Local and in-region AI hosting
The data law decides where models can run. The trigger for self-hosted or in-region inference is simple: does identifiable patient data leave the jurisdiction to reach a model processor, and how strict is that jurisdiction's residency rule. That sorts the six markets into three tiers.
Localisation effectively mandatory. UAE
Health data cannot leave the country. Run any identifiable patient inference on a self-hosted model in-country, or on in-UAE managed infrastructure. Global external APIs are acceptable only for fully de-identified or non-patient content, for example generating generic education copy offline.
Accountability or transfer-gated. Australia, EU, UK
External APIs are usable only if data stays in-region through a regional endpoint, or is covered by a transfer mechanism (SCCs, the UK IDTA) plus consent. The cleanest path is in-region managed inference, Vertex in australia-southeast1 for Australia, an EU region for the EU and UK, or self-hosting. Australia's APP 8 specifically rewards keeping data onshore.
Contract or consent-gated. India, United States
Cross-border transfer is allowed in India unless a destination is restricted, and is BAA-gated in the US. A managed API is acceptable with proper consent and, in the US, a complete BAA chain. In-region processing is still preferred for trust and to satisfy clinic contracts.
Which model runs where
The model choice is bounded by the same green and red line, not just by cost. Clinical-sounding output is an asset on the education surface and a liability on the routing path.
Patient-facing education (one-to-many). Clinical accuracy helps here, so a medically post-trained model is appropriate for the neutral knowledge surface. It still only ever educates. It never assesses an individual.
Routing, qualification, and intake paths. These must stay non-clinical, so use a general model or deterministic rules, not a medically post-trained one. A medical model's training pulls it toward clinical-sounding answers, which directly conflicts with the non-clinical legal position on these paths. Public-facing decision logic stays rules-based and deterministic, never free model reasoning.
Through the pilot. A managed general model via API is cost-optimal where residency permits. Use regional endpoints and de-identify wherever possible.
Post-pilot and strict markets. Move to self-hosted inference via vLLM on in-region GPU, or in-region managed inference, once volume or residency requires it.
Jurisdictional knowledge stays in retrieval, not weights. State-by-state ART legislation, Medicare funding, ANZARD framing, and the equivalent in each market change regularly. Keep them in the retrieval layer so a model update never freezes stale law into the product.
The stricter the jurisdiction, the more inference must run locally or in-region. UAE forces it outright. Australia, EU, and UK reward it strongly. India and the US allow more latitude but still favour it. The build should assume in-region inference is the default for any identifiable patient data, and treat the global API as the exception reserved for de-identified or non-patient work.