# GDPR and AI: Why Your Provider's Location Matters **Published**: 2026-07-05T17:03:00Z **Read time**: 12 min **Topics**: Data Protection, AI, Compliance, GDPR > Two events in June 2026 show that if you use American AI, you carry a data protection risk and an availability risk. Here is how to assess which data should stay in Europe.

GDPR and AI: Why Your Provider's Location Is Becoming a Business Risk

In short: If you send personal data to an AI service from the United States, the GDPR treats that as a transfer to a third country, and you need a legal basis for it. That basis, the adequacy decision on the EU-US Data Privacy Framework, still applies, but a US Supreme Court ruling of 29 June 2026 has put it in question again. There is also an availability risk: in June 2026 Anthropic switched off two AI models for all customers after a directive from the US government, and the Fable 5 model only came back after almost three weeks.

What happened in June 2026?

On 12 June 2026, according to Anthropic, the company received an export control directive from the US government at 17:21 Eastern Time. It required access to the Fable 5 and Mythos 5 models to be blocked for all foreign nationals, inside and outside the United States. Anthropic then switched off both models for all customers, while the company's other models remained available. According to Lawfare, the letter came from the US Department of Commerce, and the shutdown followed within hours. On 30 June the export controls were lifted, and from 1 July Fable 5 was available worldwide again.

On 29 June 2026, the US Supreme Court ruled by six votes to three in Trump v. Slaughter that the statutory protection against removal for the members of the Federal Trade Commission (FTC) is unconstitutional. This means the President can remove them without giving a reason. On the same day, Max Schrems and his organisation noyb (in German) called on the European Commission to withdraw the adequacy decision for the United States in an orderly way. The letter to the Commission is dated 30 June 2026.

Both events lead to the same question. Where your AI provider comes from determines your data protection, legal certainty and operational reliability.

What is this about, and what is it not about?

It is about a sober weighing of risks, not a blanket warning against American technology. US models are often very good, and for many purposes they are perfectly fine.

What matters is which data you hand over, to whom, under which legal regime, and how dependent you become on a single provider in a jurisdiction you have no influence over.

Why does the problem start with the prompt?

With AI language models, the sensitive data is the actual input.

An employee pastes in a customer email to have a reply drafted. Someone uploads a contract to have it summarised. A developer sends source code to a model to find a bug. HR has job applications pre-sorted.

In all these cases, personal data, trade secrets or protected code leave your organisation. If personal data ends up with a provider in the United States, that is a transfer to a third country. Chapter V of the GDPR, which covers Articles 44 to 49, allows such transfers only under certain conditions. Under Article 83(5), infringements can be fined up to €20 million or 4% of worldwide annual turnover.

Is using US AI services allowed under the GDPR in 2026?

Yes, it is currently permitted if the US recipient is on the Data Privacy Framework List kept by the US Department of Commerce. The basis is the European Commission's Implementing Decision (EU) 2023/1795 of 10 July 2023. For all other US recipients, according to the Austrian Data Protection Authority (in German), other safeguards under Article 46 GDPR are needed, such as standard contractual clauses.

On 3 September 2025, the General Court of the European Union dismissed the action brought by the Frenchman Philippe Latombe against this decision (Case T-553/23). The court expressly confirmed the adequate level of protection for the date on which the decision was adopted. On 31 October 2025, Latombe lodged an appeal with the Court of Justice. According to the Court of Justice register, Case C-703/25 P is still pending as of 11 November 2026.

The position is uncertain for two reasons. Protection against access by US intelligence agencies rests on Executive Order 14086 of 7 October 2022, which is a presidential order and not a law. And the decision refers to the independence of the FTC, which has no longer been protected by statute since Trump v. Slaughter. The European Data Protection Board therefore asked the Commission on 31 July 2026 to examine the effects of the ruling closely. noyb has also announced a lawsuit.

The Court of Justice declared the two predecessors, Safe Harbor and Privacy Shield, invalid in 2015 and in 2020 (Schrems II). If you base your data protection strategy on the current decision alone, you are betting that it will end differently this time.

Is a data centre in the EU enough protection?

No, not completely. Many providers advertise that your data is stored in Frankfurt or Dublin. That is sensible, but it only partly solves the problem.

The CLOUD Act of 23 March 2018 obliges providers that are subject to US law to hand over data in their possession or under their control, and under 18 U.S.C. § 2713 this applies regardless of whether the data is stored inside or outside the United States. So what counts is not where the server is, but who can access the data and which law that company is subject to.

If a US corporation hosts your data in a German data centre, it can therefore still be required to hand it over to US authorities.

How big is the availability risk?

It is real, as June 2026 showed. Fable 5 was blocked for all customers from 12 June to 30 June 2026. On AWS, Google Cloud and Microsoft Foundry it took even longer, according to Anthropic, and from 26 June Mythos 5 initially returned only for selected US organisations. According to Al Jazeera, the US government justified the directive with unspecified national security concerns.

Anyone who had built their processes on exactly these models had to switch to another model or wait. That is an operational risk that has nothing to do with data protection. An open model that you run yourself keeps running even if a provider has to shut down its service. A European provider is not directly affected by a directive addressed to a US company.

Does all data have to stay in Europe?

No. The sensible approach is to sort your data by how much protection it needs.

For non-critical tasks there is little reason not to use American tools. If you have a blog post drafted, collect ideas or rewrite a publicly available text, no data worth protecting is involved.

It becomes critical with personal data, contract contents, source code, and health and HR data. For this category we recommend a provider where the data does not leave Europe and which is not subject to a foreign jurisdiction.

As a rule of thumb, ask what kind of data an AI tool gets to see. Anything you would not show a stranger on the street belongs in a protected environment.

What options are there for AI from Europe?

There are essentially three options, and they can be combined.

For many mid-sized companies a combination makes sense, with non-critical work going through convenient standard tools and sensitive work through a protected European solution.

How do you go about it in five steps?

  1. Get an overview. Find out which AI tools are in use in the company and which data goes into them. Staff often use more tools than management suspects.
  2. Sort your data. Draw a clear line between non-critical data and data worth protecting.
  3. Secure your contracts. For US services, do not rely on the adequacy decision alone. Also agree the European Commission's standard contractual clauses of 4 June 2021. According to the European Data Protection Board, this includes a documented assessment of the legal situation in the recipient country.
  4. Move sensitive data. Put the category worth protecting on European or self-hosted solutions. This is where you gain the most legal certainty and operational reliability.
  5. Document your decisions. Keep a record of processing activities under Article 30 GDPR and a short risk assessment. The exemption for companies with fewer than 250 employees does not apply if the processing is regular. Also note which fallback model you will switch to in an outage.

Conclusion

Keep the data that matters in Europe. Then what the courts decide about transatlantic data transfers in the coming years becomes a side issue for you.

Where your AI provider comes from is a strategic decision for management and not a detail for IT. If you make it deliberately today, you spare yourself unpleasant surprises later.

Fragon Studios supports mid-sized companies in the DACH region in using AI in line with data protection law, from the first assessment to running European and self-hosted solutions. If you would like to know where your company actually stands, get in touch.

Frequently asked questions

Can we use US AI services with personal data?

Yes, if the transfer has a basis under Chapter V of the GDPR. At present it is enough for the provider to be on the Data Privacy Framework List. Otherwise you need other safeguards under Article 46 GDPR, for example standard contractual clauses. You also need a data processing agreement under Article 28 GDPR.

Is the EU-US Data Privacy Framework still valid?

Yes. The adequacy decision of 10 July 2023 is in force as of 11 November 2026. The General Court of the EU upheld it on 3 September 2025, and the appeal against that ruling is pending before the Court of Justice. Since the US ruling in Trump v. Slaughter of 29 June 2026, noyb has been calling for the decision to be withdrawn, and the European Data Protection Board has asked the Commission for a review.

Does a US provider's EU data centre protect against access by US authorities?

Not completely. Under the CLOUD Act of 2018, providers that are subject to US law must hand over data in their possession or under their control, even if it is stored outside the United States. A server location in the EU is a sensible measure, but it does not replace checking which law the provider is subject to.

What happened to Anthropic's AI models in June 2026?

On 12 June 2026, the US government ordered access to Fable 5 and Mythos 5 to be blocked for foreign nationals. Anthropic switched off both models for all customers, while other models remained available. On 30 June 2026 the export controls were lifted, and from 1 July Fable 5 could be used worldwide again.

What fines apply to an unlawful transfer to a third country?

Infringements of Articles 44 to 49 GDPR fall under Article 83(5). Fines can reach up to €20 million or up to 4% of worldwide annual turnover, whichever is higher. The supervisory authority decides how high a fine is in the individual case.

Sources