What you can legally use from a company in the EU or in the Republic of Moldova, where data is actually processed and what a wrong choice costs you.
Verified September 6, 2026 · 23 models · 6 platforms · 27 official sources
The question we are asked most often is not “which model is best”. It is “which am I allowed to use”. These are two different questions, and the second has answers that change monthly.
This page is a checklist, not a score comparison. For each model we say one thing: whether you can obtain the guarantee that your data are processed in the European area — and under exactly which conditions. The rest (speed, price, quality) you can find anywhere.
Everything below was checked on 6 September 2026, against the provider’s public documentation. Where the provider does not say, we write that it does not say.
Each line was checked against the provider’s documentation on the day of publication: the official pages were downloaded and their HTML tables were parsed directly, not read from memory. Where the provider does not publish the information, the line is marked "unverified" and explained.
Tabelul
Each model, with what can be demonstrated about it
Click a line for data residency, retention, processing agreement and pitfalls. No cell is filled in from memory.
23 from 23
Da
The provider publishes an option through which both storage and processing remain within the European area, and the model appears on the eligibility list.
Conditional
The model is available with a European guarantee only through certain platforms, regions, or deployment types — not by default.
Nu
The model can be called from Europe, but there is currently no published option for guaranteed processing in the European area.
Neverificat
The provider does not publish the information in a form that we could verify today.
“The model is available in the EU” and “my data stays in the EU” are two different things
This is where most buyers get it wrong, and there is an example that makes everything clear in ten seconds.
On Microsoft Foundry, `gpt-6-astra` — the model launched by OpenAI on 3 September 2026 — appears in the official availability tables for European regions: France Central, Germany West Central, Sweden Central, West Europe, and the rest. You can deploy it from a European account, in a European data centre, with an invoice in euros.
But it appears only for the **Global Standard** and **Global Provisioned Managed** deployment types. It does not appear at all in the **Data Zone Standard** table. And Microsoft’s documentation states, verbatim, what the difference means: Global types “may be processed in any Azure region”, while Data Zone types process “only within the Microsoft-specified data zone”.
In other words: the model is available in Europe, but your prompt may go anywhere. The most powerful OpenAI model with guaranteed processing inside the EU data zone on Azure today is GPT-5.6, not GPT-6 Astra.
The same pattern at Google. In Vertex AI’s official residency table, the Gemini 3.x **Flash** family (3.8, 3.7, 3.6, 3.5) has a tick for “EU multi-region”. Gemini 3.1 Pro does not appear in the table at all. The model exists in Europe; the processing guarantee does not.
The practical rule: the region name in the console is not a guarantee. The guarantee is the deployment type, the endpoint, and the eligibility table — three different things, published on three different pages.
There is another level nobody reads: **storage** and **processing** are regulated separately. OpenAI also names them separately — “data residency” for what is stored at rest, “inference residency” for where the GPU actually runs. Of the ten regions OpenAI offers at API level, only three also have regional processing: the United States, Europe (EEA + Switzerland), and the United Arab Emirates. The rest — the United Kingdom, Canada, Japan, India, Singapore, South Korea, Australia — have storage only. Choose “Canada” and your inference still goes elsewhere.
Three questions that separate marketing from contract
Where is data stored at rest? (data residency)
Where does the model actually run? (inference / ML processing residency)
What happens to abuse monitoring logs, which are a third flow, separate from the first two?
The traps that do not appear in any commercial presentation
**Video and residency are mutually exclusive at OpenAI.** The documentation says the `/v1/videos` endpoint “is currently blocked for MAM or ZDR requests”. Since European residency requires either Modified Abuse Monitoring or Zero Data Retention, today you cannot have both video generation with Sora 2 and processing in the EU in the same project.
**EU residency at OpenAI is not a button.** It requires: commercial approval (it does not activate itself), a contract amendment (“Modified Retention amendment”), approval for retention controls, and costs **10% more** for models launched after 5 March 2026. In addition, image generation in non-US regions requires separate approval for the “enhanced” variant of ZDR or MAM.
**The project is chosen only once.** The region is set when the project is created in the API. An existing project cannot be moved to Europe later.
**“EU” at AWS does not automatically mean the EEA.** In Amazon Bedrock’s official regional compatibility table, the “Geo” column is also ticked for `eu-west-2` (London) and `eu-central-2` (Zurich) — regions outside the European Economic Area. AWS documents the destination regions list on each model’s page and states that, once tied to a geography, the list no longer changes. Read it for your model before translating “eu.” as “EEA”.
**MCP servers are a hole in any residency policy.** OpenAI’s documentation says it directly: “MCP servers are third-party services. Data sent to an MCP server is subject to its data residency policies.” You can have an impeccable EU setup and still send data outside through a connected tool.
**System data does not fall under residency anywhere.** Authentication, routing, analytics, billing, metadata — for all verified providers these remain outside the regional guarantee. They do not contain client content, but they must be declared in the processing register.
**Zero Data Retention does not cover everything.** At Mistral, for example, ZDR applies only to stateless calls — chat, embeddings, moderation, classification, OCR, speech synthesis and transcription. It does not apply to agents, conversations, libraries, files or batch processing, because those products must store data in order to function.
Platforme
What you reach the model through — and what conditions each one sets
It is configured per project, when the project is created. Alternatively, you can choose processing per request by using the domain prefixed with a key from a project with global geography.
Conditions
Eligibility confirmed by the commercial team; approval for abuse monitoring controls; a signed retention amendment for any region outside the US; 10% extra in price for models launched after 5 March 2026.
The trap
An existing project cannot be moved to Europe after creation. TLS termination is done in-region through Cloudflare Regional Services, which is a useful detail to know when completing the sub-processor register.
OpenAI — ChatGPT Enterprise / Edu
The European option
Data residency in “Europe (EEA + Switzerland)” for storage at rest, plus inference residency, which is a separate setting.
How it is activated
It is activated when provisioning a new workspace. Workspace owners see the region configured in settings, but that is display, not control.
Conditions
Inference residency requires data residency to already be active in the same region.
The trap
Inference residency covers GPU execution, not all processing. Authentication, routing and analytics may remain outside the region. And if OpenAI launches new features, they are not covered by residency unless they are explicitly added to the list.
Microsoft Foundry / Azure OpenAI
The European option
The "Data Zone" deployment type, the EU zone.
How it is activated
The deployment type is chosen when the model is deployed. The documentation defines the zone textually: "European Union (data processed within any EU member nation)".
Conditions
The model must exist in the Data Zone table for your region. Not all models are there, and the list shifts with each release.
The trap
Microsoft describes the launch order: first Global, then Data Zone, then geography-linked ones — and the last ones "have no guaranteed availability date". In other words, the gap between a model’s release and its guaranteed EU availability is a platform design fact, not an accident. Plan for it.
Google Cloud Vertex AI
The European option
The multi-regional `eu` endpoint, plus individual European regions (Belgium, France, Germany, the Netherlands).
How it is activated
Google explicitly separates data at rest from ML processing and says the place of processing is determined by the chosen endpoint.
Conditions
The model must appear in the residency table, which is updated frequently (the version read today was from 2 September 2026).
The trap
The most honest wording in everything I read today, and the most useful for Moldovan law: the `eu` endpoint "strictly covers data residency within EU member states. Geographies outside the European Union political boundary, including the United Kingdom and Switzerland, are excluded." By contrast, global endpoints "don't provide regional isolation or data residency guarantees".
Amazon Bedrock
The European option
Geography-bound inference profiles, with the `eu.` prefix, plus in-region availability for some models.
How it is activated
The geography prefix is placed in the model identifier. AWS specifies that, once tied to a geography, the destination-region list of a profile never changes again — unlike global profiles, which may expand as AWS adds regions.
Conditions
The destination-region list is documented on each model’s page. It must be read per model.
The trap
Two things. First: in the compatibility table, the "Geo" column is also ticked for London and Zürich — check what this means for your model before equating `eu.` with "EEA". Second: cross-region inference profiles can route to "opt-in" regions you have not enabled, and AWS states directly that "your input prompts and output results may be stored in the opt-in Regions for abuse detection purposes". Service control policies are the mechanism by which you limit this.
Model run on own infrastructure
The European option
Total, by definition.
How it is activated
You download the weights, run them on hardware you control, in a data centre you choose.
Conditions
The model licence must be read. Apache 2.0 and MIT do not raise commercial issues; “community” licences and those marked “other” must be read in full.
The trap
You have no transfer, so you do not have Chapter V of Law 195/2024 — but you do have all the other obligations: security, record of processing activities, data subject rights. Sovereignty solves one problem on the list, not the list.
What the EU AI Regulation means for a buyer
The timetable has been changed. What is circulating in the press is, in large part, outdated.
The Regulation (AI Act) entered into force on 1 August 2024, but it applies in stages. The official timetable published by the European Commission has been **changed by “Digital Omnibus on AI”**, and the official page states this explicitly. Many press articles and many commercial presentations still cite the old version, in which the obligations for high-risk systems fell in August 2026.
What matters in practice for a buyer: if you use a chatbot, a voice agent or generate synthetic content, your main obligation is **transparency (Article 50), from 2 August 2026**. It must be clear that the interlocutor is speaking with a machine and that a piece of content is generated artificially. It is not an architecture obligation, it is an interface and process one — and it is the one that is easiest to check from the outside.
The heavy obligations, those for high-risk systems in Annex III, were postponed to **2 December 2027**. If someone is selling you “AI Act compliance” today for an ordinary use case, ask which article, with what date.
2024-08-01Entry into force
2025-02-02General provisions (definitions, AI literacy) and prohibitions
2025-08-02Rules for general-purpose AI models; governance must be functional
2026-08-02Most rules start to apply; transparency rules (art. 50) start to apply; national and EU implementation begins
2026-12-02New prohibitions (non-consensual sexual deepfake, child abuse material) + transitional period for art. 50(2)
2027-08-02Member States should have at least one operational regulatory sandbox
2027-12-02Rules for the high-risk systems in Annex III (postponed by the Digital Omnibus)
2028-08-02Rules for high-risk AI embedded in regulated products (Annex I)
Republic of Moldova: Law 195/2024 and why the EEA changes the entire calculation
The part nobody else writes about, because you need to open the text of the law.
Law No. 195/2024 on the protection of personal data has been in force since 23 August 2026 and replaces Law 133/2011. Chapter V regulates transfers to other states, and Article 44 sets out the general principle.
Paragraph (2) of Article 44 is the one that matters for any AI architecture decision made by a Moldovan company.
The consequence is direct: if you can keep processing within the European Economic Area, **you do not need to build a transfer file**. You do not need standard contractual clauses, you do not need a transfer impact assessment, you do not need authorisation from the Centre. If the data leave the EEA, you enter the rest of the chapter: either an adequacy decision issued by the Centre (Art. 45), or appropriate safeguards with standard contractual clauses, or a derogation.
From this comes a design rule worth writing on the wall: **for a company in Moldova, a well-chosen EU route is not only safer, it is cheaper in legal work.** The difference between `Data Zone Standard` and `Global Standard` on Azure is not a technical preference; it is the difference between zero transfer documents and an entire file.
Beware geography, however. The EEA means the 27 EU member states plus Iceland, Liechtenstein and Norway. **The United Kingdom is not in the EEA. Switzerland is not in the EEA.** And some commercial assurances mix exactly these jurisdictions: the OpenAI region is called “Europe (EEA + Switzerland)”, while the Azure data zone is defined, literally, as “data processed within any EU member nation” — that is, only EU member states, which is narrower than the EEA, but still entirely within it. Read the definition, not the label.
This chapter does not apply to transfers of personal data to the member states of the European Economic Area. No special authorisations are required for such transfers.
Legea nr. 195/2024, art. 44 alin. (2)Textul integral al legii, extras cu pdftotext din Legea-195-2024-TEXT-INTEGRAL-71pag.pdf, verificat pe 2026-09-06
How to choose, in six questions
This is not a decision tree that always leads to our product. Two of the branches lead to ‘run it yourself’.
1Does the data that go into the model contain personal data?
NuChoose based on quality and price. The rest of the questions are none of your concern. Just mind the trade secret, which is a contract issue, not a data protection issue.
DaTreci mai departe.
2Are you allowed to have the processing take place outside the European Economic Area?
NuRemove any “Global” implementation from the outset. Stick only to: Azure Data Zone (EU), Vertex AI multi-region EU, Bedrock with an EU geography profile verified on the model page, a provider established in the EU, or a model run on your infrastructure.
DaYou may use any global route, but you sign the standard contractual clauses and document the transfer. At OpenAI and Anthropic, the processing agreements already incorporate the clauses — check them, they are public.
3Ai nevoie exact de cel mai puternic model existent azi?
NuA model one tier lower opens up all routes with processing in the EU, at a cost that is usually half.
DaThen you are on global processing or on eu.api.openai.com with ZDR approved and 10% extra on the invoice. You cannot have both the absolute top and the EU data zone on Azure. The decision is commercial, not technical: either you accept the transfer with its safeguards, or you move down a model level.
4Do you need real-time voice?
The OpenAI Realtime family (`gpt-realtime-2.1` and the others) has regional processing in Europe at the API. However, check the platform you are using: on Azure, the Realtime models appear in European regions under Global Standard, not in the Data Zone table.
5Ai nevoie de generare video?
Then you cannot have residency as well, at least with OpenAI: `/v1/videos` is blocked for requests with MAM or ZDR, and EU residency requires them. Separate the projects or separate the providers.
6You cannot sign any transfer, under any circumstances?
Then only one category remains: open-weight models, run on your infrastructure, in the EU. The level of this category in 2026 is sufficient for most production tasks — see the open models section. The cost moves from subscription to hardware and operations, and that is calculated, not assumed.
What we could not confirm
The gaps, written as gaps
Each of these could have been filled from memory and would have looked good. They were not.
The context windows for GPT-4o, GPT-Realtime-2.1, Amazon Nova, Qwen3.8, DeepSeek-V4 and gpt-oss were not open today on each model's page. They are marked as “unverified” instead of being filled in from memory.
The training policy and processing agreement for Google Vertex AI were not verified today from a primary source. The Gemini rows are marked accordingly.
The exact composition of the destination-region list for the `eu.` profiles from AWS could not be extracted today from the individual model pages. We recorded what the compatibility table says and explicitly marked that per-model verification still remains to be done by the buyer.
Aleph Alpha does not publish a model catalogue with specifications comparable to the other providers. Five fields in its entry are “unverified”, and that is itself the relevant information.
We did not verify the tier-two providers (Cohere, AI21, xAI/Grok, FLUX), although they appear in the Azure tables for European regions. If the page expands, they are first on the list.
30 days — the lists of models eligible for regional processing change with every release.
What we do with this
This page is the material we use internally when designing an integration. We choose the route — provider, region, implementation type, what is logged and where — before we write the first line of code, because changing it later means a new project, not a new setting. At OpenAI, for example, the region cannot be changed on an existing project.
If you want to discuss the route for your specific case, the details on how we route requests and how the provider changes without rewriting the application are on the service page.