Data residency by AI provider

What the three layers of residency actually mean

Before reading the per-provider tables, internalize these distinctions. Every vendor slides between them.

• Inference geo — the data center where the GPU actually generated your response. This is the layer most exposed to compelled access by the host country's government, because the cleartext token stream is in memory there.
• Storage geo — where logs, abuse-monitoring buffers, and any persisted artifacts live at rest. Usually 30 days by default on commercial APIs, often reducible to 0 days under a ZDR addendum.
• Endpoint-services geo — where supporting infrastructure runs: tokenizers, content-safety classifiers, image pipelines, retrieval indexes, code interpreters. Easy to forget. A 'EU-resident' deployment that calls a US classifier is, in practice, sending your prompt to the US.
• Training reuse — separate from residency, but always asked in the same breath. Default on enterprise API tiers from Anthropic, OpenAI, Google, AWS, Mistral, and Cohere: customer prompts are NOT used to train base models. Verify on the order form.
• BAA scope — a Business Associate Agreement is a contract under HIPAA. The fact that a vendor 'has' a BAA does not mean every product is covered. Bedrock got added to AWS's HIPAA-eligible list on 10 February 2026; some Bedrock features remain out of scope. Read the schedule.

Provider index — at a glance

Quick scan of where each provider stood as of June 2026, drawn from their published trust portals and docs. Treat 'yes' as 'available under enterprise contract' unless noted otherwise.

Anthropic — what is actually shipping

OpenAI — global storage, US inference

Google Vertex AI and AWS Bedrock — the partner-operated lanes

If you want Claude or a major open model with the strictest residency posture available today, you almost certainly want it through one of the two big clouds. Both inherit hyperscaler compliance.

Mistral — the EU-native default

Cohere, Together, Fireworks, Replicate — the inference-platform layer

These four sit one rung below the hyperscalers in posture. All four are usable in production. Compliance scope varies sharply.

• Cohere — primary infrastructure on GCP US-Central. Annual SOC 2 Type II. HIPAA compliant with a standard BAA available. Multi-jurisdictional DPA incorporating the EU Standard Contractual Clauses (approved 4 June 2021). The honest caveat: Cohere does not currently offer true multi-region data residency outside the US — a meaningful limit if you need EU-only inference and not just EU-resident logging.
• Together AI — US-primary with EU residency live in Sweden since September 2025. Enterprise SOC 2 Type II, HIPAA on enterprise contracts, dedicated data-residency controls and ZDR on the enterprise tier. Voice platform layer adds its own enterprise compliance bundle.
• Fireworks AI — multi-region global fleet across US, Frankfurt, Iceland, Tokyo. Fully HIPAA-eligible, SOC 2 Type II, GDPR-compliant. Zero data retention by default on the enterprise product. Airgapped EKS deployment is on offer for the most regulated customers.
• Replicate — developer-and-community surface. Strong model catalog, easy experimentation. As of June 2026 not the right choice if you need formal enterprise BAA and FedRAMP — those are simply not the posture they sell. Use Replicate for prototyping and route the production workload to a hyperscaler or to Together/Fireworks.

Groq and Cerebras — the dedicated-silicon lane

Where buyers reliably get burned

How AtomEons handles this internally