AI data residency map
Thirty providers, three separate questions — inference, storage and training — and the exact wording to ask for before you sign.
Filter the map
Showing 30 of 30 providers.
Inference, storage and training are three questions
“Where is my data?” is really three questions, and a provider can answer them differently. Inference is where the prompt is processed. Storage is where logs, files and fine-tuning data sit at rest. Training is whether the provider uses your content to improve a model. A provider can process in the EU, store in the US, and train on nothing — and that combination may or may not satisfy your regulator.
| Question | What to ask | Where the answer lives |
|---|---|---|
| Inference | Which region processes the request, and can I pin it? | API docs, region list, order form |
| Storage | Where are logs and files at rest, and for how long? | DPA, retention schedule, trust centre |
| Training | Is my content used for training by default, and can I opt out? | Terms, DPA, enterprise addendum |
| Subprocessors | Which third parties touch the data, and where are they? | Subprocessor list, DPA annex |
| Support access | Can support staff in another country read my data? | Access-control annex |
The map: 30 providers
| Provider | Inference | Storage | Training on API data | Bucket | Residency option |
|---|---|---|---|---|---|
| OpenAI | US; regional processing endpoints | Customer-chosen region (regional endpoints) | No (API default) | Multi-region | Regional processing endpoints; 10% uplift on models released after 5 Mar 2026 |
| Anthropic | US; US-only inference | US | No (API default) | US | US-only inference at 1.1× input and output pricing |
| Google Vertex AI | Multi-region | Customer-chosen region | No (API default) | Multi-region | Data residency at rest and in processing |
| Microsoft Azure OpenAI | Multi-region | Customer-chosen region | No (API default) | Multi-region | Data residency and EU Data Boundary |
| Amazon Bedrock | Multi-region | Customer-chosen region | No (API default) | Multi-region | Region-pinned inference and storage |
| Mistral | EU | EU | No (API terms) | EU | EU-hosted API |
| Cohere | US, EU | US, EU | No (API terms) | US + EU | SaaS and private VPC deployment |
| AI21 | US | US | Check terms | US | Private deployment available |
| xAI | US | US | Check terms | US | US API |
| Groq | US | US | Check terms | US | US-hosted inference |
| Together AI | US | US | Check terms | US | US-hosted inference |
| Fireworks AI | US | US | Check terms | US | US-hosted inference |
| DeepInfra | US | US | Check terms | US | US-hosted inference |
| Replicate | US | US | Check terms | US | US-hosted inference |
| Hugging Face | US, EU | US, EU | Check terms | US + EU | Inference endpoints in US and EU |
| NVIDIA NIM | Customer cloud or on-prem | Customer | No (self-hosted) | Self-hosted | Self-hosted microservice |
| IBM watsonx | Multi-region | Customer-chosen | No (enterprise terms) | Multi-region | Regional deployment |
| OVHcloud AI | EU (France) | EU (France) | No (API terms) | EU | EU-hosted |
| Scaleway | EU (France) | EU (France) | No (API terms) | EU | EU-hosted |
| Nebius | EU (Finland) | EU (Finland) | No (API terms) | EU | EU-hosted |
| CoreWeave | US, EU | Customer-chosen | No (self-hosted) | US + EU | GPU cloud in US and EU |
| Plugsky | Cloud, VPC, on-prem, air-gapped | Deployment-dependent | No (API default) | Multi-region | Deployment model sets the region |
| OpenRouter | Global routing | Provider-dependent | Provider-dependent | Global routing | No single residency guarantee |
| Perplexity | US | US | Check terms | US | US API |
| DeepSeek | China | China | Check terms | China | China-hosted API |
| Alibaba Model Studio | China, Singapore | China, Singapore | Check terms | China + APAC | Regional endpoints |
| Moonshot (Kimi) | China | China | Check terms | China | China-hosted API |
| Zhipu (GLM) | China | China | Check terms | China | China-hosted API |
| Baidu Qianfan | China | China | Check terms | China | China-hosted API |
| Meta Llama (hosted) | Host-dependent | Host-dependent | Host-dependent | Self-hosted | Depends on the host, not Meta |
What “EU residency” usually means
Three different promises get sold as EU residency. The weakest is an EU endpoint that routes to a global backend: the request lands in Frankfurt and is processed somewhere else. The middle promise is EU processing with global storage, which satisfies some regulators and not others. The strongest is EU processing and EU storage, with support access restricted and subprocessors disclosed. Ask which one is in the contract, not which one is on the marketing page.
“US-only inference” is the same: Anthropic prices it at 1.1× input and output, and OpenAI’s regional processing endpoints carry a 10% uplift on models released after 5 March 2026. Residency is a cost line before it is a compliance line.
The 20 regions we track
The filter buckets collapse 20 regions into six groups so the table stays readable. The full list behind the groups is: US East, US West, Canada Central, Ireland, Frankfurt, Paris, London, Zurich, UAE, Saudi Arabia, Mumbai, Singapore, Tokyo, Seoul, Sydney, São Paulo, Johannesburg, China North, Hong Kong, and “customer cloud” for self-hosted deployments.
A listed region is not always available for every model or tier. Check the model page, not only the region page.
How to verify a claim
- Find the region list for the exact model and endpoint you call, not the provider’s global list.
- Read the DPA for storage location, retention and subprocessors. The DPA governs when the marketing page and the contract disagree.
- Ask about support access. A regional data plane with a global support team is not regional.
- Test it. Send a request with a known marker and check the response headers, logs and egress region.
- Record the date. Residency options change with model launches.
Methodology
Rows are compiled from provider documentation, pricing pages and DPAs checked on 20 November 2026. The bucket is the narrowest published inference region, not the broadest. Where a provider publishes nothing, the row says “check terms” rather than guessing.
The map is not a legal opinion and does not determine whether a transfer mechanism is valid. It is a starting list for a data-protection review, with the questions that review needs to ask. For the next step, run a single provider through the data residency checker, then read the residency guide and the enterprise deployment notes.
Sources
Links checked 20 November 2026. We cite the primary page for each figure. Where a provider publishes a price behind a login, we say so and do not guess.
- OpenAI API pricing — regional processing endpoints and the 10% uplift rule
- Anthropic pricing — US-only inference at 1.1× input and output pricing
- Google Cloud regions — the region inventory behind the 20-region list
- AWS global infrastructure — region and availability-zone reference
FAQ
Does EU residency mean my data never leaves the EU?
Not automatically. An EU endpoint can route to a global backend, and support access may come from another country. Ask for EU processing and EU storage, restricted support access and a subprocessor list, and check that all four are in the contract.
Which providers offer US-only inference?
Anthropic publishes a US-only inference option at 1.1× input and output pricing. OpenAI offers regional processing endpoints with a 10% uplift on models released after 5 March 2026. Other providers vary by model and contract.
Is training on API data the same as storage location?
No. A provider can store your data in the EU and still use it for model improvement, or store it in the US and never train on it. Training, storage and inference are separate commitments and should be read separately in the DPA.
Why do some rows say “check terms”?
Because we could not verify a public default for that provider on 20 November 2026. It does not mean the provider trains on your data; it means the answer is in the contract, and we do not guess.
How often is the map updated?
Monthly, and when a provider changes a region list or a residency option. Each row carries the date it was checked. Region availability is per model, so re-check the model page before relying on a row.
Run the same check on your own traffic
Plugsky's API is OpenAI-compatible, with published plan caps and 39 models on one key. Start free with no card.
Start free → See plans