All posts AI Risk & Regulation

EU data residency for Claude, OpenAI, Gemini and Mistral: what each offers and what it does not solve

The short answer

Claude data residency in Europe is not available from Anthropic directly: its API accepts only “global” or “us” as an inference location, and “us” is the only workspace geo for stored data. To keep Claude in the EU, buy it through Amazon Bedrock’s EU inference profile or Google-managed Claude on Google Cloud’s “eu” multi-region endpoint; for Fable only Google works, because Bedrock offers Fable in Europe only on its global profile. Microsoft Foundry offers Claude only as Global Standard or a US data zone.

OpenAI’s API keeps data in the EEA and Switzerland once OpenAI approves a retention control. Azure OpenAI’s EU Data Zone, Google’s EU multi-region and Mistral’s EU endpoint keep inference in Europe, each with its own map of Europe. The Gemini Developer API has no documented regional option, and in Microsoft 365 Copilot, Anthropic models are “currently excluded from the EU Data Boundary”. Checked on 29 October 2026.

Key takeaways

  • Claude’s EU option lives with the cloud provider, not the lab, and on Bedrock it does not cover Fable.
  • “EU” means something different at each vendor. Google excludes Switzerland, OpenAI includes it, and two Microsoft pages describe it differently.
  • Residency covers six classes of data, not one, and the carve-outs cluster on what agents produce.
  • Microsoft 365 Copilot has three switches that move processing out of the EU, two on by default for tenants created after 25 March 2026, plus a separate opt-in for Anthropic models with Data Retention.
  • A US court can still order the data from a US-parented provider holding it in Frankfurt, and GDPR Article 48 does not make that order enforceable in Europe.

Which AI vendors keep data in the EU? Claude, OpenAI, Gemini, Mistral and Copilot compared

“Left outside” is what the vendor’s own page excludes from its residency promise. The last column is what an order could reach, if it reaches the vendor: the CLOUD Act question.

OfferingStored at restWhere inference runsZero or limited retentionLeft outsideWhat a US order could reach
Anthropic API, also Claude Platform on AWS and Anthropic-managed Claude on Google CloudUS only“global” (default) or “us”; no EU; the cloud region does not pin itZDR on eligible features; not Managed Agents or Covered ModelsMemory stores, even with a self-hosted sandboxMemory stores, agent transcripts, Covered Model copies
Claude apps (Team, Enterprise)USUS, Europe, Asia, Australia by default; US-only optionNot ZDR-eligible, except Claude Code on EnterpriseConnectorsStored conversations
Claude on Amazon BedrockYour RegionEU profile across six EU Regions; single-Region in London, Stockholm or Ireland for some models, depending on endpoint; Fable: global onlyPer-Region ZDR; Fable kept up to 30 daysAbuse-detection copies in the destination RegionYour stored data and Fable copies, at AWS
Claude on Google Cloud (Google managed)Your location“eu” multi-region, EU member states only, Fable includedFable and Mythos logged up to 30 days, shared with AnthropicGlobal endpointYour stored data; Fable and Mythos logs, at Google and Anthropic
Claude in Microsoft FoundryYour Azure geography, if hosted on AzureGlobal Standard or US Data Zone; no EU zoneAnthropic’s terms, as independent processorFlagged content and usage metadata go to AnthropicFlagged content at Anthropic; the deployment at Microsoft
OpenAI APIEEA and SwitzerlandIn region on eu.api.openai.com, for listed models and modesEU needs ZDR, modified abuse monitoring, Private Retention or Safety RetentionSystem data; Agents API (beta) is US onlyAbuse logs up to 30 days unless approved; conversations and vector stores until deleted
ChatGPT EnterpriseA supported region you chooseSeparate capability, specified regionsAPI ZDR does not set ChatGPT Work retentionIntegrations, synced indexes, dots (beta)Stored workspace content
Azure OpenAIYour geographyGlobal: any region. EU Data Zone: EU, can include EFTAModified abuse monitoring on applicationGrounding with Bing; new models arrive lateAbuse store, unless modified monitoring is approved
Gemini Developer APINo residency option documentedNo residency option documentedGoogle points to Vertex AINo residency promise to carve fromUploaded files until deleted or expired; search grounding logs 30 days
Gemini on Vertex AIYour locationEU multi-region or an EU regionPer feature, by customer actionGlobal endpoint and its cacheConversation state; search grounding logs up to three days; Deep Research sessions seven days
Gemini in WorkspaceEU data regionIn the EU only on the top Data Regions versionsPrompts and responses kept 90 days to indefinite, set by adminsGemini Notebook, logs, cached contentPrompts and responses, for as long as admins keep them
Mistral APIEU by defaultEU and EFTA on api.eu.mistral.ai onlyZDR on pay-as-you-go, stateless calls onlyControl plane; Agents, Batch and FilesAgent and library data, if US courts have jurisdiction; US endpoint traffic
Mistral Vibe (formerly Le Chat)EU by defaultNo committed location documentedNo ZDR for Vibe Work or Chat; Vibe CLI follows API ZDRFeature-dependent transfersStored conversations, on the same condition
Microsoft 365 CopilotEU Data BoundaryIn the boundary unless flex routing or an Anthropic setting moves itAnthropic models with Data Retention: up to 30 days with Anthropic, longer if flaggedAnthropic models; remote access; Multi-Geo tenantsCopilot data at Microsoft; retention-model copies at Anthropic

Checked on 29 October 2026 against each vendor’s own documentation, linked below. Storage and processing are separate promises almost everywhere.

What does “EU” mean in each vendor’s residency promise?

Something different at almost every vendor, and for a group with Swiss, Norwegian or UK entities this is the decision.

OfferingWhat “EU” coversSwitzerland and Norway
Google “eu” multi-regionEU member states onlyExcluded, with the UK
OpenAI Europe regionEEA and SwitzerlandIncluded
Azure EU Data ZoneEU, and EFTA on one pageDepends on the page
Mistral EU endpointEU and EFTA data centresPossibly
Bedrock EU profileSix EU Regions, plus the calling Region; none for FableZurich only when called from Zurich
Microsoft 365 EU Data BoundaryEU and EFTAIncluded

Microsoft’s deployment types page says the EU Data Zone “can include European Free Trade Association (EFTA) countries and regions such as Norway and Switzerland”; its Azure OpenAI privacy page says EU Data Zone prompts “may be processed in that or any other European Union Member Nation”. Write down which EU a contract means.

Does Claude offer data residency in Europe?

Not from Anthropic itself. Its data residency documentation says “Two independent settings govern this”. Inference geo, set per request, offers only “us” and “global”. Workspace geo “Controls where data is stored at rest and where endpoint processing (such as image transcoding and code execution) occurs”, and “us” is the only option.

The Claude apps follow. Anthropic’s privacy centre says: “Note that data is stored in the US.” By default Anthropic “may route customer traffic to select countries in the US, Europe, Asia and Australia”. Usage-based Enterprise plans can switch on US-only inference, which does not govern connectors. The Team and Enterprise interfaces “are not ZDR-eligible”, except Claude Code on Claude Enterprise (Anthropic).

Agents sit furthest out. The same retention page says “In Claude Managed Agents, no feature is eligible for ZDR or HIPAA readiness”, and the Covered Models (Fable and Mythos) “require 30-day data retention”. Even with a self-hosted sandbox, “the contents of attached memory stores remain stored by Anthropic” (Anthropic).

Anthropic’s regional compliance page adds confusion: its FAQ offers “country-specific deployment options through AWS Bedrock, GCP Vertex, and Microsoft Foundry”, while its own table marks Foundry “Coming soon”. In the cloud documentation, Foundry has no EU option, and “country-specific” holds on Google Cloud only for Sonnet 4.6 and earlier, and on Bedrock only for some models, in London, Stockholm or Ireland.

Two more traps. The Anthropic-operated Claude Platform on AWS is not Bedrock: its AWS region “does not pin where model inference runs”. And Anthropic-managed Claude on Google Cloud follows Anthropic’s rules: “The location is separate from where inference runs”. Only the Google-managed offering has the “eu” endpoint.

Claude on AWS Bedrock, Google Vertex AI or Microsoft Foundry: which keeps data in the EU?

Bedrock and Google-managed Claude on Google Cloud do, except Fable on Bedrock. Microsoft Foundry does not.

Bedrock. The Claude Sonnet 5.5 model card lists an EU geographic profile across Frankfurt, Stockholm, Milan, Spain, Ireland and Paris, and single-Region (In-Region) inference in none of them. On bedrock-runtime, the Opus 5.5 card says In-Region inference with the bare model ID “is available in eu-west-2 (London) only”; the Sonnet 5, Opus 4.6 and Sonnet 4.6 cards mark London in their availability tables but give N/A as the bedrock-runtime In-Region endpoint URL. On bedrock-mantle, the Sonnet 5, Opus 5, Opus 4.8, Opus 4.7 and Haiku 4.5 cards mark In-Region in Stockholm and Ireland. London is outside the EU. Called from Zurich or London, the EU profile adds that Region. “To the extent we store data for abuse detection, your input prompts and output results will be stored in the destination region” (AWS).

Fable has no EU profile on Bedrock. The Fable 5 and Fable 5.1 cards list only a US geographic profile and a global one, so in Europe Fable’s prompts and abuse-detection copy can be processed and stored outside the EU. That copy is not optional: “all traffic will be retained for up to 30 days for automated offline abuse detection” (AWS), with zero retention for eligible Enterprise Frontier Safeguards customers only “through December 31, 2026”, and it is “not shared with third-party model providers”. Other models can use Bedrock’s per-Region zero data retention mode.

Google Cloud. Anthropic’s Google-managed Claude page says “Regional endpoints serve Claude Sonnet 4.6 and earlier models only”. Newer models reach the EU through the “eu” multi-region, which Google’s data residency page, updated on 28 October 2026, lists for Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5 and Fable 5.1. “Geographies outside the European Union political boundary, including the United Kingdom and Switzerland, are excluded from this endpoint”, so a Swiss, Norwegian or UK entity that needs in-country processing gets nothing from it.

Google’s abuse monitoring page logs prompts to the models it designates “Advanced AI” (Fable, Mythos, and Opus 4.7+ or Sonnet 5+ traffic enrolled in Anthropic’s Cyber Verification Program) “for up to 30 days for the sole purpose of monitoring for abuse”, in your selected region, without customer-managed encryption. For Fable 5 and Mythos 5, “as required by Anthropic, you must enable sharing this data with Anthropic for abuse monitoring”. Anthropic’s retention page says retained data on Bedrock and Google Cloud stays within the cloud provider’s environment; AWS agrees, Google’s page does not. Ask both vendors in writing where the shared copy is stored.

Microsoft Foundry. Anthropic’s Foundry page says “Claude is available in Global Standard and US Data Zone Standard deployment types in Foundry resources”, and Microsoft’s partner model tables list no Claude model in a European data zone. Global types “May be processed in any Azure region” (Microsoft). Flagged content and usage metadata go to Anthropic, which “acts as an independent data processor for prompts and outputs associated with Claude models” (Microsoft).

Does OpenAI offer EU data residency for the API and ChatGPT Enterprise?

Yes for both, with conditions. OpenAI’s your-data guide lists a Europe region for the EEA and Switzerland at eu.api.openai.com, where “TLS termination and HTTPS decryption occur within the selected processing region.” The UK region stores but does not process: “Support for regional storage does not imply support for regional processing.”

It is not self-serve: “To use data residency with any region other than the United States, you must be approved for abuse monitoring controls, and execute a Modified Retention amendment.” Processing stays in region only for listed models and modes. “Data residency does not apply to system data, which may be processed and stored outside the selected region”, and ZDR-ineligible endpoints “may retain application state”. The Agents API, in beta, “currently supports data residency only in the United States and does not support Zero Data Retention (ZDR)”.

For ChatGPT Enterprise, OpenAI’s trust portal says: “Eligible ChatGPT Enterprise and Edu customers can choose from supported regions for storage of covered customer content. Inference residency is a separate capability available in specified regions.” In ChatGPT Work, “Connected apps, external providers, and some processing or synced indexes can follow separate location rules”, and API zero data retention does not define ChatGPT Work retention (OpenAI). During the Enterprise beta, “dots do not support data residency or inference residency” (OpenAI).

Azure OpenAI data zones: does the EU Data Zone keep inference in the EU?

Yes, inside Microsoft’s definition of the EU. Microsoft’s deployment types page says “Data stored at rest remains in the designated Azure geography” for every type, and for Data Zone types “The service processes data only within the Microsoft-specified data zone (US, EU, or Asia Pacific (APAC)).” The EU zone follows the EU Data Boundary, and “Microsoft can add regions to either data zone without prior notice”.

Abuse monitoring data stays in your geography, reviewed for EEA deployments by Microsoft employees “located in the European Economic Area” (Microsoft), unless you are approved for modified abuse monitoring.

Two catches. New models reach Data Zone only after Global: “Geography-based deployment types arrive last, have no guaranteed availability date”. And for Grounding with Bing, “your data flows outside the Azure compliance and Geo boundary.” With the standard agent setup, by contrast, “Agent states (conversations, responses) are stored in your own Azure resources”, and so is the retrieval index.

Gemini EU data residency: Gemini API vs Vertex AI vs Gemini in Workspace

Vertex AI and Workspace offer an EU option; the Developer API documents none.

Gemini Developer API. Google documents no regional processing option and points enterprise workloads elsewhere: “If your workload requires guaranteed zero data retention or enterprise data processing agreements, use Vertex AI” (Google).

Vertex AI, which Google’s documentation now calls Gemini Enterprise Agent Platform, offers the “eu” multi-region or an EU region, where machine learning processing “stays within a specific jurisdictional boundary” (Google). Zero retention takes work per feature, and on the global endpoint cached data “can be served from any region” (Google).

Gemini in Workspace. Admins choose a data region: US, EU or no preference. Keeping processing there needs the Enterprise or Assured Controls version of Data Regions, included in Enterprise Plus, Enterprise Essentials Plus, Frontline Plus and the Business Continuity editions and sold as an add-on to editions with Fundamental Data Regions, and with it “some advanced features might become unavailable”. Data Regions does not cover “Gemini Notebook (formerly NotebookLM), or logs or cached content”. The privacy hub keeps prompts and responses for “90 days to indefinite, as determined by admins”.

Where does Mistral store and process data?

In the EU by default for storage, and in the EU for processing only if you call the EU endpoint. Mistral’s help centre says data is hosted in the EU unless you explicitly use the US endpoint, and that “Depending on the feature you use, your data can be temporarily transferred outside of the European Union”.

Mistral’s regional inference documentation separates the two. Of the default endpoint: “Mistral does not commit to a specific inference location for requests sent to this endpoint.” The EU endpoint costs “a 10% upcharge”, does not keep all control-plane data such as API keys and billing in region, and does not serve Agents, Batch or the Files API.

Vibe, formerly Le Chat, sits outside those guarantees, which are written for API endpoints; Mistral documents no committed inference location for it. Zero retention mostly misses it too: ZDR “is available only with pay-as-you-go and only for stateless API calls”, and “ZDR is not available for Vibe Work, Chat, libraries, agents, conversations, batch files, or other stateful products that must store data to work”. Vibe CLI calls follow the underlying API’s ZDR rules when ZDR is enabled on pay-as-you-go (Mistral).

Mistral’s strongest residency answer is to run the model yourself: its pricing page offers deployment “self-hosted, in your private cloud, or on-premises; your data stays in your environment”, and shows “Contact us” for the Enterprise plan rather than a price. When a cloud provider serves the models, Mistral’s partner terms say “We do not have access to the Infrastructure or Your Data”. Residency then follows the cloud provider, and so does jurisdiction.

Is Microsoft 365 Copilot inside the EU Data Boundary?

It depends on when the tenant was created. Microsoft’s Copilot privacy page says “EU traffic stays within the EU Data Boundary”, but with flex routing “LLM inferencing may occur in the United States, Canada, or Australia during times of peak demand”, and flex routing “is on by default for eligible tenants that were created after March 25, 2026”. With it on, “limited pseudonymized data” may be stored outside the boundary. Tenants that bought or used Multi-Geo are outside the EU Data Boundary for Microsoft 365 altogether.

Support access is the class nobody counts: Microsoft treats remote access from outside the boundary as a transfer under European privacy law, even when the data stays in EU datacentres. Agents reading Fabric data have a further exit.

Are Anthropic models in Microsoft 365 Copilot inside the EU Data Boundary?

No. Microsoft’s AI subprocessor page says Anthropic models “are currently excluded from the EU Data Boundary, and when applicable, in-country processing commitments”. “Off by default in the EU” is true of one setting out of two.

SettingApplies toDefaultProcessing
Global Anthropic subprocessorCopilot, Copilot Studio, Researcher, Power PlatformOff: “Customers within the EU Data Boundary and customers in the UK have Anthropic models disabled by default”Outside the EU Data Boundary
Anthropic in Word, Excel and PowerPointCopilot in those appsOn for EU, EFTA and UK tenants created after 25 March 2026; older tenants check the Message Center“outside of the Microsoft EU Data Boundary (EUDB)”
Flex routingCopilot inferencing at peakOn for eligible EU and EFTA tenants created after 25 March 2026 (UK not in scope)US, Canada or Australia

The Word, Excel and PowerPoint page sets that default and calls the setting separate from the global one; whether it applies while Anthropic is not enabled as a subprocessor is unclear, so confirm in the admin center. For Anthropic “models with Data Retention”, a separate opt-in off by default, the subprocessor page says “data is stored by Anthropic and not subject to your Microsoft Customer Agreement including commitments in the Product Terms and DPA”.

Does data residency cover prompts, agent logs and retrieval indexes?

Not as one switch. Sorted by what the vendor documentation covers, there are six classes of data, each with its own rule.

Data classWhat an EU option usually coversExceptions the vendors document
Prompts in flightBedrock EU profile, Google “eu”, Azure EU Data Zone, OpenAI and Mistral EU endpointsAnthropic API: global or US only. Fable on Bedrock: global only. Copilot flex routing. Any global deployment
Stored conversations and stateAt rest in the chosen regionAnthropic: US only. OpenAI: state on ZDR-ineligible endpoints. Mistral: stateful features off the EU endpoint
Agent logs and memoryAzure’s standard setup keeps agent state in your resourcesAnthropic memory stores; OpenAI Agents API US only; Workspace logs not covered
Retrieval indexesAn index in your own EU accountOpenAI vector stores not ZDR-eligible; ChatGPT synced indexes; Gemini Notebook; Mistral libraries
Abuse-monitoring copiesAzure and Google: your regionBedrock: destination Region. Google: Fable and Mythos shared with Anthropic. OpenAI: up to 30 days unless approved. Copilot: Anthropic retention models outside Microsoft’s DPA
Support access and system dataRarely coveredOpenAI system data; Mistral control plane; Microsoft remote access counted as a transfer; Anthropic support and safety review wherever it or its affiliates operate

Inside your company each row usually has a different owner, and an impact assessment naming one region for “the AI system” has answered the first row and assumed the other five. The intelligence exhaust audit asks who owns these artefacts; this table asks where they sit. Training defaults are a separate question.

Does EU data residency cost more?

Where it is priced, yes, and the premium is similar everywhere: one more line in the monthly run cost. OpenAI’s pricing page says “Regional processing (data residency) endpoints are charged a 10% uplift for models released on or after March 5, 2026.” For Google-managed Claude on Sonnet 4.5 and later models, “Multi-region and regional endpoints cost 10 percent more than global endpoints” (Anthropic). AWS frames it as a saving: global routing offers “Approximately 10% savings” over the geographic profiles it recommends “when you have data residency requirements” (AWS). Mistral charges its 10% upcharge. Anthropic has no EU price because it has no EU option; on Claude 4.6 and later models, its US-only inference “is priced at 1.1x the standard rate across all token pricing categories” (Anthropic).

US CLOUD Act vs GDPR: can a US provider be made to hand over data stored in the EU?

Yes, a US court can order it, and GDPR does not make that order enforceable in Europe. The provider is caught between the two.

18 U.S.C. 2713 requires a provider to disclose data in its control wherever it is stored, and an AI API sold to the public plausibly qualifies as a remote computing service, “the provision to the public of computer storage or processing services” (18 U.S.C. 2711). 18 U.S.C. 2703(h) lets the provider move to quash within 14 days, for a customer that is not a US person and does not reside in the United States, where disclosure would risk violating the laws of a “qualifying foreign government”: one “with which the United States has an executive agreement that has entered into force under section 2523”.

The Justice Department’s CLOUD Act resources page, last updated in October 2023, lists agreements only with the United Kingdom and Australia, and for the EU only negotiations, resumed on 3 March 2023. The European Commission’s e-evidence page still says “negotiations with the United States are ongoing”. Until an agreement enters into force, the statutory motion is not available for a conflict with EU law, leaving the common-law comity argument, which the Act’s rule of construction leaves untouched.

Announcing the final Guidelines 02/2024 on Article 48 GDPR on 5 June 2025, the EDPB said “judgements or decisions from third country authorities cannot automatically be recognised or enforced in Europe” (EDPB). The final version addresses two cases that fit AI vendors: a request that lands on a processor, and a parent outside Europe asking its European subsidiary for the data.

It happens rarely, not never. Microsoft’s transparency report for July to December 2025 says it “provided content data to U.S. law enforcement related to 3 non-U.S. enterprise customers whose data was stored outside the U.S.”, and “One of the 3 customers was located in the EU/EFTA.” Residency is not jurisdiction makes the argument for Microsoft; it holds for every US-parented row above.

A US order reaches what the provider holds: the 30-day abuse-monitoring copy, the agent transcript kept until deleted, the managed vector store, Anthropic’s memory store. Zero retention leaves less to produce. An index in your own EU account sits outside the model vendor’s control, though not your cloud provider’s, an exposure the sovereign and EU-owned options narrow.

Mistral’s data processing addendum names it “a French limited joint-stock corporation, incorporated in Paris”, which lowers the exposure without settling it: 2703(h) covers “a foreign electronic communication service or remote computing service”, and the Justice Department’s CLOUD Act FAQ says jurisdiction is not limited to US companies but turns on a provider’s contacts with the United States. Mistral runs a US endpoint; ask which entity operates it.

What should a residency clause name?

The exit-clause checklist and the four questions a European buyer should add cover export, deletion, legal entity and deployer duties. Residency needs four asks of its own:

  1. A region per data class: prompts, stored state, agent logs and memory, retrieval indexes, abuse-monitoring copies and backups.
  2. Subprocessor regions for abuse monitoring and support, with notice before they change, and a written answer on any copy shared with the model maker.
  3. Notice of foreign government access requests, where the law permits.
  4. Notice before a residency setting changes, and the right to keep the old one: data zones gain regions without notice, and new tenants get new defaults.

When is EU residency enough, and when is it not the question?

For an EU-hosted assistant for staff, buy the vendor’s EU option and stop: Gemini in Workspace with an EU region and in-use processing, or Microsoft 365 Copilot with flex routing and both Anthropic settings off. Claude’s own apps have no EU option, and Claude inside Copilot runs outside the EU Data Boundary. For agents on Claude in the EU, use Bedrock or Google-managed Claude, and Google alone for Fable; for OpenAI models, an Azure EU Data Zone or OpenAI’s Europe region. We build for EU residency, but we do not sell it as a product, and nobody needs a transformation partner to tick a region box.

Residency stops being the question when your agents depend on what your company knows: procedures, decisions taken in email, experts’ corrections. Those become prompts, agent logs, memory and indexes, and the tables show where vendors keep them, not whether you can take them to the next model.

Residency is one layer. Who owns the context is the other.

The second layer is covered in whether a sovereign cloud is enough to own your intelligence and where your company’s know-how should live. A context engine in your own EU accounts keeps indexes, agent logs and memory where you chose while the model on top changes; inference still runs wherever that model runs, and the accounts are only as far from US law as their cloud. Settle residency per data class before the first agent reaches its production gate.

FAQ

Does Claude have data residency in Europe? Not from Anthropic directly: its API offers only “global” or “us” inference and “us” storage. In the EU, use Amazon Bedrock’s EU inference profile or Google-managed Claude on its “eu” endpoint, and only Google for Fable, which Bedrock serves in Europe only globally. Microsoft Foundry has no EU option for Claude.

Does OpenAI offer data residency in the EU? Yes, with conditions. The API’s Europe region covers the EEA and Switzerland, only for listed models and modes, only after OpenAI approves a retention control, and never for system data. Eligible ChatGPT Enterprise customers can choose a storage region; inference residency is separate.

Is Microsoft 365 Copilot inside the EU Data Boundary? Mostly. Two exceptions are on by default for tenants created after 25 March 2026: flex routing at peak demand, and Anthropic models in Word, Excel and PowerPoint (confirm in the admin center). Multi-Geo tenants are outside the boundary altogether.

Are Anthropic models in Microsoft 365 Copilot inside the EU Data Boundary? No, Microsoft currently excludes them. The global subprocessor setting is off by default in the EU, EFTA and UK, but the Word, Excel and PowerPoint setting is on by default for tenants there created after 25 March 2026 (confirm in the admin center).

Does data residency cover agent logs and retrieval indexes? Often not. Anthropic keeps memory store contents even with a self-hosted sandbox, OpenAI’s Agents API supports residency only in the United States, and Workspace Data Regions does not cover logs or Gemini Notebook. Ask for a region per class of data.

Does EU data residency cost more? Where it is priced, yes. OpenAI charges a 10% uplift on residency endpoints for models released on or after 5 March 2026, Google-managed Claude costs 10 percent more on regional and multi-region endpoints for Sonnet 4.5 and later models, Bedrock’s global routing is about 10% cheaper than its geographic profiles, and Mistral’s EU endpoint carries a 10% upcharge.

Does the US CLOUD Act override GDPR? No, and GDPR does not override the CLOUD Act. US authorities can compel a provider to disclose data it holds wherever it is stored; under GDPR Article 48, as the EDPB reads it, that order cannot automatically be recognised in Europe. The provider is caught between the two.

Is Mistral outside the reach of the US CLOUD Act? Not by nationality alone. Mistral AI is French, but the CLOUD Act’s motion-to-quash provision expressly covers foreign providers. What matters is whether US courts have jurisdiction over the provider, or over the cloud provider serving Mistral’s models.



Nucleo is an AI-native transformation company for European mid-caps. Our forward-deployed engineers build the context engine your company runs on, in your own accounts and model-agnostic, so agent logs, memory and retrieval indexes sit in accounts and a region you chose, whichever model runs on top. See how an engagement works, or write to us.

Sources, checked on 29 October 2026: vendor documentation from Anthropic, AWS, Google, OpenAI, Microsoft and Mistral, linked in each section; 18 U.S.C. 2711, 2713 and 2703; US Department of Justice, CLOUD Act resources and CLOUD Act FAQ; EDPB, Guidelines 02/2024 on Article 48 GDPR; European Commission, e-evidence; Microsoft, law enforcement requests report.

The future belongs to those who see it before it is obvious, and build it while everyone else is still out there scouting for the best AI tools.