$0.40 | \$0.025 | \$0.05 / \$0.20 | #### Open-Weight Models OpenAI has released limited open-weight reasoning models under Apache 2.0: | Model | Parameters | License | Purpose | | ---------------------- | ---------- | ---------- | --------------------- | | gpt-oss-120b | 120B | Apache 2.0 | General reasoning | | gpt-oss-20b | 20B | Apache 2.0 | Lightweight reasoning | | gpt-oss-safeguard-120b | 120B | Apache 2.0 | Safety classification | | gpt-oss-safeguard-20b | 20B | Apache 2.0 | Safety classification | All flagship models (GPT-5.x, o-series) remain closed-weight. #### Deployment Model and Sovereign Position OpenAI operates via its own API and Azure OpenAI Service. The Microsoft exclusivity arrangement was removed in April 2026 -- OpenAI can now partner with other cloud providers. Azure Sovereign Cloud / Azure Local offers on-premises control planes for government and defence workloads. The NEXTDC partnership ("OpenAI for Australia") involves an AUD \$7+ billion hyperscale AI campus at Eastern Creek, Sydney (S7, 650MW total campus capacity, with OpenAI as initial offtaker at approximately 550MW). Phase 1 is expected H2 2027. However, this is OpenAI sovereign compute infrastructure, not customer-controlled infrastructure -- the distinction matters. OpenAI has so far avoided direct export control restrictions. However, industry expectation is that export control obligations will extend across multiple providers over the next 12-24 months as models exceed capability thresholds. #### Financials Approximately \$25 billion annualised revenue, approximately 900 million weekly users, projected \$14 billion loss in 2026 (inference costs dominate). Training run estimates for frontier models: \$500M+ per run. Stargate Abilene cluster coming online in phases. *** ### Google (Gemini Family) #### Architecture All Gemini models from 2.5 onward use a **sparse Mixture-of-Experts (MoE)** architecture built on a dense Transformer backbone. The Gemini 2.5 Pro technical report (the only one with confirmed architecture details) describes: approximately 200 billion total parameters, decoder-only transformer, 80 layers, 16,384 hidden dimensions, 128 self-attention heads, MoE layers every other block with 64 experts per block and 8 active per token (approximately 12.5% of parameters active per inference). This yields roughly 1.6x compute/capacity efficiency over purely dense models. Gemini 3.x adds a "DeepThink System 2" deliberation layer with three-tier reasoning control (Low/Medium/High). Parameter counts for the 3.x family are not disclosed. Google trains entirely on custom TPU hardware (v5e, v6e Trillium) with no NVIDIA GPU fallback for Gemini models. The newly announced TPU 8t delivers 121 exaflops per superpod with 9,600 chips. #### Current Model Lineup | Model | Release | Context | Input/MTok | Output/MTok | Key Benchmark | | --------------------- | ----------------- | ------- | ------------- | --------------- | ----------------------------- | | Gemini 3.5 Flash | May 2026 | 1M | \$1.50 | \$9.00 | Terminal-Bench 2.1: 76.2% | | Gemini 3.5 Pro | Previewed, not GA | -- | -- | -- | -- | | Gemini 3.1 Pro | Feb 2026 | 1M | \$2.00/\$4.00 | \$12.00/\$18.00 | SWE-Bench: 80.6%, GPQA: 94.3% | | Gemini 3.1 Flash-Lite | -- | -- | \$0.25 | \$1.50 | Budget frontier | | Gemini 2.5 Pro | GA | 1M | \$1.25/\$2.50 | \$10.00/\$15.00 | SWE-Bench: 63.8% | | Gemini 2.5 Flash | GA | 1M | \$0.30 | \$2.50 | -- | | Gemini 2.5 Flash-Lite | GA | 1M | \$0.10 | \$0.40 | Cheapest | Pricing tiers with "/" indicate \<=200K / >200K context pricing. Batch mode runs at 50% of standard pricing across all models. #### Key Benchmarks (Gemini 3.1 Pro) | Benchmark | Score | | --------------------------- | ----------------------------------- | | SWE-Bench Verified | 80.6% | | GPQA Diamond | 94.3% (highest at launch) | | ARC-AGI-2 | 77.1% | | Humanity's Last Exam | 44.4% (text + multimodal, no tools) | | MRCR v2 (128K long-context) | 84.9% | | MMMU-Pro | 80.5% | | MMMLU | 92.6% | #### Specialised Models Google maintains the broadest multimodal portfolio: Veo 3.1 (video generation up to 4K), Lyria 3 (music generation), Imagen 4 (image generation, being replaced by Gemini-native generation), Gemini Computer Use Preview, Gemini Robotics-ER 1.6, real-time audio dialogue, real-time translation, and Deep Research. #### Deployment Model and Sovereign Position Google offers the most mature sovereign deployment story among the three major closed-model providers: * **Gemini Developer API and Vertex AI**: Standard cloud access with enterprise SLAs. * **Google Distributed Cloud (GDC)**: Full on-premises AI deployment with managed infrastructure. Gemini models and Gemma open models available. Confidential external key management for regulated organisations. A new "sovereign agentic AI architecture" announced at Cloud Next 2026 ensures agentic workflows execute entirely within customer organisation boundaries. * **Forrester recognition**: Named a Leader in The Forrester Wave Sovereign Cloud Platforms, Q2 2026. No reports of Google models being export-restricted as of June 2026, but the Anthropic precedent means all frontier providers face potential future restrictions. #### Open Models: Gemma 4 | Variant | Parameters | Context | Architecture | VRAM (Q4) | | ------- | --------------------- | ------- | -------------------- | --------- | | E2B | \~2B effective | 128K | Dense multimodal | \~1.5 GB | | E4B | \~4B effective | 128K | Dense multimodal | \~5 GB | | 12B | 12B | 128K+ | Dense (encoder-free) | \~8 GB | | 26B-A4B | 26B total / 4B active | 256K | MoE | \~18 GB | | 31B | 31B dense | 256K | Dense | \~20 GB | Licensed under **Apache 2.0** -- Google's first truly open-source family under this license. Gemma 4 31B posts 89.2% AIME 2026 and 80.0% LiveCodeBench, competitive with some closed frontier models. The E2B model runs on phones. #### Financials Google guided \$175-185 billion in 2026 capex, majority AI-related. Gemini 1.0 Ultra training cost approximately \$191 million (Stanford AI Index / Epoch AI). Gemini 3.x training costs are not disclosed but estimated in the several-hundred-million-dollar range per model. Google's use of custom TPUs significantly reduces marginal compute costs compared to competitors renting NVIDIA GPUs. *** ### Open-Source / Open-Weight Models The open-weight ecosystem is where the sovereign AI opportunity lives. Multiple model families now offer frontier-class performance under permissive licenses, deployable on sovereign infrastructure without any foreign provider dependency. #### Meta Llama 4 | Model | Active Params | Total Params | Experts | Context | | -------- | ------------- | ------------ | ------- | ------------ | | Scout | 17B | 109B | 16 | 10M tokens | | Maverick | 17B | 400B | 128 | 512K tokens | | Behemoth | 288B | \~2T | 16 | Not released | Architecture: MoE with alternating dense and MoE layers. 128 routed experts plus one shared expert per MoE layer; each token activates the shared expert plus one routed expert. **License: Custom (Llama 4 Community License Agreement)**. This is not Apache 2.0 or MIT. The EU is **explicitly excluded** -- rights do not extend to individuals domiciled in, or companies with principal place of business in, the European Union. Companies with >700 million MAU require a separate license. Government agencies must request exceptions case-by-case. This makes Llama 4 unsuitable for sovereign deployment in any EU member state and introduces legal uncertainty elsewhere. Hardware: Scout fits \~61 GB VRAM at Q4\_K\_M (single H100). Maverick needs \~224 GB (4x H100). Behemoth was never publicly released. #### Mistral AI | Model | Total Params | Active Params | Architecture | License | | ----------------------- | ------------ | ------------- | --------------------------- | ---------- | | Mistral Large 3 | 675B | 41B | MoE | Apache 2.0 | | Mistral Small 4 | 119B | 6B | MoE (128 experts, 4 active) | Apache 2.0 | | Ministral 3 (3B/8B/14B) | 3-14B | Dense | Dense | Apache 2.0 | All models Apache 2.0. No MAU thresholds, no geographic exclusions. Mistral is the de facto **European sovereign AI champion**: framework agreement with the French Ministry of Armed Forces (2026-2030), EUR 2.1 billion in state investment, data centres in France with thousands of H100 GPUs, partnership with SAP and French/German governments for sovereign public administration AI. Key benchmarks: Large 3 posts 73.11% MMLU-Pro and 93.60% MATH-500. Ministral 3 14B reasoning variant achieves 85% on AIME 2025. #### Qwen (Alibaba) | Model | Total Params | Active Params | Architecture | Context | License | | ------------------ | ------------ | ------------- | ---------------- | --------------- | ---------- | | Qwen3-235B-A22B | 235B | 22B | MoE | -- | Apache 2.0 | | Qwen3-Coder 480B | 480B | 35B | MoE | -- | Apache 2.0 | | Qwen 3.5-397B-A17B | 397B | 17B | MoE (GDN hybrid) | 262K (ext. 1M+) | Apache 2.0 | | Qwen 3.6-35B-A3B | 35B | 3B | MoE | 1M native | Apache 2.0 | | Qwen 3.6-27B | 27B | 27B | Dense + vision | 1M | Apache 2.0 | | Smaller models | 0.6B-9B | Dense | -- | -- | Apache 2.0 | The broadest size range (0.6B to 480B) under Apache 2.0. Qwen 3.5 introduced Gated Delta Networks (GDN) fused with sparse MoE -- 8.6x faster than Qwen3-Max at 32K context, 19x faster at 256K. 201-language support. Key benchmarks: Qwen3-235B-A22B posts 95.6 ArenaHard and 85.7 AIME'24. Qwen 3.6-35B-A3B runs on a single consumer GPU (\~21 GB at Q4\_K\_M, 30 tok/s) and won coding benchmarks vs Gemma 4 26B-A4B by 21 points. **Geopolitical note**: Chinese origin. Self-hosted deployments involve no data flowing to China and weights are openly inspectable. The Qwen Chat web interface applies Chinese content restrictions, but this is a platform restriction, not a license restriction on the downloadable weights. For nations without anti-China procurement policies, Qwen is arguably the most versatile open model family available. #### DeepSeek | Model | Total Params | Active Params | Architecture | Context | License | | ----------------- | ------------ | ------------- | ------------------- | ------- | ------- | | DeepSeek-V4-Pro | 1.6T | 49B | MoE + MLA + CSA/HCA | 1M | MIT | | DeepSeek-V4-Flash | 284B | 13B | MoE + MLA + CSA/HCA | 1M | MIT | | DeepSeek-R1 | 671B | 37B | MoE + MLA + RL | 128K | MIT | | DeepSeek-V3 | 671B | 37B | MoE + MLA | 128K | MIT | MIT license -- the most permissive possible. No restrictions of any kind. DeepSeek-V3's training cost of approximately \$5.6 million (2,048 H800 GPUs) sent shockwaves through the industry in January 2025, demonstrating frontier-class models could be trained at 10-20x lower cost than assumed. Key benchmarks (V4-Pro): SWE-bench Verified 80.6%, LiveCodeBench Pass\@1 93.5% (highest among all models evaluated), Codeforces rating 3206, MMLU-Pro 87.5%. **Geopolitical note**: Chinese origin. Subject to PRC laws requiring cooperation with intelligence agencies. DeepSeek reportedly used tens of thousands of NVIDIA chips restricted from export to China. Self-hosted weights are safe -- MIT license, no data flows to China, fully inspectable. The hosted API service applies Chinese content regulations and stores data under PRC law. #### Other Notable Open Models | Model | Parameters | Architecture | License | Standout Feature | | -------------------------- | ---------- | ------------------------ | ---------------- | ------------------------------------------------------ | | Gemma 4 (Google) | 2B-31B | Dense/MoE | Apache 2.0 | Runs on phones (E2B), 89.2% AIME 2026 (31B) | | Phi-4 (Microsoft) | 3.8B-15B | Dense | MIT | 93.7% GSM8K at 14B, surpasses many 70B models on maths | | Falcon H1 (TII, Abu Dhabi) | 3B-34B | Hybrid Mamba-Transformer | Apache 2.0-based | Best Arabic LLM, 4x input throughput | | gpt-oss (OpenAI) | 20B/120B | -- | Apache 2.0 | Reasoning-focused, safety classification | | Cohere Command R+ | 104B | -- | CC-BY-NC | RAG-optimised with grounding citations | *** ## Comparison Matrix ### Closed / API-Only Models | Model | Architecture | Params (Total/Active) | Context | SWE-Bench Verified | GPQA Diamond | Input \$/MTok | Output \$/MTok | On-Prem | Export Risk | Fine-Tunable | | --------------------- | ---------------------- | --------------------- | ------- | ------------------ | ------------ | ------------- | --------------- | --------------- | ------------- | ------------ | | Fable 5 | MoE (unconfirmed) | Undisclosed | 1M | 95.0%\* | -- | \$10.00 | \$50.00 | No | **SUSPENDED** | No | | Opus 4.8 | Undisclosed | Undisclosed | 1M | 88.6% | 93.6% | \$5.00 | \$25.00 | No | Medium | No | | Sonnet 4.6 | Undisclosed | Undisclosed | 1M | 72.7-79.6% | \~83% | \$3.00 | \$15.00 | No | Medium | No | | Haiku 4.5 | Undisclosed | Undisclosed | 200K | 73.3% | -- | \$1.00 | \$5.00 | No | Medium | No | | GPT-5.5 | MoE | Est. 2-5T active | 1.05M | 88.7% | 93.6% | \$5.00 | \$30.00 | Via Azure Local | Medium | No | | GPT-5.4 | MoE | Undisclosed | 1M | 74.9% | 92.8% | \$2.50 | \$15.00 | Via Azure Local | Medium | No | | GPT-5.4 mini | MoE | Undisclosed | 400K | -- | -- | \$0.75 | \$4.50 | Via Azure Local | Medium | No | | GPT-5.4 nano | MoE | Undisclosed | -- | -- | -- | \$0.20 | \$1.25 | Via Azure Local | Low | No | | o3 | Reasoning chain | Undisclosed | 200K | -- | -- | \$2.00 | \$8.00 | No | Medium | No | | o3-pro | Reasoning chain | Undisclosed | 200K | -- | 86% | \$20.00 | \$80.00 | No | Medium | No | | o4-mini | Reasoning chain | Undisclosed | 200K | 68.1% | -- | \$1.10 | \$4.40 | No | Medium | No | | Gemini 3.5 Flash | Sparse MoE | Undisclosed | 1M | -- | -- | \$1.50 | \$9.00 | Via GDC | Low | No | | Gemini 3.1 Pro | Sparse MoE + DeepThink | Undisclosed | 1M | 80.6% | 94.3% | \$2.00/\$4.00 | \$12.00/\$18.00 | Via GDC | Low | No | | Gemini 2.5 Pro | Sparse MoE | \~200B total | 1M | 63.8% | -- | \$1.25/\$2.50 | \$10.00/\$15.00 | Via GDC | Low | No | | Gemini 2.5 Flash-Lite | Sparse MoE | Undisclosed | 1M | -- | -- | \$0.10 | \$0.40 | Via GDC | Low | No | \*Vendor-scaffold score; independent evaluations typically run 10-30 points lower. ### Open-Weight Models | Model | Architecture | Params (Total/Active) | Context | SWE-Bench Verified | License | On-Prem | Export Risk | Fine-Tunable | EU Deployable | Min VRAM (Q4) | | ----------------- | ------------------------ | --------------------- | ------- | ------------------ | ---------------- | ------- | ---------------- | ------------ | ------------- | ------------- | | Llama 4 Scout | MoE | 109B / 17B | 10M | -- | Custom | Yes | None (published) | Yes | **NO** | \~61 GB | | Llama 4 Maverick | MoE | 400B / 17B | 512K | -- | Custom | Yes | None (published) | Yes | **NO** | \~224 GB | | Mistral Large 3 | MoE | 675B / 41B | 256K | -- | Apache 2.0 | Yes | None | Yes | Yes | Multi-GPU | | Mistral Small 4 | MoE | 119B / 6B | -- | -- | Apache 2.0 | Yes | None | Yes | Yes | \~25 GB | | Ministral 3 14B | Dense | 14B / 14B | -- | -- | Apache 2.0 | Yes | None | Yes | Yes | \~10 GB | | Qwen3-235B-A22B | MoE | 235B / 22B | -- | -- | Apache 2.0 | Yes | None | Yes | Yes | Multi-GPU | | Qwen 3.6-35B-A3B | MoE | 35B / 3B | 1M | -- | Apache 2.0 | Yes | None | Yes | Yes | \~21 GB | | Qwen 3.6-27B | Dense | 27B / 27B | 1M | -- | Apache 2.0 | Yes | None | Yes | Yes | \~20 GB | | DeepSeek V4-Pro | MoE + MLA | 1.6T / 49B | 1M | 80.6% | MIT | Yes | None (published) | Yes | Yes | \~1 TB+ | | DeepSeek V4-Flash | MoE + MLA | 284B / 13B | 1M | -- | MIT | Yes | None (published) | Yes | Yes | \~80 GB (FP8) | | Gemma 4 31B | Dense | 31B / 31B | 256K | -- | Apache 2.0 | Yes | None | Yes | Yes | \~20 GB | | Gemma 4 12B | Dense | 12B / 12B | 128K | -- | Apache 2.0 | Yes | None | Yes | Yes | \~8 GB | | Gemma 4 E4B | Dense | \~4.5B | 128K | -- | Apache 2.0 | Yes | None | Yes | Yes | \~5 GB | | Gemma 4 E2B | Dense | \~2.3B | 128K | -- | Apache 2.0 | Yes | None | Yes | Yes | \~1.5 GB | | Phi-4 14B | Dense | 14B / 14B | 128K | -- | MIT | Yes | None | Yes | Yes | \~10 GB | | Phi-4-mini | Dense | 3.8B / 3.8B | 128K | -- | MIT | Yes | None | Yes | Yes | \~3 GB | | gpt-oss-120b | -- | 120B | -- | -- | Apache 2.0 | Yes | None | Yes | Yes | Multi-GPU | | gpt-oss-20b | -- | 20B | -- | -- | Apache 2.0 | Yes | None | Yes | Yes | \~15 GB | | Falcon H1 34B | Hybrid Mamba-Transformer | 34B / 34B | -- | -- | Apache 2.0-based | Yes | None | Yes | Yes | \~22 GB | *** ## Architectural Approaches The AI model landscape has diversified beyond "make the model bigger." Four distinct architectural philosophies now compete, each with different implications for cost, capability, and sovereign deployability. ### Monolithic Scaling The original paradigm: train a single dense transformer with as many parameters as possible. GPT-4 (2023) was the high-water mark. By mid-2026, no frontier lab still uses purely dense architectures for their largest models -- the compute cost scales linearly with parameter count, making trillion-parameter dense models economically impractical. Dense architectures remain optimal at smaller scales (Gemma 4 12B, Phi-4, Ministral 3) where every parameter earns its keep. ### Sparse Mixture-of-Experts (MoE) Now the dominant frontier architecture. Used by Anthropic (likely), OpenAI (GPT-5.x), Google (Gemini), DeepSeek, Mistral, Meta (Llama 4), and Qwen. The key insight: a model with 1.6 trillion total parameters but only 49 billion active per inference gets the knowledge capacity of the larger model at the inference cost of the smaller one. DeepSeek V4-Pro exemplifies this: 1.6T total parameters, 49B active, posting frontier-class benchmarks at dramatically lower training and inference costs. The efficiency gain is roughly 1.6-2x over equivalent dense models (per Google's Gemini 2.5 technical report). ### Reasoning Chains (Inference-Time Compute) Pioneered by OpenAI's o-series and adopted by Google (DeepThink System 2) and others. Instead of making the model larger, make it think longer on hard problems. The model generates explicit reasoning steps before producing a final answer, trading latency for accuracy. o3-pro achieves 98% on AIME 2025 through extended reasoning. This approach is additive -- reasoning chains work on top of MoE or dense architectures. The cost model shifts from "pay for parameters" to "pay for thinking time," which is controllable per-query (Google offers Low/Medium/High tiers). ### Multi-Model Orchestration Rather than routing tokens to experts within a single model, route entire tasks to specialist models. Annie's Hierarchical Mixture-of-Experts architecture is one approach: twelve specialist small language models (250M to 27B parameters) orchestrated through a messaging backbone, with classification, expert selection, judgment panels, and verification. This approach trades single-model coherence for composability, cost efficiency, and sovereign deployability (each specialist model can run on consumer hardware). The research supports this: ensembles of smaller models can outperform single large models with both higher accuracy and fewer total FLOPs, and the gap widens as models become large. The limitation is orchestration complexity and latency from multi-hop routing. ### Architectural Comparison ```mermaid theme={null} graph TB subgraph "Monolithic Dense" MD_IN[Input] --> MD_ALL[All Parameters Active
e.g. 14B dense] MD_ALL --> MD_OUT[Output] MD_NOTE["Every token uses every parameter
Simple, efficient at small scale
Cost scales linearly with size"] end subgraph "Sparse MoE (Single Model)" MOE_IN[Input] --> MOE_ROUTER[Router Network] MOE_ROUTER --> MOE_E1[Expert 1] MOE_ROUTER --> MOE_E2[Expert 2] MOE_ROUTER -.-> MOE_E3[Expert 3
inactive] MOE_ROUTER -.-> MOE_EN[Expert N
inactive] MOE_E1 --> MOE_OUT[Output] MOE_E2 --> MOE_OUT MOE_NOTE["1.6T total, 49B active per token
Knowledge of large, cost of small
Dominant frontier architecture"] end subgraph "Reasoning Chains" RC_IN[Input] --> RC_MODEL[Base Model
MoE or Dense] RC_MODEL --> RC_THINK["Thinking Step 1
Thinking Step 2
...
Thinking Step N"] RC_THINK --> RC_ANSWER[Final Answer] RC_NOTE["Same model, more compute per query
Trades latency for accuracy
Controllable per-query budget"] end subgraph "Multi-Model Orchestration" MM_IN[Input] --> MM_CLASS[Classifier] MM_CLASS --> MM_S1[Specialist 1
250M-3B] MM_CLASS --> MM_S2[Specialist 2
7B-14B] MM_CLASS --> MM_S3[Specialist 3
14B-27B] MM_S1 --> MM_JUDGE[Judgment
Panel] MM_S2 --> MM_JUDGE MM_S3 --> MM_JUDGE MM_JUDGE --> MM_VERIFY[Verification] MM_VERIFY --> MM_OUT[Output] MM_NOTE["Many small models, one orchestrator
Each specialist runs on consumer HW
Ensemble accuracy can exceed single large"] end style MD_NOTE fill:#f9f9f9,stroke:#999,stroke-dasharray: 5 5 style MOE_NOTE fill:#f9f9f9,stroke:#999,stroke-dasharray: 5 5 style RC_NOTE fill:#f9f9f9,stroke:#999,stroke-dasharray: 5 5 style MM_NOTE fill:#f9f9f9,stroke:#999,stroke-dasharray: 5 5 style MOE_E3 fill:#eee,stroke:#999,stroke-dasharray: 5 5 style MOE_EN fill:#eee,stroke:#999,stroke-dasharray: 5 5 ``` ### Cost-Capability Tradeoffs | Approach | Training Cost | Inference Cost | Peak Capability | Sovereign Deployability | | ------------------------ | ------------------------------- | ---------------------------------- | ------------------------------------ | --------------------------------------- | | Monolithic Dense (small) | \$2K-\$500K | Lowest per token | Limited by parameter count | Excellent (laptop to single GPU) | | Sparse MoE (large) | \$5M-\$500M+ | Low per token (only active params) | Highest (frontier) | Poor to moderate (datacenter) | | Reasoning Chains | Adds to base model cost | Variable (controllable) | Highest on hard problems | Same as base model | | Multi-Model Ensemble | Sum of specialists (\$10K-\$2M) | Moderate (multiple models) | Approaches frontier on defined tasks | Excellent (consumer hardware per model) | *** ## The Sovereign AI Imperative ### What Sovereignty Means in Practice Sovereign AI is a nation's or organisation's ability to develop, deploy, and control AI using its own infrastructure, data, talent, and governance frameworks without critical dependencies on foreign providers. It spans four dimensions: 1. **Data sovereignty**: Data collected, stored, and processed according to local laws without unauthorised foreign access. 2. **Model sovereignty**: Ownership of model weights, training capability, and inference control. 3. **Compute sovereignty**: Infrastructure under national jurisdiction on domestic soil. 4. **Interaction sovereignty**: Prompts, queries, and outputs remain within sovereign boundaries. Before June 12, 2026, most organisations treated sovereignty as a compliance checkbox. After June 12, it is an operational resilience requirement. ### The Deployment Spectrum ```mermaid theme={null} graph LR subgraph "Level 1: Dependent" L1["API calls to US provider
No control over model or infra
Kill-switch demonstrated"] end subgraph "Level 2: Resident" L2["Workloads in-jurisdiction
Foreign entity operates
e.g. Azure sovereign region"] end subgraph "Level 3: Controlled" L3["Actor holds keys and decisions
Can disconnect without permission
e.g. Open-weight on own infra"] end subgraph "Level 4: Capable" L4["Domestic production capability
Can build, modify, or replace
e.g. Fine-tuned specialists"] end subgraph "Level 5: Self-Determining" L5["Complete autonomy
Can lose every external dependency
Only 3 nations can sustain"] end L1 -->|"Increasing sovereignty"| L2 L2 --> L3 L3 --> L4 L4 --> L5 style L1 fill:#ff6b6b,stroke:#c92a2a,color:#fff style L2 fill:#ffa94d,stroke:#e67700 style L3 fill:#ffd43b,stroke:#f08c00 style L4 fill:#69db7c,stroke:#2b8a3e style L5 fill:#4dabf7,stroke:#1864ab,color:#fff ``` Most practical sovereign AI strategies target Level 3-4 on critical dimensions while accepting Level 2 on others. Targeting Level 5 across all dimensions is economically prohibitive -- only the US, China, and possibly the EU as a bloc can sustain it. ### Hardware Requirements by Sovereignty Tier | Tier | What You Need | Hardware | Approx. Cost | | --------------------------------- | ---------------------------------------- | ---------------------------------- | -------------------------------------------------------------------------- | | **L1: API consumer** | Internet connection | None | \$0.10-\$50/MTok (recurring) | | **L2: Sovereign cloud tenant** | Contract with sovereign cloud provider | Provider-managed | \$50K-\$500K/year | | **L3: Self-hosted open models** | Open-weight models on own infrastructure | 1-8 GPUs per model | \$5K-\$200K hardware + \$105K-\$210K/year electricity (AU rates, per rack) | | **L3+: Fine-tuned specialists** | Domain adaptation of open models | Same as L3 + training compute | Additional \$2K-\$500K per model for fine-tuning | | **L4: Domestic foundation model** | Train from scratch, 1B-7B parameters | 8x RTX 4090 to 64x A100 | \$2K-\$500K per model | | **L4+: Sovereign language model** | National-language foundation model | H100 cluster | \$8M-\$32M (Brazil/Mexico research) | | **L5: Frontier-competitive** | Full-scale foundation model training | Thousands of GPUs, dedicated power | \$100M-\$1B+ per model | ### Cost Comparison: API vs Self-Hosted The break-even depends entirely on utilisation. Bursty, low-volume usage favours APIs. Constant high-throughput workloads favour self-hosting. | Scenario | API Cost | Self-Hosted Cost | Winner | | ------------------------------ | -------------------------- | ------------------------------------------ | ---------------------- | | Light usage (1M tokens/day) | \$3-\$50/day | \$10-\$50/day (amortised hardware + power) | API | | Medium usage (100M tokens/day) | \$300-\$5,000/day | \$50-\$200/day | Self-hosted | | Heavy usage (1B+ tokens/day) | \$3,000-\$50,000/day | \$200-\$500/day | Self-hosted by 10-100x | | Frontier capability required | Only option for some tasks | Open models lag on hardest 5-10% of tasks | API (for now) | Frontier API costs are increasing: GPT-5.5 costs over 3x what GPT-5 cost 8 months ago; Gemini 3.5 Flash tripled pricing versus its predecessor. ### The Australian Context **Government policy**: The National AI Plan (March 2026) confirmed reliance on existing laws and sector regulators rather than a standalone AI Act. Defence released binding governance for AI use across ADF. New DTA Cloud Policy (effective July 1, 2026) mandates APS entities prioritise cloud computing. The AI Safety Institute is operational with AUD \$29.9 million in funding. **Defence spending**: \$1.2 billion in the 2025-26 budget for sovereign capability development in AI and autonomous systems. Defence Innovation Hub has funded 80+ AI-related projects. ASD-AWS "Top Secret Cloud" partnership worth approximately AUD \$2 billion over a decade. **Data centre infrastructure**: Three main sovereign providers are building AI-capable facilities: * **CDC Data Centres**: 200MW AI campus near Perth (AUD \$415M first stage, operational 2026). * **Macquarie Data Centres**: IC3 Super West, 47MW AI data centre in Sydney (AUD \$350M, opening September 2026). Partnering with Dell for Sovereign AI Factories powered by NVIDIA. * **NEXTDC**: S7 site at Eastern Creek, Sydney (650MW capacity, partnered with OpenAI, Phase 1 expected H2 2027). **Regulatory landscape**: ASIC requires AI in financial services to align with responsible lending and market integrity obligations. TGA released guidance on AI-based software as a medical device. New privacy obligations effective December 2026 require disclosure of automated decision-making. **The gap**: Australia has sovereign compute infrastructure under construction and defence funding in place, but lacks a domestic foundation model programme. The Fable 5 suspension demonstrated that Five Eyes membership provides no exemption from US export controls. Australia's current position is Level 1-2 for frontier AI (API-dependent on US providers) with infrastructure being built for Level 2-3. *** ## Implications for Annie This section identifies what the landscape means for Annie's positioning. The full competitive analysis is in doc 03. ### Where the Gaps Are 1. **The 80% problem**: For 80% of production use cases, a well-tuned specialist model works as well as a frontier model and costs 95% less. But the tooling, orchestration, and confidence to run multi-model systems does not exist as a product. Every organisation doing this today is building it from scratch. 2. **The sovereignty gap is operational, not theoretical**: Before June 12, sovereign AI was a compliance discussion. Now it is about whether your AI infrastructure survives a single government directive. There is no product that packages sovereign AI deployment as a turnkey solution with the user experience of a frontier API. 3. **The ensemble evidence is strong but unexploited**: Research consistently shows that ensembles of smaller models can outperform single large models with higher accuracy and fewer total FLOPs. No commercial product operationalises this finding. 4. **Fine-tuning at the bottom, frontier at the top, nothing in between**: You can fine-tune a 7B model for under \$5 or pay \$50/MTok for Fable 5. There is no product that intelligently routes between a portfolio of specialists and frontier fallbacks based on task complexity. ### What the Export Controls Create as Opportunity The Fable 5 suspension created three market conditions that did not exist two weeks ago: 1. **Enterprise demand for multi-model resilience**: 81% of enterprises now run three or more AI model families (up from 13% a year ago), and every procurement conversation now includes "what happens if we lose access." A system architecturally designed for multi-model orchestration is no longer a nice-to-have. 2. **Government demand for sovereign AI that actually works**: More than 60 nations have published AI strategies, over 30 have committed funding, and the sovereign AI infrastructure market is projected to reach \$301.6 billion by 2040. But most sovereign AI initiatives are infrastructure plays (data centres, GPU clusters) without the model-layer product to run on them. 3. **The open-weight window**: Published open-weight models are currently exempt from US export controls under ECCN 4E091. This regulatory posture could change. Sovereign entities should be downloading and fine-tuning open models now. A product that makes this easy has a time-limited but significant advantage. ### Why Small Specialist Models Matter Now * Serving a 7B specialist is 10-30x cheaper than running a 70B-175B general model. * Training a 1B specialist costs \$2K-\$15K. Training a 7B specialist costs \$50K-\$500K. Fine-tuning a 7B model for a specific domain costs under \$5. * Small models (250M to 27B) run on hardware ranging from phones to single consumer GPUs. No datacenter required. * India's Bhashini programme demonstrates the sovereign small-model strategy at national scale: purpose-built language models serving 140 million users across 22 languages on domestic sovereign infrastructure. * Research shows performance gains decrease exponentially beyond certain parameter thresholds, making smaller models more cost-effective for most defined tasks. The limitation is real: small models lag significantly on complex tasks requiring deeper reasoning or nuanced understanding. They match large models in specific, well-defined scenarios but not in general-purpose reasoning. This is precisely where intelligent orchestration -- routing easy tasks to cheap specialists and hard tasks to capable models -- closes the gap. ### The Cost and Accessibility Advantage The frontier labs are spending staggering amounts: Anthropic projects \$29 billion in 2026 losses, OpenAI projects \$14 billion, Google guided \$175-185 billion in capex. These economics require massive scale to justify and produce products priced accordingly (Fable 5 at \$50/MTok output, GPT-5.5 Pro at \$180/MTok output). A system built from twelve specialist models in the 250M-27B range, each fine-tuned for its domain, running on hardware costing \$5K-\$50K total, with intelligent routing to minimise frontier API fallback, could deliver comparable task performance at 1-2 orders of magnitude lower cost. The total training cost for the specialist portfolio would be a rounding error in a frontier lab's monthly electricity bill. This is not a hypothetical. The models exist (Gemma 4, Qwen 3.6, Phi-4, Mistral Small 4). The hardware exists (consumer GPUs). The research supports ensemble approaches. What does not yet exist is the product that makes it work reliably and is simple enough for organisations to adopt. *** ## Sources ### Anthropic * [Claude Fable 5 and Mythos 5 announcement](https://www.anthropic.com/news/claude-fable-5-mythos-5) * [Statement on export control suspension](https://www.anthropic.com/news/fable-mythos-access) * [Claude Opus 4.8 announcement](https://www.anthropic.com/news/claude-opus-4-8) * [Claude Fable product page](https://www.anthropic.com/claude/fable) * [Claude Opus product page](https://www.anthropic.com/claude/opus) * [Responsible Scaling Policy v3.0](https://anthropic.com/responsible-scaling-policy/rsp-v3-0) * [MorphLLM: Claude Benchmarks 2026](https://www.morphllm.com/claude-benchmarks) * [Weights & Biases: Fable 5 Benchmark Scores](https://wandb.ai/byyoung3/ml-news/reports/Claude-Fable-5-Benchmark-Scores--VmlldzoxNzE3NTE3MQ) * [Tom's Hardware: Claude Fable 5 review](https://www.tomshardware.com/tech-industry/artificial-intelligence/claude-fable-5-brings-mythos-to-the-masses-anthropics-next-frontier-model-is-state-of-the-art-on-nearly-all-tested-benchmarks) * [Fortune: Anthropic disables Fable/Mythos](https://fortune.com/2026/06/13/anthropic-disables-fable-mythos-export-controls-national-security-threat/) * [Nextgov: Export control order details](https://www.nextgov.com/artificial-intelligence/2026/06/anthropic-suspends-top-ai-models-after-us-export-control-order/414173/) * [Fortune: Sovereign AI scramble](https://fortune.com/2026/06/16/anthropic-shutdown-sparks-global-scramble-for-sovereign-ai/) * [Washington Post: House demands answers](https://www.washingtonpost.com/technology/2026/06/18/house-members-want-answers-export-controls-placed-anthropic-fable/) * [Klover.ai: Anthropic IPO infrastructure economics](https://www.klover.ai/anthropic_ipo_infrastructure_economics_of_compute_supply_chain_indepth_analysis_2026/) * [SaaStr: Anthropic revenue vs training spend](https://www.saastr.com/anthropic-just-passed-openai-in-revenue-while-spending-4x-less-to-train-their-models/) * [Axios: Mythos-class safeguards](https://www.axios.com/2026/06/09/anthropic-mythos-class-safeguards) * [CNBC: Anthropic Mythos release](https://www.cnbc.com/2026/06/09/anthropic-mythos-claude-fable-5.html) * [Simon Willison: Fable 5 impressions](https://simonwillison.net/2026/Jun/9/claude-fable-5/) * [Al Jazeera: US asks Anthropic to block global access](https://www.aljazeera.com/news/2026/6/14/us-asks-anthropic-to-block-global-access-to-top-ai-models-why-it-matters) * [Digital Applied: Fable 5 vs GPT-5.5](https://www.digitalapplied.com/blog/claude-fable-5-vs-gpt-5-5-frontier-comparison-2026) * [Artificial Analysis: Fable 5 Intelligence Index](https://artificialanalysis.ai/articles/claude-fable-5-mythos-intelligence-index) ### OpenAI * [GPT-5.5 Docs](https://developers.openai.com/api/docs/models/gpt-5.5) * [Introducing GPT-5.5](https://openai.com/index/introducing-gpt-5-5/) * [Codex Changelog](https://developers.openai.com/codex/changelog) * [Codex Models](https://developers.openai.com/codex/models) * [Introducing GPT-5.3-Codex](https://openai.com/index/introducing-gpt-5-3-codex/) * [Codex for Almost Everything](https://openai.com/index/codex-for-almost-everything/) * [Introducing o3 and o4-mini](https://openai.com/index/introducing-o3-and-o4-mini/) * [OpenAI Pricing](https://developers.openai.com/api/docs/pricing) * [Open Weight Models](https://help.openai.com/en/articles/11870455-openai-open-weight-models-gpt-oss) * [Introducing gpt-oss](https://openai.com/index/introducing-gpt-oss/) * [Open Weights and AI for All](https://openai.com/global-affairs/open-weights-and-ai-for-all/) * [NEXTDC Partnership](https://www.nextdc.com/news/building-the-next-generation-of-sovereign-ai-infrastructure-in-australia) * [OpenAI for Australia](https://openai.com/global-affairs/openai-for-australia/) * [O-mega Complete Guide](https://o-mega.ai/articles/gpt-5-5-the-complete-guide-2026) * [TokenMix Review](https://tokenmix.ai/blog/gpt-5-5-spud-review-88-swe-bench-2026) * [AI Pricing Guru](https://www.aipricing.guru/openai-pricing/) * [DeployBase Pricing](https://deploybase.ai/articles/openai-api-pricing-2026) * [Sam Altman AGI Shift](https://www.startuphub.ai/ai-news/ai-figures/2026/figure-sam-altman-public-position-evolution-2026-06-13) * [Epoch AI Training Compute](https://epoch.ai/gradient-updates/why-gpt5-used-less-training-compute-than-gpt45-but-gpt6-probably-wont) * [AI Inference Cost Crisis](https://aiautomationglobal.com/blog/ai-inference-cost-crisis-openai-economics-2026) * [Microsoft Sovereign Cloud](https://azure.microsoft.com/en-us/blog/microsoft-strengthens-sovereign-cloud-capabilities-with-new-services/) * [Microsoft-OpenAI Non-Exclusive](https://www.vaasblock.com/news/microsoft-openai-non-exclusive-partnership-azure-2026/) ### Google * [Gemini Developer API Pricing](https://ai.google.dev/gemini-api/docs/pricing) * [Gemini API Models](https://ai.google.dev/gemini-api/docs/models) * [Gemini 3.1 Pro Model Card](https://deepmind.google/models/model-cards/gemini-3-1-pro/) * [Gemini 3.5 Flash and Pro: Google I/O 2026](https://pasqualepillitteri.it/en/news/2984/gemini-3-5-flash-pro-google-io-2026) * [Gemini 3.5 Flash Complete Guide](https://www.nxcode.io/resources/news/gemini-3-5-flash-complete-guide-benchmarks-pricing-api-2026) * [Gemini 3.1 Pro Benchmarks](https://layerlens.ai/blog/gemini-3-1-pro-benchmark-review) * [Gemini 2.5 Technical Report (arxiv 2507.06261)](https://arxiv.org/pdf/2507.06261) * [Gemma 4 -- Google DeepMind](https://deepmind.google/models/gemma/gemma-4/) * [Gemma 4 Complete Guide](https://dev.to/aniruddhaadak/gemma-4-complete-guide-2026-architecture-benchmarks-deployment-3en9) * [Gemma 4 Apache 2.0](https://www.mindstudio.ai/blog/what-is-google-gemma-4-apache-open-weight) * [Google Distributed Cloud at Next '26](https://cloud.google.com/blog/topics/hybrid-cloud/google-distributed-cloud-at-next26) * [Forrester Wave Sovereign Cloud 2026](https://cloud.google.com/blog/products/identity-security/a-leader-in-forrester-wave-sovereign-cloud-platform-2026) * [Google \$75B AI Infrastructure Spend](https://valueaddvc.com/blog/google-75b-ai-infrastructure-spend-data-centers-tpus-and-the-gemini-bet) * [AI Infrastructure at Next '26](https://cloud.google.com/blog/products/compute/ai-infrastructure-at-next26) * [DeepMind Scaling Philosophy](https://eu.36kr.com/en/p/3619378357748743) * [ML Training Cost Statistics 2026](https://www.aboutchromebooks.com/machine-learning-model-training-cost-statistics/) ### Open-Source / Open-Weight * [Meta Llama 4 Blog](https://ai.meta.com/blog/llama-4-multimodal-intelligence/) * [Llama 4 License](https://www.llama.com/llama4/license/) * [Llama 4 EU Exclusion](https://dionwiggins.substack.com/p/llama-4-is-banned-in-the-eu-open) * [Llama FAQ](https://www.llama.com/faq/) * [Unsloth Llama 4 Guide](https://unsloth.ai/docs/models/tutorials/llama-4-how-to-run-and-fine-tune) * [Mistral Models Overview](https://docs.mistral.ai/models/overview) * [Mistral Defense Deal](https://www.analyticsinsight.net/amp/story/news/mistral-ai-signs-defense-deal-with-france-as-europe-pushes-for-sovereign-ai) * [Mistral Sovereign AI](https://www.raconteur.net/global-business/mistral-bets-big-on-european-sovereign-ai) * [France NVIDIA Hub](https://windowsnews.ai/article/france-fires-up-sovereign-ai-engine-mistrals-nvidia-powered-hub-goes-live-to-fuel-eus-open-model-amb.427898) * [Mistral Small 4](https://serenitiesai.com/articles/mistral-ai-models-2026-complete-guide) * [Qwen3 GitHub](https://github.com/QwenLM/Qwen3) * [Qwen 3.6 GitHub](https://github.com/QwenLM/Qwen3.6) * [Qwen 3.5 Blog](https://qwen.ai/blog?id=qwen3.5) * [Qwen Licensing](https://qwenimage.art/blog/qwen-ai-license-explained) * [VentureBeat Qwen3](https://venturebeat.com/ai/alibaba-launches-open-source-qwen3-model-that-surpasses-openai-o1-and-deepseek-r1) * [Qwen 3.6 VRAM Guide](https://willitrunai.com/blog/qwen-3-6-vram-requirements) * [DeepSeek-V3 Technical Report](https://arxiv.org/html/2412.19437v1) * [DeepSeek V4 Benchmarks](https://deepseekai.guide/news/deepseek-benchmarks-2026/) * [DeepSeek V4 Review](https://www.morphllm.com/deepseek-v4) * [DeepSeek V4 MIT License](https://framia.converge.ai/page/en-US/news/deepseek-v4-open-source-mit-license) * [CSIS DeepSeek Analysis](https://www.csis.org/analysis/delving-dangers-deepseek) * [DeepSeek V4 Self-Hosting Guide](https://lushbinary.com/blog/deepseek-v4-self-hosting-guide-vllm-hardware-deployment/) * [DeepSeek V4 VRAM](https://codersera.com/blog/deepseek-v4-vram-gpu-requirements-2026/) * [Gemma 4 Blog](https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/) * [Gemma 4 Hardware Guide](https://www.gemma4.wiki/requirements/gemma-4-hardware-requirements) * [Phi-4-mini HuggingFace](https://huggingface.co/microsoft/Phi-4-mini-instruct) * [Microsoft Phi Models](https://azure.microsoft.com/en-us/products/phi) * [Phi-4-reasoning-vision](https://siliconangle.com/2026/03/04/microsoft-open-sources-multimodal-reasoning-model-15b-parameters/) * [Falcon H1 Launch](https://www.businesswire.com/news/home/20250521901857/en/) * [Cohere Models](https://docs.cohere.com/docs/models) ### Sovereign AI and Export Controls * [Fable 5 Suspension Facts and Timeline](https://www.eisneramper.com/insights/artificial-intelligence-insights/fable-5-suspension-facts-and-timeline-0626/) * [Enterprise Impact](https://www.fifthrow.com/blog/us-export-control-order-and-global-suspension-of-fable-5-mythos-5-operationalizing-compliance-as-a-live-mandate) * [Security Team Implications](https://snyk.io/blog/fable-mythos-suspension-security-takeaways/) * [Enterprise AI Under Export Controls](https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-model-export-controls-enterprise-govern/) * [Fable 5 Ban Update](https://www.techtimes.com/articles/318760/20260620/fable-5-ban-update-trump-softens-directive-stands-refund-deadline-closes-today.htm) * [Fable 5 Full Story](https://www.explainx.ai/blog/us-government-bans-fable-5-mythos-5-anthropic-export-control-2026) * [Europe Wake-Up Call](https://www.euronews.com/2026/06/13/wake-up-call-europe-reacts-to-anthropic-halting-access-to-its-fable-5-and-mythos-5-ai-mode) * [Europe AI Sovereignty Crisis G7](https://www.techtimes.com/articles/318611/20260618/europe-ai-sovereignty-crisis-g7-offers-platform-kill-switch-fears-grow.htm) * [EU AI Sovereignty Push at G7](https://www.computing.co.uk/news/2026/at-g7-euro-ai-sovereignty-push-intensifies) * [Washington AI Kill Switch](https://aiweekly.co/newsletters/ai-geopolitics/washingtons-ai-kill-switch-broke-the-alliance) * [Al Jazeera: US Export Ban Strains Alliances](https://www.aljazeera.com/news/2026/6/19/us-export-ban-on-anthropics-ai-models-further-strains-alliances) * [EU Insider: Washington Cuts Europe Off](https://www.euinsider.eu/news/us-bans-europeans-anthropic-ai-sovereignty-2026) * [SmartCompany: Why Australians Lost Access](https://www.smartcompany.com.au/artificial-intelligence/why-australians-lost-access-to-anthropic-fable-5-mythos-5/) * [Kate Carruthers: Sovereign AI Got Real](https://katecarruthers.com/anthropic-fable-5-and-why-sovereign-ai-just-got-real/) * [AIMadeTools: Sovereign AI Models 2026](https://www.aimadetools.com/blog/sovereign-ai-models-2026) * [McKinsey: Sovereign AI Ecosystems](https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/sovereign-ai-building-ecosystems-for-strategic-resilience-and-impact) * [Sovereign AI Definition and Maturity Model](https://zeroandone.me/blogs/sovereign-ai.html) * [Stanford HAI: AI Sovereignty Definitional Dilemma](https://hai.stanford.edu/news/ai-sovereigntys-definitional-dilemma) * [CNAS Sovereign AI Index](https://interactives.cnas.org/reports/sovereign-ai-index/) * [Sovereign AI Infrastructure Guide 2026](https://fluxhuman.com/en/blog/sovereign-ai-infrastructure-the-2026-guide) * [BIS Export Controls](https://www.wilmerhale.com/en/insights/publications/20250205-bis-issues-long-awaited-export-controls-on-ai) * [Hogan Lovells Analysis](https://www.hoganlovells.com/en/publications/us-department-of-commerce-expands-controls-on-advanced-semiconductors-and-establishes) * [MindStudio: Export Controls Explained](https://www.mindstudio.ai/blog/ai-export-controls-claude-fable-5-enterprise-implications) * [Google/OpenAI Push to Ease Controls](https://www.bankinfosecurity.com/google-openai-push-urges-trump-to-ease-ai-export-controls-a-27739) * [RAND Analysis](https://www.rand.org/pubs/perspectives/PEA3776-1.html) ### Australian Context * [Australia National AI Plan](https://www.6clicks.com/resources/blog/australias-national-ai-plan-sovereign-ai-compliance-leaders) * [Australia Defence AI Policy](https://www.6clicks.com/resources/blog/australia-defence-ai-policy-sovereign-grc-compliance) * [APS AI Plan 2025](https://www.digital.gov.au/policy/ai/australian-public-service-ai-plan-2025/what-we-plan-achieve) * [Australia Sovereign AI Governance-Led](https://securitybrief.com.au/story/australia-s-sovereign-ai-adoption-stays-governance-led) * [UNSW Defence AI Project](https://www.unsw.edu.au/newsroom/news/2026/01/unsw-sydney-lands-3m-defence-project-to-build-next-gen-ai) * [AI Regulation in Australia 2026](https://www.theadaptavistgroup.com/blog/ai-regulation-in-australia) * [Financial Services AI Compliance](https://www.insidetechlaw.com/blog/2026/03/regulating-ai-in-australian-financial-services-practical-guidance-for-compliance) * [Australia Gov Cloud Market 2026](https://vocal.media/trader/australia-government-cloud-market-2026-sovereign-cloud-surge-whole-of-government-policy-and-ai-integration) * [CDC Data Centres](https://cdc.com/) * [CDC Perth AI Campus](https://www.datacenterdynamics.com/en/news/australias-cdc-plans-200mw-data-center-campus-outside-perth/) * [Macquarie Data Centres](https://www.macquariedatacentres.com/) * [Macquarie IC3 Super West](https://aimagazine.com/news/macquaries-ic3-super-west-ai-data-centre-for-huge-gpu-loads) * [Macquarie Dell Sovereign AI Factory](https://www.cyberdaily.au/digital-transformation/12963-macquarie-data-centres-tops-out-ai-focused-sydney-cloud-ai-data-centre) * [NEXTDC S7 Campus](https://datacentremagazine.com/news/how-will-nextdc-ai-campus-drive-openai-for-australia) * [NEXTDC Sovereign Data Centres](https://www.nextdc.com/blog/data-centre-sovereignty-for-control-and-compliance) ### Hardware, Costs, and Small Models * [VRAM Requirements 2026](https://localaimaster.com/blog/vram-requirements-2026) * [LLM Hardware Requirements 2026](https://overchat.ai/ai-hub/llm-hardware-requirements) * [Best GPU for LLM 2026](https://bizon-tech.com/blog/best-gpu-llm-training-inference) * [Local AI vs Cloud AI 2026](https://www.mindstudio.ai/blog/local-ai-vs-cloud-ai-2026) * [AI API Pricing Comparison June 2026](https://devtk.ai/en/blog/ai-api-pricing-comparison-2026/) * [Frontier AI Cost Crisis](https://www.figuringoutwithai.com/playbooks/frontier-ai-cost-crisis-local-model-migration-playbook-2026) * [AI Data Center Power Requirements 2026](https://techplustrends.com/power-requirements-ai-data-centers/) * [AI Inference Power Consumption](https://www.spheron.network/blog/ai-inference-power-electricity-cost-2026/) * [Small Language Models Guide 2026](https://machinelearningmastery.com/introduction-to-small-language-models-the-complete-guide-for-2026/) * [Small Language Models Enterprise Cost Guide](https://iterathon.tech/blog/small-language-models-enterprise-2026-cost-efficiency-guide) * [Training Sovereign Language Models](https://arxiv.org/html/2510.19801) * [AI Model Training Costs 2026](https://localaimaster.com/blog/ai-model-training-costs-2025-analysis) * [India Sovereign AI Status 2026](https://explainx.ai/blog/india-sovereign-ai-status-indiaai-mission-2026) * [Bhashini Migration to Sovereign Cloud](https://www.crnasia.com/india/news/2026/yotta-enables-bhashini-s-migration-to-sovereign-ai-cloud-on-indian-infrastructure) * [Ensemble vs Large Models](https://arxiv.org/pdf/2005.00570) * [AI Model Size vs Performance 2026](https://localaimaster.com/blog/ai-model-size-vs-performance-analysis-2025) * [Sovereign AI Enterprise Guide 2026](https://www.advisori.de/en/blog/sovereign-ai-vendor-lock-in-enterprise-guide-2026) * [Sovereign AI Infrastructure Market](https://www.rootsanalysis.com/sovereign-ai-infrastructure-market) * [The High Cost of Sovereignty](https://www.idc.com/resource-center/blog/the-high-cost-of-sovereignty-in-the-age-of-ai/) * [The Sovereignty Illusion](https://www.knectiq.com/sovereign-ai/) * [SoftBank EUR 75B French AI Investment](https://fortune.com/2026/05/30/softbank-75-billion-investment-french-ai-data-centers-masayoshi-son-emmanuel-macron/) *** *Document prepared June 22, 2026 by Annie. The AI model landscape is evolving rapidly. Benchmark figures, pricing, and availability are subject to change. Where parameter counts or architecture details are not officially confirmed, this is noted explicitly.* # Add export Source: https://docs.quiva.ai/api-reference/account-management/add-export /api-reference/endpoint/accounts/openapi.json post /accounts/export Adds a new export to the account. - Enables an account to make its internal resources (such as data streams or services) available to other accounts. # Add import Source: https://docs.quiva.ai/api-reference/account-management/add-import /api-reference/endpoint/accounts/openapi.json post /accounts/import Adds a new import to the account. - When an account adds an import, it's essentially creating a mapping that allows it to access and utilize resources (such as data streams or services) that are exported by another account. This request creates a trust relationship between accounts where one account can consume resources produced by another. # Cancel account deletion Source: https://docs.quiva.ai/api-reference/account-management/cancel-account-deletion /api-reference/endpoint/accounts/openapi.json post /accounts/cancel-account-deletion Cancels the deletion of an account. - Allows the root user of an account to reverse a scheduled account deletion if the current date is before *deletion_date*. # Confirm email Source: https://docs.quiva.ai/api-reference/account-management/confirm-email /api-reference/endpoint/accounts/openapi.json post /accounts/confirm-email Confirms a user's email address using a token or code - This process ensures that the person registering or using an email address actually has access to it. When a user registers a new account or adds a new email address to their profile, the system sends a confirmation email containing either a token (a long, randomly generated string) or a numerical code. This function processes the verification when the user clicks the confirmation link in the email or manually enters the provided code. It accepts either: - a) A token parameter, which is included in the confirmation link - b) A code parameter, which is a shorter numerical sequence that can be manually entered # Create account Source: https://docs.quiva.ai/api-reference/account-management/create-account /api-reference/endpoint/accounts/openapi.json post /accounts - Creates a new account and a user if the provided email does not exist on the system. - Associates the user with the account with root privileges. - Sends an email to the user to verify their email address if the email is not already verified on the system. - A user can be a member of any number of accounts but they can be a root of at most one account. # Delete account Source: https://docs.quiva.ai/api-reference/account-management/delete-account /api-reference/endpoint/accounts/openapi.json delete /accounts/account Deletes the current account. - Allows the root user of an account to delete the entire account from the platform. - The root user must request a verification code to be able to finalize the process (see **/accounts/email-code**) - This is a significant operation that removes all account data and revokes access for all users associated with the account. - Users, however, are not deleted and they can continue to sign in to the platform if they are members of any other active accounts. - There is an option to delete the account immediately or in 14 days. If the latter, the deletion can be unscheduled (see **/accounts/cancel-account-deletion**) # Get account information Source: https://docs.quiva.ai/api-reference/account-management/get-account-information /api-reference/endpoint/accounts/openapi.json get /accounts/info Retrieves comprehensive information about the authenticated user's current account in a public-facing format (no sensitive data is returned). # List accounts by user Source: https://docs.quiva.ai/api-reference/account-management/list-accounts-by-user /api-reference/endpoint/accounts/openapi.json get /accounts/list Lists all accounts for the current user # Remove export Source: https://docs.quiva.ai/api-reference/account-management/remove-export /api-reference/endpoint/accounts/openapi.json delete /accounts/export Removes an export from the account. - Allows an account to discontinue sharing a previously exported resource with other accounts. This operation effectively revokes external access to the specified resource. # Remove import Source: https://docs.quiva.ai/api-reference/account-management/remove-import /api-reference/endpoint/accounts/openapi.json delete /accounts/import Removes an import from the account. Allows an account to discontinue the integration with resources from another account. This operation effectively severs the connection to external resources that an account was consuming. # Rename account Source: https://docs.quiva.ai/api-reference/account-management/rename-account /api-reference/endpoint/accounts/openapi.json patch /accounts/rename Renames an account - Allows the root user of an account to change the display name and internal name of their account - They must request a verification code to be able to finalize the process (see **/accounts/email-code**) - Once an account name is changed, all user sessions will be automatically revoked. Affected users will be notified about the change by email. # Resend email confirmation Source: https://docs.quiva.ai/api-reference/account-management/resend-email-confirmation /api-reference/endpoint/accounts/openapi.json post /accounts/resend-confirmation Resends email confirmation link to user. Always returns success for security reasons. # Set emergency credit Source: https://docs.quiva.ai/api-reference/account-management/set-emergency-credit /api-reference/endpoint/accounts/openapi.json patch /accounts/set-emergency-credit Sets the emergency credit amount for an account. - Allows users with appropriate permissions (root, admin, or billing roles) to configure a maximum monthly emergency budget. - For accounts with fixed resource allocations, emergency credit funds the provisioning of additional resources when regular allocations are exhausted. # Submit contact form Source: https://docs.quiva.ai/api-reference/account-management/submit-contact-form /api-reference/endpoint/accounts/openapi.json post /accounts/submit-contact-form Submits a contact form - The function provides a structured mechanism for authenticated users to send inquiries, feedback, or support requests directly to the quiva.ai team or specific team members. This communication channel is integrated within the platform, allowing users to reach out for assistance without leaving the application environment. # Update account Source: https://docs.quiva.ai/api-reference/account-management/update-account /api-reference/endpoint/accounts/openapi.json patch /accounts Updates account information. - An update can be made to any of the non-critical editable account properties only by root and admin users. - Critical account properties such as name, ID, and root email address cannot be changed with this request - To update *name*, see **/accounts/rename** - To update *root_email*, see **/accounts/request-email-address-change** # Delete API key Source: https://docs.quiva.ai/api-reference/api-keys/delete-api-key /api-reference/endpoint/accounts/openapi.json delete /accounts/api-key Delete an API key. - A key can be delted by providing the full key or by id. - Deleted API keys cannot be recovered # Get API credentials Source: https://docs.quiva.ai/api-reference/api-keys/get-api-credentials /api-reference/endpoint/accounts/openapi.json get /accounts/api-creds Retrieves credentials for a specific API key. -Allows users to retrieve the full credential string that can be used with the CLI. # Issue API key Source: https://docs.quiva.ai/api-reference/api-keys/issue-api-key /api-reference/endpoint/accounts/openapi.json post /accounts/api-key Issues a new API key for the specified user - This function nables users to generate programmatic access credentials for quiva.ai API as well as credentials for CLI access. This feature allows developers to integrate the platform's capabilities into external applications, automation scripts, or CI/CD pipelines without requiring interactive login sessions. # View API keys Source: https://docs.quiva.ai/api-reference/api-keys/view-api-keys /api-reference/endpoint/accounts/openapi.json get /accounts/api-keys Retrieves the metadate for all API keys for the specified user. - The function allows users to view the metadata for all active API keys for their own account - Root and administrator users can view API keys metadata associated with other users on the same account. # Authenticate with API key Source: https://docs.quiva.ai/api-reference/authentication/authenticate-with-api-key /api-reference/endpoint/accounts/openapi.json post /accounts/auth-with-api-key Authenticates a user using an API key. # Authenticate with password Source: https://docs.quiva.ai/api-reference/authentication/authenticate-with-password /api-reference/endpoint/accounts/openapi.json post /accounts/auth-with-password Authenticates a user with their email and password. If MFA is enabled, returns an MFA token that must be used with a subsequent authentication request. # Authenticate with recovery code Source: https://docs.quiva.ai/api-reference/authentication/authenticate-with-recovery-code /api-reference/endpoint/accounts/openapi.json post /accounts/auth-with-recovery-code Complete authentication using a recovery code after initial password verification. Recovery codes are one-time use. # Authenticate with TOTP Source: https://docs.quiva.ai/api-reference/authentication/authenticate-with-totp /api-reference/endpoint/accounts/openapi.json post /accounts/auth-with-totp Complete authentication using a Time-based One-Time Password (TOTP) after initial password verification. # Begin passkey authentication Source: https://docs.quiva.ai/api-reference/authentication/begin-passkey-authentication /api-reference/endpoint/accounts/openapi.json post /accounts/begin-auth-with-passkey Initiates the WebAuthn passkey authentication flow. Returns a challenge for the client to sign with the user's passkey. # Complete passkey authentication Source: https://docs.quiva.ai/api-reference/authentication/complete-passkey-authentication /api-reference/endpoint/accounts/openapi.json post /accounts/complete-auth-with-passkey Completes the WebAuthn passkey authentication flow by verifying the signed challenge. # Get email verification code Source: https://docs.quiva.ai/api-reference/authentication/get-email-verification-code /api-reference/endpoint/accounts/openapi.json get /accounts/email-code Sends a verification code to the user's email # Get session tokens Source: https://docs.quiva.ai/api-reference/authentication/get-session-tokens /api-reference/endpoint/accounts/openapi.json get /accounts/session-tokens Retrieves all active session tokens for the current user # Initiate OAuth login flow Source: https://docs.quiva.ai/api-reference/authentication/initiate-oauth-login-flow /api-reference/endpoint/accounts/openapi.json get /accounts/oauth/login Generates OAuth URL for authentication with external providers (Auth0) # OAuth callback handler Source: https://docs.quiva.ai/api-reference/authentication/oauth-callback-handler /api-reference/endpoint/accounts/openapi.json get /accounts/oauth/callback Handles OAuth callback from external authentication provider # Refresh authentication token Source: https://docs.quiva.ai/api-reference/authentication/refresh-authentication-token /api-reference/endpoint/accounts/openapi.json post /accounts/refresh-token Issues a new authentication token using a valid refresh token. # Request password reset Source: https://docs.quiva.ai/api-reference/authentication/request-password-reset /api-reference/endpoint/accounts/openapi.json post /accounts/request-reset-password Requests a password reset for a user # Reset password Source: https://docs.quiva.ai/api-reference/authentication/reset-password /api-reference/endpoint/accounts/openapi.json post /accounts/reset-password Resets a user's password # Revoke all tokens Source: https://docs.quiva.ai/api-reference/authentication/revoke-all-tokens /api-reference/endpoint/accounts/openapi.json delete /accounts/tokens Revokes all tokens for a specific user # Revoke auth token Source: https://docs.quiva.ai/api-reference/authentication/revoke-auth-token /api-reference/endpoint/accounts/openapi.json delete /accounts/auth-token Revokes a specific authentication token # Revoke refresh token Source: https://docs.quiva.ai/api-reference/authentication/revoke-refresh-token /api-reference/endpoint/accounts/openapi.json delete /accounts/refresh-token Revokes a specific refresh token # Set password Source: https://docs.quiva.ai/api-reference/authentication/set-password /api-reference/endpoint/accounts/openapi.json post /accounts/set-password Sets a password for a new user # Validate token Source: https://docs.quiva.ai/api-reference/authentication/validate-token /api-reference/endpoint/accounts/openapi.json post /accounts/token-validate Validates an authentication or refresh token. # Create a new collection Source: https://docs.quiva.ai/api-reference/collections/create-a-new-collection /api-reference/endpoint/hub-flows/openapi.json post /hub/collections Creates a new collection for organizing nodes or workflows # Delete a collection Source: https://docs.quiva.ai/api-reference/collections/delete-a-collection /api-reference/endpoint/hub-flows/openapi.json delete /hub/collections/{collection_type}/{collection_topic} Deletes a specific collection by subject # Get a collection Source: https://docs.quiva.ai/api-reference/collections/get-a-collection /api-reference/endpoint/hub-flows/openapi.json get /hub/collections/{collection_type}/{collection_topic} Retrieves a specific collection # List collections Source: https://docs.quiva.ai/api-reference/collections/list-collections /api-reference/endpoint/hub-flows/openapi.json get /hub/collections Retrieves a list of collections # Update a Workflow or Node collection Source: https://docs.quiva.ai/api-reference/collections/update-a-workflow-or-node-collection /api-reference/endpoint/hub-flows/openapi.json patch /hub/collections/{collection_type}/{collection_topic} Updates an existing collection's properties # Create a gateway trigger Source: https://docs.quiva.ai/api-reference/gateway-triggers/create-a-gateway-trigger /api-reference/endpoint/trigger/openapi.json post /trigger/gateway Creates a new gateway trigger for API endpoint management. A gateway trigger establishes an HTTP endpoint that can be accessed from outside the system and routes requests to appropriate handlers. **Usage Notes:** - The system automatically generates a unique topic for your gateway trigger - Default timeout is 30 seconds (30000ms) if not specified - Gateway mappings define the HTTP method, path, and access controls - Setting `is_public: true` allows the endpoint to be accessed without authentication - Rate limits can be configured per endpoint - After creation, the gateway endpoint is immediately available for traffic - The response includes the full subject identifier which may be needed for other operations # Create a new gateway Source: https://docs.quiva.ai/api-reference/gateways/create-a-new-gateway /api-reference/endpoint/gateway/openapi.json post /gateway Creates a new gateway with a unique topic and URL. # Delete a gateway Source: https://docs.quiva.ai/api-reference/gateways/delete-a-gateway /api-reference/endpoint/gateway/openapi.json delete /gateway/{gateway_topic} Marks a gateway as deleted (poison-pilled). # Get a specific gateway Source: https://docs.quiva.ai/api-reference/gateways/get-a-specific-gateway /api-reference/endpoint/gateway/openapi.json get /gateway/{gateway_topic} Retrieves details about a specific gateway. # Get all gateways Source: https://docs.quiva.ai/api-reference/gateways/get-all-gateways /api-reference/endpoint/gateway/openapi.json get /gateway/all Retrieves a list of all available gateways. # Update a gateway Source: https://docs.quiva.ai/api-reference/gateways/update-a-gateway /api-reference/endpoint/gateway/openapi.json patch /gateway/{gateway_topic} Updates an existing gateway's configuration. # List all indexers Source: https://docs.quiva.ai/api-reference/indexers/list-all-indexers /api-reference/endpoint/storage/openapi.json get /storage/indexers Retrieves a list of all indexers # Create a new KV bucket Source: https://docs.quiva.ai/api-reference/kv-buckets/create-a-new-kv-bucket /api-reference/endpoint/storage/openapi.json post /storage/kv Creates a new key-value bucket with optional indexing configuration # Create or update a KV bucket Source: https://docs.quiva.ai/api-reference/kv-buckets/create-or-update-a-kv-bucket /api-reference/endpoint/storage/openapi.json put /storage/kv/{bucket} Creates a new key-value bucket with optional indexing or updates it if one does not exist # Delete KV bucket Source: https://docs.quiva.ai/api-reference/kv-buckets/delete-kv-bucket /api-reference/endpoint/storage/openapi.json delete /storage/kv/{bucket} Deletes a key-value bucket and all its contents # Get KV bucket details Source: https://docs.quiva.ai/api-reference/kv-buckets/get-kv-bucket-details /api-reference/endpoint/storage/openapi.json get /storage/kv/{bucket} Retrieves information about a specific KV bucket # List all KV buckets Source: https://docs.quiva.ai/api-reference/kv-buckets/list-all-kv-buckets /api-reference/endpoint/storage/openapi.json get /storage/kv Retrieves a list of all key-value buckets in the system # Purge KV bucket Source: https://docs.quiva.ai/api-reference/kv-buckets/purge-kv-bucket /api-reference/endpoint/storage/openapi.json delete /storage/kv/{bucket}/purge Purges all data from a key-value bucket while keeping the bucket configuration # Create or update KV entry Source: https://docs.quiva.ai/api-reference/kv-entries/create-or-update-kv-entry /api-reference/endpoint/storage/openapi.json put /storage/kv/{bucket}/entries/{key} Creates or updates a key-value entry in the specified bucket # Delete KV entry Source: https://docs.quiva.ai/api-reference/kv-entries/delete-kv-entry /api-reference/endpoint/storage/openapi.json delete /storage/kv/{bucket}/entries/{key} Deletes a specific entry from a key-value bucket # Get KV entry Source: https://docs.quiva.ai/api-reference/kv-entries/get-kv-entry /api-reference/endpoint/storage/openapi.json get /storage/kv/{bucket}/entries/{key} Retrieves a specific entry from a key-value bucket # Get latest KV entries Source: https://docs.quiva.ai/api-reference/kv-entries/get-latest-kv-entries /api-reference/endpoint/storage/openapi.json get /storage/kv/{bucket}/entries Retrieves the latest entries from a key-value bucket # Search KV entries Source: https://docs.quiva.ai/api-reference/kv-search/search-kv-entries /api-reference/endpoint/storage/openapi.json get /storage/kv/{bucket}/search Searches for entries in a KV bucket based on indexed fields # Search KV entries by key pattern Source: https://docs.quiva.ai/api-reference/kv-search/search-kv-entries-by-key-pattern /api-reference/endpoint/storage/openapi.json get /storage/kv/{bucket}/search-by-key Searches for KV entries by key pattern (e.g., 'user*') # Create a new mapping version Source: https://docs.quiva.ai/api-reference/mapping-versions/create-a-new-mapping-version /api-reference/endpoint/gateway/openapi.json post /gateway/{gateway_topic}/mapping/version Creates a new version for an existing mapping. # Delete a mapping version Source: https://docs.quiva.ai/api-reference/mapping-versions/delete-a-mapping-version /api-reference/endpoint/gateway/openapi.json delete /gateway/mapping-version/{gateway_topic}/{version_number} Marks a mapping version as deleted (poison-pilled). # Get a specific mapping version Source: https://docs.quiva.ai/api-reference/mapping-versions/get-a-specific-mapping-version /api-reference/endpoint/gateway/openapi.json get /gateway/mapping-version/{gateway_topic}/{version_number} Retrieves information about a specific mapping version. # Get all mapping versions Source: https://docs.quiva.ai/api-reference/mapping-versions/get-all-mapping-versions /api-reference/endpoint/gateway/openapi.json get /gateway/mapping-version/all Retrieves a list of all available mapping versions. # Get all versions for a mapping Source: https://docs.quiva.ai/api-reference/mapping-versions/get-all-versions-for-a-mapping /api-reference/endpoint/gateway/openapi.json get /gateway/{gateway_topic}/mapping/version Retrieves all versions defined for a specific mapping. # Update a mapping version Source: https://docs.quiva.ai/api-reference/mapping-versions/update-a-mapping-version /api-reference/endpoint/gateway/openapi.json patch /gateway/mapping-version/{gateway_topic}/{version_number} Updates an existing mapping version configuration. # Create a new mapping for a gateway Source: https://docs.quiva.ai/api-reference/mappings/create-a-new-mapping-for-a-gateway /api-reference/endpoint/gateway/openapi.json post /gateway/{gateway_topic}/mapping Creates a new path mapping within the specified gateway. # Delete a mapping Source: https://docs.quiva.ai/api-reference/mappings/delete-a-mapping /api-reference/endpoint/gateway/openapi.json delete /gateway/{gateway_topic}/mapping/details Marks a gateway mapping as deleted (poison-pilled). # Get a specific mapping Source: https://docs.quiva.ai/api-reference/mappings/get-a-specific-mapping /api-reference/endpoint/gateway/openapi.json get /gateway/{gateway_topic}/mapping/details Retrieves information about a specific gateway mapping. # Get all mappings Source: https://docs.quiva.ai/api-reference/mappings/get-all-mappings /api-reference/endpoint/gateway/openapi.json get /gateway/mapping/all Retrieves a list of all available gateway mappings. # Get all mappings for a gateway Source: https://docs.quiva.ai/api-reference/mappings/get-all-mappings-for-a-gateway /api-reference/endpoint/gateway/openapi.json get /gateway/{gateway_topic}/mapping Retrieves all mappings defined for the specified gateway. # Update a mapping Source: https://docs.quiva.ai/api-reference/mappings/update-a-mapping /api-reference/endpoint/gateway/openapi.json patch /gateway/{gateway_topic}/mapping/details Updates an existing gateway mapping configuration. # Update mesh nodes Source: https://docs.quiva.ai/api-reference/mesh-nodes/update-mesh-nodes /api-reference/endpoint/accounts/openapi.json put /accounts/mesh-nodes Updates the number of mesh nodes and the available regions # Deletes stream messages Source: https://docs.quiva.ai/api-reference/messages/deletes-stream-messages /api-reference/endpoint/storage/openapi.json delete /storage/streams/{name}/messages Deletes messages from a stream by sequence(seq). # Publish message to stream Source: https://docs.quiva.ai/api-reference/messages/publish-message-to-stream /api-reference/endpoint/storage/openapi.json post /storage/streams/{name}/messages Publishes a message to a stream # Get current metric value Source: https://docs.quiva.ai/api-reference/metrics/get-current-metric-value /api-reference/endpoint/monitoring/openapi.json get /monitoring/metrics/{metric}/current Retrieves the latest value for a specific metric # Get metric statistics over time Source: https://docs.quiva.ai/api-reference/metrics/get-metric-statistics-over-time /api-reference/endpoint/monitoring/openapi.json get /monitoring/metrics/{metric}/stats Retrieves metric values over a specified time range with a given interval # Get multiple current metric values Source: https://docs.quiva.ai/api-reference/metrics/get-multiple-current-metric-values /api-reference/endpoint/monitoring/openapi.json post /monitoring/metrics/current Retrieves the latest values for multiple metrics simultaneously # Begin passkey registration Source: https://docs.quiva.ai/api-reference/multi-factor-authentication/begin-passkey-registration /api-reference/endpoint/accounts/openapi.json post /accounts/begin-passkey-registration Initiates the passkey registration process # Complete passkey registration Source: https://docs.quiva.ai/api-reference/multi-factor-authentication/complete-passkey-registration /api-reference/endpoint/accounts/openapi.json post /accounts/complete-passkey-registration Completes the passkey registration process # Confirm TOTP registration Source: https://docs.quiva.ai/api-reference/multi-factor-authentication/confirm-totp-registration /api-reference/endpoint/accounts/openapi.json post /accounts/confirm-totp-registration Confirms a TOTP registration by validating a code # Generate TOTP key Source: https://docs.quiva.ai/api-reference/multi-factor-authentication/generate-totp-key /api-reference/endpoint/accounts/openapi.json post /accounts/generate-totp-key Generates a TOTP key for multi-factor authentication # Get user MFA info Source: https://docs.quiva.ai/api-reference/multi-factor-authentication/get-user-mfa-info /api-reference/endpoint/accounts/openapi.json get /accounts/user-mfa-info Retrieves MFA information for a user # Issue recovery codes Source: https://docs.quiva.ai/api-reference/multi-factor-authentication/issue-recovery-codes /api-reference/endpoint/accounts/openapi.json post /accounts/recovery-codes Generates new recovery codes for a user # Remove all passkeys Source: https://docs.quiva.ai/api-reference/multi-factor-authentication/remove-all-passkeys /api-reference/endpoint/accounts/openapi.json delete /accounts/all-passkeys Removes all passkeys for a user # Remove passkey Source: https://docs.quiva.ai/api-reference/multi-factor-authentication/remove-passkey /api-reference/endpoint/accounts/openapi.json delete /accounts/passkey Removes a specific passkey # Remove TOTP Source: https://docs.quiva.ai/api-reference/multi-factor-authentication/remove-totp /api-reference/endpoint/accounts/openapi.json delete /accounts/totp Disables TOTP authentication for a user # Update MFA settings Source: https://docs.quiva.ai/api-reference/multi-factor-authentication/update-mfa-settings /api-reference/endpoint/accounts/openapi.json patch /accounts/mfa-settings Updates multi-factor authentication settings for a user # View recovery codes Source: https://docs.quiva.ai/api-reference/multi-factor-authentication/view-recovery-codes /api-reference/endpoint/accounts/openapi.json get /accounts/recovery-codes Retrieves existing recovery codes for a user # Create a new node Source: https://docs.quiva.ai/api-reference/nodes/create-a-new-node /api-reference/endpoint/hub-flows/openapi.json post /hub/nodes Creates a new node template within a collection # Delete a node Source: https://docs.quiva.ai/api-reference/nodes/delete-a-node /api-reference/endpoint/hub-flows/openapi.json delete /hub/nodes/{collection_topic}/{node_topic} Deletes a specific node by subject # Get a node Source: https://docs.quiva.ai/api-reference/nodes/get-a-node /api-reference/endpoint/hub-flows/openapi.json get /hub/nodes/{collection_topic}/{node_topic} Retrieves a specific node by subject # List node templates Source: https://docs.quiva.ai/api-reference/nodes/list-node-templates /api-reference/endpoint/hub-flows/openapi.json get /hub/nodes Retrieves a list of node templates, optionally filtered by collection and type # Update a node Source: https://docs.quiva.ai/api-reference/nodes/update-a-node /api-reference/endpoint/hub-flows/openapi.json patch /hub/nodes/{collection_topic}/{node_topic} Updates an existing node's properties # Update multiple nodes Source: https://docs.quiva.ai/api-reference/nodes/update-multiple-nodes /api-reference/endpoint/hub-flows/openapi.json patch /hub/nodes/batch Updates multiple existing nodes in a single request # Delete an OAuth configuration. Source: https://docs.quiva.ai/api-reference/oauth/delete-an-oauth-configuration /api-reference/endpoint/hub-flows/openapi.json delete /hub/integrations/{integration_id}/oauth-config [quiva.ai admins only] Deletes an OAuth configuration for a specific integration # Delete OAuth connection Source: https://docs.quiva.ai/api-reference/oauth/delete-oauth-connection /api-reference/endpoint/hub-flows/openapi.json delete /hub/integrations/{integration_id}/oauth-connections/{connection_id} Delete an OAuth connection for a specific integration # Get integration connections Source: https://docs.quiva.ai/api-reference/oauth/get-integration-connections /api-reference/endpoint/hub-flows/openapi.json get /hub/integrations/{integration_id}/oauth-connections Gets integration connections by integration_id # Get OAuth configuration Source: https://docs.quiva.ai/api-reference/oauth/get-oauth-configuration /api-reference/endpoint/hub-flows/openapi.json get /hub/integrations/{integration_id}/oauth-config Get OAuth configuration for a specific integration # Get OAuth connection Source: https://docs.quiva.ai/api-reference/oauth/get-oauth-connection /api-reference/endpoint/hub-flows/openapi.json get /hub/integrations/{integration_id}/oauth-token/{connection_id} Get OAuth connection details for a specific integration and connection # List OAuth configurations Source: https://docs.quiva.ai/api-reference/oauth/list-oauth-configurations /api-reference/endpoint/hub-flows/openapi.json get /hub/integrations/oauth-config Get list of all OAuth configurations # List OAuth connections Source: https://docs.quiva.ai/api-reference/oauth/list-oauth-connections /api-reference/endpoint/hub-flows/openapi.json get /hub/integrations/oauth-connections Get list of OAuth connections for the authenticated user # OAuth callback Source: https://docs.quiva.ai/api-reference/oauth/oauth-callback /api-reference/endpoint/hub-flows/openapi.json get /hub/integrations/oauth-callback Handle OAuth callback after user authorization # Refresh OAuth token Source: https://docs.quiva.ai/api-reference/oauth/refresh-oauth-token /api-reference/endpoint/hub-flows/openapi.json get /hub/integrations/{integration_id}/oauth-refresh-token/{connection_id} Refresh an OAuth access token using the refresh token # Update an OAuth configuration. Source: https://docs.quiva.ai/api-reference/oauth/update-an-oauth-configuration /api-reference/endpoint/hub-flows/openapi.json put /hub/integrations/{integration_id}/oauth-config [quiva.ai admins only] Update the OAuth configuration for a specific integration - full configuration replacement # Create object store bucket Source: https://docs.quiva.ai/api-reference/object-store/create-object-store-bucket /api-reference/endpoint/storage/openapi.json post /storage/obj Creates a new object store bucket # Create or update object store Source: https://docs.quiva.ai/api-reference/object-store/create-or-update-object-store /api-reference/endpoint/storage/openapi.json put /storage/obj/{bucket} Creates a new object store bucket if one does not exist or updates it otherwise # Delete an object store bucket Source: https://docs.quiva.ai/api-reference/object-store/delete-an-object-store-bucket /api-reference/endpoint/storage/openapi.json delete /storage/obj/{bucket} Deletes an object store bucket and all its contents # Delete an object store entry Source: https://docs.quiva.ai/api-reference/object-store/delete-an-object-store-entry /api-reference/endpoint/storage/openapi.json delete /storage/obj/{bucket}/entries/{key} Deletes a specific entry from an object store bucket # Get an object store bucket Source: https://docs.quiva.ai/api-reference/object-store/get-an-object-store-bucket /api-reference/endpoint/storage/openapi.json get /storage/obj/{bucket} Retrieves detailed information about a specific object store, including its configuration and current state. # List all object store buckets Source: https://docs.quiva.ai/api-reference/object-store/list-all-object-store-buckets /api-reference/endpoint/storage/openapi.json get /storage/obj Retrieves a list of all object store buckets # List object store entries Source: https://docs.quiva.ai/api-reference/object-store/list-object-store-entries /api-reference/endpoint/storage/openapi.json get /storage/obj/{bucket}/entries Retrieves entries from an object store bucket # Purge object store bucket Source: https://docs.quiva.ai/api-reference/object-store/purge-object-store-bucket /api-reference/endpoint/storage/openapi.json delete /storage/obj/{bucket}/purge Purges all data from an object store bucket while keeping the bucket configuration # Cancel subscription Source: https://docs.quiva.ai/api-reference/payment/cancel-subscription /api-reference/endpoint/accounts/openapi.json delete /accounts/payment/subscription Cancels the current subscription # Create free subscription Source: https://docs.quiva.ai/api-reference/payment/create-free-subscription /api-reference/endpoint/accounts/openapi.json post /accounts/payment/free-subscription Creates a free subscription # Get checkout URL Source: https://docs.quiva.ai/api-reference/payment/get-checkout-url /api-reference/endpoint/accounts/openapi.json post /accounts/payment/checkout-url Generates a Stripe checkout URL for subscription # Get dashboard URL Source: https://docs.quiva.ai/api-reference/payment/get-dashboard-url /api-reference/endpoint/accounts/openapi.json get /accounts/payment/dashboard-url Generates a Stripe customer portal URL # Get prices Source: https://docs.quiva.ai/api-reference/payment/get-prices /api-reference/endpoint/accounts/openapi.json get /accounts/payment/prices Retrieves all available prices # Get products Source: https://docs.quiva.ai/api-reference/payment/get-products /api-reference/endpoint/accounts/openapi.json get /accounts/payment/products Retrieves all available products # Get subscription Source: https://docs.quiva.ai/api-reference/payment/get-subscription /api-reference/endpoint/accounts/openapi.json get /accounts/payment/subscription Retrieves the current subscription # Get subscription items Source: https://docs.quiva.ai/api-reference/payment/get-subscription-items /api-reference/endpoint/accounts/openapi.json get /accounts/payment/subscription-items Retrieves all items in the current subscription # Process payment webhook Source: https://docs.quiva.ai/api-reference/payment/process-payment-webhook /api-reference/endpoint/accounts/openapi.json post /accounts/payment/webhook Processes payment webhooks from Stripe # Update subscription Source: https://docs.quiva.ai/api-reference/payment/update-subscription /api-reference/endpoint/accounts/openapi.json patch /accounts/payment/subscription Updates the current subscription # Search an index Source: https://docs.quiva.ai/api-reference/search/search-an-index /api-reference/endpoint/storage/openapi.json get /storage/search/{index} Performs a search on an index # Create or update a secret Source: https://docs.quiva.ai/api-reference/secrets/create-or-update-a-secret /api-reference/endpoint/secrets-manager/openapi.json put /secrets Stores a new secret or updates an existing one in the secrets bucket # Delete a secret Source: https://docs.quiva.ai/api-reference/secrets/delete-a-secret /api-reference/endpoint/secrets-manager/openapi.json delete /secrets/{key} Removes a specific secret by its key from the secrets bucket # Get a secret Source: https://docs.quiva.ai/api-reference/secrets/get-a-secret /api-reference/endpoint/secrets-manager/openapi.json get /secrets/{key} Retrieves a specific secret by its key from the secrets bucket # List secrets Source: https://docs.quiva.ai/api-reference/secrets/list-secrets /api-reference/endpoint/secrets-manager/openapi.json get /secrets Retrieves a list of secrets from the secrets bucket. By default, only key names are returned unless reveal=true is specified. # Search stream messages by subject pattern Source: https://docs.quiva.ai/api-reference/stream-search/search-stream-messages-by-subject-pattern /api-reference/endpoint/storage/openapi.json get /storage/streams/{stream}/search-by-subject Searches for stream messages by subject pattern # Create a stream trigger Source: https://docs.quiva.ai/api-reference/stream-triggers/create-a-stream-trigger /api-reference/endpoint/trigger/openapi.json post /trigger/stream Creates a new stream trigger for data processing and routing. A stream trigger connects a data source to a target, enabling real-time data flow between components. **Usage Notes:** - The system automatically creates necessary import/export pairs - Stream triggers define both source and target configurations - The source configuration specifies where data originates - The target configuration specifies where data should be delivered - Response types can be Singleton (single response), Stream (continuous), or Chunked (paginated) - Source and target types determine how data flows through the system - The trigger is active immediately after creation - Both synchronous and asynchronous data processing patterns are supported # Create a new stream Source: https://docs.quiva.ai/api-reference/streams/create-a-new-stream /api-reference/endpoint/storage/openapi.json post /storage/streams Creates a new stream for message publishing # Create or update a stream Source: https://docs.quiva.ai/api-reference/streams/create-or-update-a-stream /api-reference/endpoint/storage/openapi.json put /storage/streams/{name} Creates a new stream or updates an existing stream with the provided configuration. If the stream does not exist, it will be created. If it exists, it will be updated with the new configuration. Note that some properties cannot be changed after a stream is created, such as the storage type. If you attempt to change these properties, the request will fail with a 400 error. # Delete a stream Source: https://docs.quiva.ai/api-reference/streams/delete-a-stream /api-reference/endpoint/storage/openapi.json delete /storage/streams/{name} Deletes a stream with all the messages and consumers attached to it. # Get information about a stream Source: https://docs.quiva.ai/api-reference/streams/get-information-about-a-stream /api-reference/endpoint/storage/openapi.json get /storage/streams/{name} Retrieves detailed information about a specific stream, including its configuration and current state. This endpoint provides metrics such as the number of messages, bytes used, first and last sequence numbers, and more. # List all streams Source: https://docs.quiva.ai/api-reference/streams/list-all-streams /api-reference/endpoint/storage/openapi.json get /storage/streams Retrieves a list of all streams # Purge stream messages Source: https://docs.quiva.ai/api-reference/streams/purge-stream-messages /api-reference/endpoint/storage/openapi.json delete /storage/stream/{name}/purge Purges messages from a stream with optional filters # Delete a trigger Source: https://docs.quiva.ai/api-reference/triggers/delete-a-trigger /api-reference/endpoint/trigger/openapi.json delete /trigger/triggers/{trigger_type}/{topic} Marks a trigger as deleted (poison-pilled). This operation logically removes the trigger from the system while preserving the record. **Usage Notes:** - This operation cannot be undone through the API - The `trigger_type` and `topic` parameters together identify the exact trigger to delete - The system constructs the full subject using these parameters: `ms.trigger.{trigger_type}.{topic}` - For gateway triggers, associated mappings remain but are no longer accessible - For stream triggers, connections between source and target are terminated - In-flight operations may still complete, but new operations will be rejected - The trigger remains in the system's history for audit purposes # Retrieve a specific trigger Source: https://docs.quiva.ai/api-reference/triggers/retrieve-a-specific-trigger /api-reference/endpoint/trigger/openapi.json get /trigger/triggers/{trigger_type}/{topic} Retrieves detailed information about a specific trigger identified by its type and topic. This endpoint returns comprehensive configuration data including all trigger properties, associated resources, and metadata. **Usage Notes:** - The `trigger_type` parameter identifies the category of trigger (gateway, stream, subject) - The `topic` parameter is the unique name identifier for the specific trigger - The system constructs the full subject using these parameters: `ms.trigger.{trigger_type}.{topic}` - Different trigger types will return different sets of properties in the response - The response includes the full subject identifier which may be needed for other operations # Retrieve all triggers Source: https://docs.quiva.ai/api-reference/triggers/retrieve-all-triggers /api-reference/endpoint/trigger/openapi.json get /trigger/triggers Retrieves all triggers matching a specific subject pattern. This endpoint provides a comprehensive view of all available triggers that match the given pattern, including their configurations, types, and associated metadata. **Usage Notes:** - Use wildcards in the subject pattern to match multiple triggers - Results include all trigger types (gateway, stream, subject) - Each trigger includes its complete configuration - The subject pattern forms part of the unique identifier for triggers - The response includes total count of matching triggers # Complete email address change Source: https://docs.quiva.ai/api-reference/user-management/complete-email-address-change /api-reference/endpoint/accounts/openapi.json post /accounts/change-email-address-action Completes an email address change process # Create/Add user Source: https://docs.quiva.ai/api-reference/user-management/createadd-user /api-reference/endpoint/accounts/openapi.json post /accounts/users Creates a new user and sends an invitation email if the user does not exist. Adds the new or existing user to the current account. If the user already exists and is already a member of the account, an error is returned - If the user already exists but isn't a member of the account, they are added without creating a new user record - The user won't be able to access the account until they complete the activation process by setting their password - For security, new users must set their own password via the activation link - The function schedules cleanup tasks to remove unactivated user accounts after a certain period of time (48 hours) # Delete profile image Source: https://docs.quiva.ai/api-reference/user-management/delete-profile-image /api-reference/endpoint/accounts/openapi.json delete /accounts/profile-image Deletes the user's profile image # Delete user Source: https://docs.quiva.ai/api-reference/user-management/delete-user /api-reference/endpoint/accounts/openapi.json delete /accounts/user Deletes the current user # Get user login data Source: https://docs.quiva.ai/api-reference/user-management/get-user-login-data /api-reference/endpoint/accounts/openapi.json get /accounts/user-login-data Retrieves login data for the current user # List users by account Source: https://docs.quiva.ai/api-reference/user-management/list-users-by-account /api-reference/endpoint/accounts/openapi.json get /accounts/users/list Lists all users in the current account # Remove user from account Source: https://docs.quiva.ai/api-reference/user-management/remove-user-from-account /api-reference/endpoint/accounts/openapi.json delete /accounts/remove-user Removes a user from the current account # Request email address change Source: https://docs.quiva.ai/api-reference/user-management/request-email-address-change /api-reference/endpoint/accounts/openapi.json post /accounts/request-email-address-change Initiates an email address change process # Suspend user Source: https://docs.quiva.ai/api-reference/user-management/suspend-user /api-reference/endpoint/accounts/openapi.json patch /accounts/suspend-user Suspends or unsuspends a user in an account. Suspended users have no access to any of the resources and services associated with the account # Update user Source: https://docs.quiva.ai/api-reference/user-management/update-user /api-reference/endpoint/accounts/openapi.json patch /accounts/user Updates user information # Update user role Source: https://docs.quiva.ai/api-reference/user-management/update-user-role /api-reference/endpoint/accounts/openapi.json patch /accounts/role Updates a user's role in an account # Upload profile image Source: https://docs.quiva.ai/api-reference/user-management/upload-profile-image /api-reference/endpoint/accounts/openapi.json post /accounts/profile-image Uploads a profile image for the user # Create a new workflow Source: https://docs.quiva.ai/api-reference/workflows/create-a-new-workflow /api-reference/endpoint/hub-flows/openapi.json post /hub/workflows Creates a new workflow for defining node sequences # Delete a workflow Source: https://docs.quiva.ai/api-reference/workflows/delete-a-workflow /api-reference/endpoint/hub-flows/openapi.json delete /hub/workflows/{collection_topic}/{flow_topic} Deletes a specific workflow by subject # Get a workflow Source: https://docs.quiva.ai/api-reference/workflows/get-a-workflow /api-reference/endpoint/hub-flows/openapi.json get /hub/workflows/{collection_topic}/{flow_topic} Retrieves a specific workflow by subject # List workflows Source: https://docs.quiva.ai/api-reference/workflows/list-workflows /api-reference/endpoint/hub-flows/openapi.json get /hub/workflows Retrieves a list of workflows, optionally filtered by collection and type # Publish a workflow Source: https://docs.quiva.ai/api-reference/workflows/publish-a-workflow /api-reference/endpoint/hub-flows/openapi.json post /hub/workflows/publish Publishes a draft workflow to make it available for execution # Run a workflow Source: https://docs.quiva.ai/api-reference/workflows/run-a-workflow /api-reference/endpoint/hub-flows/openapi.json post /hub/workflows/{collection_topic}/{flow_topic}/run Executes a workflow with the specified trigger data # Update a workflow Source: https://docs.quiva.ai/api-reference/workflows/update-a-workflow /api-reference/endpoint/hub-flows/openapi.json patch /hub/workflows/{collection_topic}/{flow_topic} Updates an existing draft workflow's properties # Workflow Configuration History Source: https://docs.quiva.ai/api-reference/workflows/workflow-configuration-history /api-reference/endpoint/hub-flows/openapi.json get /hub/workflows/{collection_topic}/{flow_topic}/history Get workflow configuration history # Assistant Best Practices Source: https://docs.quiva.ai/assistants/best-practices Proven patterns and strategies for building high-performing assistants # Assistant Best Practices This guide compiles proven patterns, strategies, and lessons learned from building successful agents. Follow these practices to create agents that perform reliably, cost-effectively, and delight users. ## Design Principles ### Start Simple, Add Complexity
Behavior: The flow waits for the agent to finish processing before moving to the next step or returning a response.
Use when:
- Building API endpoints that need to return agent results
- Creating real-time chat experiences
- Form submissions that show agent responses
- Any synchronous interaction where users wait for results
Pros:
- ✅ Users get immediate responses
- ✅ Easier to handle errors and retries
- ✅ Simpler flow logic
- ✅ Can pass agent output to subsequent steps
Cons:
- ⚠️ User waits for agent to complete (may be 5-30 seconds)
- ⚠️ Request times out if agent takes too long
- ⚠️ Not suitable for very long-running tasks
Behavior: The flow responds immediately and the agent runs asynchronously. The trigger doesn't wait for completion.
Use when:
- Processing emails or documents
- Running scheduled tasks
- Handling webhooks from external systems
- Long-running analysis or content generation
- Any async workflow where immediate response isn't needed
Pros:
- ✅ Instant trigger response (no waiting)
- ✅ Can handle very long-running tasks
- ✅ Won't timeout on complex processing
- ✅ Better for high-volume processing
Cons:
- ⚠️ Can't return agent results to trigger
- ⚠️ More complex to notify users of completion
- ⚠️ Harder to handle errors in real-time
- ⚠️ Need separate mechanism to show results