SovereigntyResidency guarantees

Vertex AI data residency: GDPR rules and EU alternatives

Assess Vertex AI data residency against GDPR and the US CLOUD Act. Compare sovereign EU AI inference alternatives including IONOS, Nebius, and OVHcloud.

, AI Researcher9 min read
In this article8
  1. How Google Cloud structures Vertex AI data residency
  2. The GDPR compliance layer, SCCs, and zero retention
  3. The US CLOUD Act conflict and US jurisdiction exposure
  4. Evaluating European sovereign alternatives for AI inference
  5. Nebius and OVHcloud: Owned EU infrastructure
  6. IONOS AI Model Hub: ISO-certified German inference
  7. Frequently asked questions
  8. Sources
A Google data centre at Eemshaven in the Netherlands with wind turbines behind it
Google's Eemshaven site. An EU region answers where the data sits, not which jurisdiction can compel it. Photo: Wvdp, CC0, via Wikimedia Commons

How Google Cloud structures Vertex AI data residency

Google Cloud organizes data residency for artificial intelligence workloads by differentiating between data stored at rest and machine learning (ML) inference processing. Under Google Cloud's data residency commitments, enterprise customers using its generative AI services can choose to store their data at rest in customer-selected locations, consistent with Google Cloud's General Service Terms. For generative AI on Vertex AI, Google lists European data-store options for the Netherlands, France, the UK, Germany and Belgium, while Vertex AI Search and Vertex AI Conversation each offer an EU multi-region data-store option.[1]

A critical distinction lies between the enterprise Vertex AI platform and consumer-facing Gemini offerings. Consumer Gemini (accessed via web interfaces or consumer accounts) operates under general consumer terms where prompt inputs may be routed globally and processed without geographic guarantees. By contrast, Vertex AI ties the guarantee to the endpoint the request is sent to, and Google's documentation splits the question in two (retrieved 19 August 2026). Data stored at rest stays in the location the customer chose, independent of the endpoint called. The location of ML processing - inference and tuning - is decided by the endpoint: jurisdictional multi-region endpoints keep processing inside one geography, and Google states that the European Union multi-region (eu) endpoint "strictly covers data residency within EU member states", with the United Kingdom and Switzerland excluded.[2] Locational endpoints such as europe-west3 keep processing within that jurisdiction, but Google adds that local in-country processing varies by model type, and it publishes the ML-processing commitment model by model rather than platform-wide. Global endpoints - the hostnames with no region in them, such as aiplatform.googleapis.com - "route and process data anywhere globally" and, in Google's own words, "don't provide regional isolation or data residency guarantees".[2]

  • Generative AI on Vertex AI offers European data-store options for the Netherlands, France, the UK, Germany and Belgium.[1]
  • Vertex AI Search and Vertex AI Conversation each offer an EU multi-region data-store option.[1]
  • Contractual commitments apply strictly to enterprise Vertex AI instances rather than consumer Gemini accounts.

The GDPR compliance layer, SCCs, and zero retention

For European organizations processing personal data under Regulation (EU) 2016/679 (GDPR), Vertex AI operates under the Google Cloud Data Processing Addendum (CDPA). In that addendum Google acts as processor and the customer as controller (or processor), and the international transfer commitments that govern where customer data may be processed sit in its specific privacy law appendix.[3] Under the training restrictions in the service specific terms, Google does not use customer data to train or fine-tune its AI/ML models without prior permission or instruction. The addendum's own default cuts the other way on location: subject to Google's data location commitments under the Service Specific Terms and the transfer commitments in Appendix 3, customer data may be processed in any country where Google or its subprocessors maintain facilities.[3] Residency on Vertex AI is therefore something the customer configures and Google commits to per service and per model, not a property the contract confers by default.[4]

However, technical defaults require active configuration to satisfy strict data minimization mandates under Article 5(1)(c) of the GDPR. By default, Gemini models cache customer inputs, outputs and derived data in memory to cut latency: that data stays in memory rather than at rest, is isolated at project level, and carries a 24-hour time-to-live, and caching can be switched off per project.[4] To avoid data retention, teams must also leave request/response logging disabled, set store to false on Interactions API requests (the default is true), and obtain an approved abuse-monitoring exemption where prompt logging for abuse monitoring applies.[4]

Configuration LayerDefault SettingZero-Retention RequirementGDPR Impact
Prompt CachingIn-memory cache of inputs, outputs and derived data with a 24-hour TTLDisable caching at project level via the cacheConfig endpointEnforces strict data minimization under Article 5
Contractual TermsCloud Data Processing Addendum, with Google as processor and the customer as controllerSign the Google Cloud agreement that incorporates the Cloud Data Processing AddendumEstablishes Article 28 processor obligations
Model TrainingGoogle does not use customer data to train or fine-tune models without prior permission or instructionMaintain default enterprise tenant isolationPrevents unauthorized secondary processing

The US CLOUD Act conflict and US jurisdiction exposure

Physical server location in Frankfurt or Eemshaven does not establish legal immunity from foreign statutory claims. The Clarifying Lawful Overseas Use of Data (CLOUD) Act was adopted by the US Congress on 23 March 2018 and amends the Stored Communications Act to make explicit the obligation of US service providers to preserve and produce data they control regardless of where it is stored.[5] The amended statute requires a provider to preserve, back up or disclose the contents of communications and related records within its possession, custody or control, regardless of whether that information is located inside or outside the United States.[6]

That mandate sits unresolved against Article 48 of the GDPR, under which a third-country court or administrative order to disclose personal data is not recognised or enforceable unless it rests on an international agreement such as a Mutual Legal Assistance Treaty. The conflict is a live legal argument rather than a settled one, and it applies to the US hyperscaler as a corporate entity, not to any particular Vertex AI region setting.

  1. Extraterritorial reach: the obligation attaches to data in a US provider's possession, custody or control, wherever it is stored.[6]
  2. Conflict of laws: Direct disclosure under US warrants circumvents the MLAT process and conflicts with GDPR Article 48.
  3. Residual exposure: custody of encryption keys is a separate question from residency and is not settled by an endpoint setting.

Evaluating European sovereign alternatives for AI inference

European buyers therefore evaluate European alternatives that maintain corporate headquarters, operational personnel, and physical infrastructure entirely within EU member states. Despite significant enterprise interest in digital sovereignty, market concentration remains pronounced: Synergy Research Group data reported by IT Pro puts European providers' share of their own regional cloud market at around 15%.[7]

Deploying open-weights foundation models on sovereign European cloud infrastructure narrows exposure to foreign extraterritorial discovery requests, because the argument turns on the operator's corporate nexus rather than on the hardware. What separates candidates is how far down the stack each will commit in writing: Google states its EU machine-learning-processing commitment model by model rather than platform-wide, and the smaller EU operators are no different.

Infrastructure ModelJurisdictional ExposureContractual BasisHardware Ownership
US Hyperscaler EU RegionReachable under the US CLOUD Act because the provider is subject to US jurisdictionCloud Data Processing Addendum with Standard Contractual ClausesHyperscaler proprietary datacenter
Independent EU AI CloudEU/EEA jurisdiction exclusivelyDirect GDPR Article 28 processor agreementOwned or co-located European facilities
On-Premises DeploymentComplete internal jurisdictionDirect controller governanceEnterprise-owned bare-metal compute

Nebius and OVHcloud: Owned EU infrastructure

Specialized European infrastructure operators have scaled proprietary compute clusters to provide high-performance alternatives for enterprise workloads requiring true data sovereignty. Nebius contracts through Nebius B.V., a company incorporated in the Netherlands, and through its Token Factory platform provides API inference for more than 60 open-source models, including Llama, DeepSeek and Qwen,[9] with zero data retention available as an opt-in rather than as the default.[8]

Similarly, France-based OVHcloud delivers AI compute through company-owned datacenters across Europe, offering bare-metal GPU instances alongside managed AI software stacks. Because OVHcloud is French-incorporated with no US parent, it sits outside the corporate nexus the CLOUD Act relies on. That is a jurisdictional argument rather than a technical guarantee, and it should be checked against the contract and the sub-processor list like any other.

  • Nebius offers zero data retention as an opt-in and publishes the processing location per model - EU (Finland and France), Israel or the US - noting it can change for public endpoints.[8]
  • OVHcloud operates vertically integrated European facilities from custom hardware design to localized colocation.
  • Hardware ownership makes the physical and hypervisor layers auditable.

IONOS AI Model Hub: ISO-certified German inference

The IONOS AI Model Hub is the turnkey option in this group. Built and managed by Germany-based IONOS, it provides managed access to open-weights models from providers such as Meta and Mistral without requiring underlying infrastructure orchestration.[10]

All AI Model Hub services and vector database integrations run inside ISO/IEC 27001-certified datacenters located exclusively in Germany.[11] IONOS states that data processing is confined within Germany and that customer data, including prompts, inputs and documents uploaded for retrieval-augmented generation, is not used for training, fine-tuning or otherwise improving any AI models, under its Data Processing Agreement. Combining German incorporation with local compute puts the operator outside the US corporate nexus the CLOUD Act depends on, which is the one thing a Vertex AI EU endpoint cannot offer at any configuration.

  • ISO/IEC 27001-certified German datacenters ensure rigorous operational and technical controls.
  • Turnkey managed API endpoints for Llama, Mistral, and other open-weight models in the IONOS catalogue.[10]
  • Keeps governance and physical hosting within Germany, removing the US corporate nexus that CLOUD Act process relies on.

Frequently asked questions

Does Vertex AI keep my data in the EU?

Only if you configure it. Google lists five European data-at-rest countries for generative AI on Vertex AI - the Netherlands, France, the UK, Germany and Belgium - and data stored there stays there regardless of which endpoint you call. Where the model actually runs is a separate setting: the EU multi-region endpoint keeps ML processing inside EU member states and excludes the UK and Switzerland, while the default global endpoints provide no regional isolation or data residency guarantee at all. Google publishes the ML-processing commitment model by model. Neither setting insulates the data from US jurisdiction.

How long does Vertex AI retain prompt data?

By default, Vertex AI uses in-memory caching for Gemini models to minimize latency, which retains prompt data at the project level for up to 24 hours. Enterprise administrators can disable this caching via the API to achieve zero data retention.

Does the US CLOUD Act apply to Google Cloud servers in Europe?

Yes. The CLOUD Act requires companies subject to US jurisdiction to disclose data in their possession upon government demand, regardless of physical location. Because Google is headquartered in the United States, its European servers remain legally exposed to these access requests.

Is Google Workspace Gemini compliant with the GDPR?

Both operate under an enterprise Data Processing Addendum. That covers the Article 28 processor obligations, but it does not resolve the conflict between a US provider's CLOUD Act duties and GDPR Article 48.

What are the sovereign alternatives to Vertex AI in Europe?

IONOS, Nebius and OVHcloud all run open-weight models on European infrastructure without a US parent. That changes the jurisdictional question rather than the technical one, and each still has to be checked on residency, retention and contract terms.

Sources

  1. Google Cloud generative AI data residency guarantees for data stored at-rest cloud.google.com
  2. Data residency | Gemini Enterprise Agent Platform docs.cloud.google.com
  3. Cloud Data Processing Addendum cloud.google.com
  4. Gemini Enterprise Agent Platform and zero data retention docs.cloud.google.com
  5. The CLOUD Act eurojust.europa.eu
  6. 18 U.S. Code § 2713 - Required preservation and disclosure of communications and records law.cornell.edu
  7. US companies dominate the European cloud market – regional players are left fighting for scraps itpro.com
  8. Legal Quick Guide docs.tokenfactory.nebius.com
  9. Nebius Token Factory nebius.com
  10. Models | Products docs.ionos.com
  11. European Union AI Act | Products docs.ionos.com

Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score

Related reading

All Sovereignty →

Best serverless AI for European teams

Compare ten European serverless AI inference providers based on data residency, certification, latency, and per-token pricing to make an auditable choice.

A comparison table layout showing European flags and data nodes, representing the evaluation of serverless AI inference providers across the continent.