Sovereign inference and the end of US default

Position, signed by Priya Raman, 10 July 2026. This is the lab's view on the us-non-dominance field: whether, and how fast, the assumption that inference means a US model on US infrastructure stops holding for our clients. It draws on three validation runs, the gateway and open-weight experiments (x-gateway-routing, x-open-weight-parity), the TEE experiment in flight (x-tee-inference), and the demand signal from banking and government engagements.

What is demonstrated. The open-weight tier — Qwen, DeepSeek, Llama, Mistral, Gemma — is at parity with frontier on classification and extraction on our evals and can be hosted in an AU region by three providers today (c-us-non-dominance-1, c-open-weight-models-1). Two of the three best open-weight models are Chinese; one is French. Clients have noticed. In the first half of 2026, four banking and government RFPs required an AU-resident inference option, up from zero in 2025 (c-us-non-dominance-2). Sovereignty is now a procurement criterion, not a talking point. And the capability ceiling is still American: on agentic loops the frontier lead is 11–18 points and belongs to Anthropic, OpenAI and Google DeepMind.

What is hype. 'Sovereign AI' as a national model programme. The Australian-trained frontier model is not coming, and nothing about the field's economics suggests it should. Also hype: on-prem as the sovereignty answer. g-onprem-h100-cluster is the graveyard entry, and the lesson is that residency does not require ownership; it requires a boundary. Confidential inference on a TEE in an AU region may draw that boundary more cheaply than a rack does, which is what x-tee-inference is testing. And it is hype that Chinese open-weight models are a security problem by construction; the weights are inspectable, the hosting is ours, and the assessment is the same as for any dependency (c-us-non-dominance-3).

What would have to be true for the position to be wrong. If the US frontier labs offered AU-resident inference with contractual data-residency at parity pricing — one has announced it, none has delivered it — much of the demand for the open-weight tier would move back. If open-weight parity reversed on the tasks where it currently holds, the routing decision would collapse to 'frontier for everything'. If export controls or an equivalent AU regulatory move restricted the Chinese open-weight models, the tier would thin to Mistral and Meta and lose its parity. Each of these is watchable, and the field's depends-on edges carry the triggers.

What we would do. Design every new pattern for a model garden with at least one AU-resident open-weight route, through a gateway we control (r-gateway-default). Treat sovereignty as a routing policy — this task type stays inside this boundary — rather than a hardware decision. Keep the position's dependency on x-tee-inference explicit: if confidential inference works for banking PII, it becomes the recommended residency route and on-prem stays in the graveyard. Watch for the announced AU-resident frontier endpoint and re-run the position when it ships.

Strength is moderate. The demand evidence is strong and the parity evidence is tested, but the position rests on a trajectory — that the residency requirement spreads and the open-weight tier holds — and trajectories are what positions are for. Review in a year or when the TEE experiment concludes.