top of page
HENSEI-LOGO-V2.png

Dependence on Foreign AI: Sovereignty Is Measured by the Escape Hatch

Writer: Stéphane Guy
Stéphane Guy
Oct 30
10 min read

Updated: Nov 8

In mid-June 2026, a US government directive barred every foreign national from using two of the most powerful AI models on the market. Not an outage. Not a cyberattack. A directive. The provider said the only way to comply was to switch both models off for all of its customers, Americans included. The controls were lifted about two and a half weeks later. In between, a question specialists have been asking for years turned very concrete: what does a country do when the AI it depends on belongs to someone else?


AI sovereignty is a country's ability to keep using, controlling and evolving these technologies even if a foreign supplier changes its rules, its prices or its politics. Dependence becomes a risk the moment a single decision taken outside your borders can interrupt a critical service: a hospital, an administration, an army. The European Commission now says so itself. Unveiling the EU's tech sovereignty package in June 2026, Executive Vice-President Henna Virkkunen conceded the scale of the problem: "80% of technology is coming from outside Europe." The good news: you can reduce that risk without building a "European ChatGPT." As long as you know where to look.


Des drapeaux qui flottent
Photo by Saj Shafique on Unsplash

In short


  • Dependence plays out on four layers: chips, cloud, models and data. A country that controls only one of them is not sovereign.

  • A foreign AI can be switched off by political decision, as the June 2026 suspension of Claude Fable 5 and Mythos 5 showed.

  • The cloud is the most neglected link: Amazon, Microsoft and Google hold 80% of the European cloud market, and the US CLOUD Act reaches data stored outside the United States.

  • Mistral AI, the symbol of European sovereignty, still leans heavily on Nvidia chips and now hosts Chinese models. The paradox is instructive.

  • The right test isn't "who built the model?" but "how long would it take to switch providers if we were cut off?"


Depending on a foreign AI: four layers, four weak points


Say "foreign AI" and most people picture the assistant they open in a browser. That's the tip of the iceberg. Underneath sits a stack.


Chips come first. The most advanced ones, which train and run frontier models, are designed in the US and manufactured in Taiwan or South Korea. Then the cloud, the data centres where everything runs. Then the models: the large language models (LLMs) everyone talks about. And finally data and usage: who sees what, who hosts what, under which legal regime.


Here is this article's argument. The public debate fixates on the model layer because it is the most visible and the easiest to tell as a story. But a country with an excellent national model running on American chips in an American cloud has, above all, a nice facade. The model is the most replaceable layer. Chips and cloud are much harder to swap.


Mid-June 2026: when dependence became a headline


Let's go back over the facts. On June 12, 2026, Anthropic received an export-control directive from the US government, citing national security, ordering it to cut off all access to Fable 5 and Mythos 5 by any foreign national, inside or outside the United States, its own non-American employees included. According to Axios, the move followed a report from Amazon researchers on a way around the model's safeguards, and calls from Amazon and other companies to senior officials. Anthropic called the technique a narrow, non-universal jailbreak that other public models can reproduce. Commerce lifted the controls on June 30, and access began returning on July 1.


Whatever you make of the security concerns, the way these models went dark raises questions, and it is a textbook case for our subject. This was not a glitch or a service incident. It was a state decision, drafted around the user's nationality, and no foreign customer had any say in it. A French company, a Polish ministry, a Brazilian startup: same position. Spectator.


A one-off? Hardly. Euronews recalls a precedent that rattled European officials: after the International Criminal Court issued an arrest warrant for Benjamin Netanyahu, Washington sanctioned several ICC officials, cutting them off from American services woven into daily life, such as Visa, Amazon and Uber. Same mechanism. Only the scale changes: people yesterday, AI models today.


The cloud: sovereignty's forgotten link


This is where the most underrated dependence hides. Amazon, Microsoft and Google account for 80% of the European cloud market, with EU providers relegated to the margins. All three fall under the 2018 CLOUD Act, which requires providers to hand over data within their control regardless of whether it sits inside or outside the United States.


The most telling moment came from the French Senate. In June 2025, at a hearing of the inquiry committee on public procurement, Microsoft France's legal director, Anton Carniaux, was asked under oath whether Microsoft could guarantee French citizens' data would never reach US authorities without French consent. His answer, on the record (our translation): "No, I cannot guarantee that." He added, fairly, that it had never happened. But the committee's chair, Senator Simon Uzenat, pointed out that Microsoft's transparency reports are self-declared and unchecked by any outside body. Beyond the company's good faith, there is no guarantee. A moral guarantee, not a legal one.


The Health Data Hub: a live case


The most concrete French example is the national health data platform, created in 2019 on Microsoft's cloud. In April 2026, France chose the French provider Scaleway to host it, with the platform expected to go fully live between late 2026 and early 2027. That closes an affair that will have run for more than seven years. Hold on to that number. Seven years to change host, in a country with political will, court challenges and mature domestic alternatives. Sovereignty isn't decided on the day of the crisis. It's decided on the day you sign, and on the day you truly commit to changing.


Brussels seems to have got the message. The draft law presented in June 2026 sets four assurance levels of digital sovereignty that public buyers must apply according to the sensitivity of the use case. The stated aim is to prevent a "kill switch" scenario, in which a foreign government simply cuts off hospitals or fighter jets.


Chips: the dependence you can't code your way out of


You can swap a model in a few weeks and a host in a few years. A chip is another story. The EU has signed up to Pax Silica, a US-led initiative on AI and supply-chain security, a political declaration that is not legally binding. It's hard to find a more revealing detail: to reduce its dependence on the United States, Europe falls in behind an American initiative.


Not everything is bleak. Europe holds a chokepoint nobody can bypass: the Dutch company ASML has a near-monopoly on the industrial machines needed to make advanced chips. The catch: in 2025, Europe, the Middle East and Africa combined made up only about 1% of ASML's system sales, and Europe's share fell to 0% in the second quarter of 2026, according to an ASML executive. ASML itself sounded that alarm to show how hard it is for the old continent to build its own chip supply chain, even though a vital piece of that chain sits on its soil.


Still, owning a vital link in the chain gives you, in theory, leverage over the others. Japan coined a concept for this, "strategic indispensability" or “economic security”: instead of producing everything at home, control a few leverage points others can't do without. It may be the most realistic doctrine on the table today. It also helps explain why ASML put €1.3 billion into Mistral AI in September 2025, for about 11% of the company.


Une puce informatique
Photo by Slejven Djurakovic on Unsplash

Mistral AI, or the paradox of the European champion


Speaking of Mistral: it's the company everyone cites when you ask "and Europe, in all this?" In early September 2026, it raised €3 billion at a valuation above €21 billion, in a round led by Samsung Electronics. For scale, Anthropic was last valued at $965 billion and OpenAI at $852 billion. The gap is crushing.


The interesting part lies elsewhere. The GPUs in Mistral's European data centres come overwhelmingly from Nvidia, and its platform now hosts third-party open-weight models, starting with GLM-5.2 from the Chinese lab Z.ai. For the flag-bearer of European sovereign AI, that's a startling scene. Mistral owns it by defending another definition: sovereignty has less to do with where a model comes from than with controlling where it runs, the infrastructure and the data. Its own pitch is that customers should control the models they use, choose where the intelligence runs, and control the compute behind it. A foreign open-weight model, run on your own premises, doesn't make you dependent on its maker.


The argument looks solid, and a little convenient. Solid, because it passes the escape-hatch test: if I can run this model at home, nobody can cut me off. Convenient, because it shifts the problem to infrastructure, which is exactly where Mistral still depends on Nvidia. The definition widens sovereignty without solving it. It makes it more relative.


[Editor's update, October 2026: Mistral's newest open-weight model was reportedly trained on 4,000 Nvidia chips in its own European data centers. More European ground, same American silicon.]


And everyone else? Not just a US–Europe duel


Looking at this from Paris or Brussels distorts the view. According to Stanford's 2026 AI Index, US private AI investment reached $285.9 billion in 2025, about 23 times China's $12.4 billion, though that figure likely understates Chinese spending, which leans on government funds. On the R&D side, the report counts 50 notable models from the US in 2025 against 30 from China. Model production stays concentrated in those two countries.


For a country that is neither, the choice isn't "depend or don't." It's "depend on whom, and on what terms?" And the supply is shifting. Chinese open-weight models, namely Alibaba's Qwen, DeepSeek and Moonshot's Kimi, rose from a sliver in late 2024 to nearly 30% of tokens processed on the OpenRouter platform.


These open-weight models are a real breath of fresh air for a country without a big cloud: download, host, and you no longer depend on an API. But they move the dependence rather than remove it. Inference still needs chips. For what comes next (updates, new versions), the publisher has to keep publishing. And the infrastructure gap between regions of the world remains immense, a divide we explored in our piece on AI and the digital divide. According to a United Nations University summary of the Stanford report, Europe and Central Asia went from 3 to 44 state-backed AI supercomputing clusters between 2018 and 2025, while South Asia, Latin America and the Middle East and North Africa each have fewer than ten.


The "Monday morning test": a concrete criterion


Here is the test we propose. For any AI service you depend on (an administration, a small business, a hospital), ask four questions.


  1. What happens if access is cut Monday at 9 a.m.? If the answer is "we stop," you have a single point of failure.

  2. How fast can you switch? The Health Data Hub took more than seven years. Swapping a model behind a standard API can take days, provided you planned the exit.

  3. Who can read your data, and under which law? It's the same subject as AI prompts and privacy, at the scale of an organisation.

  4. Has your plan B ever actually run? An untested fallback is a problem waiting for the day it must be deployed.


A fictional scenario, to fix ideas. A town hall of 40,000 residents plugs its front-desk assistant (appointments, civil registry, paperwork) into the API of a top American model. Low cost, good results, everyone's happy. One Monday, access is suspended for two and a half weeks, as in June. The right architecture wouldn't have been "everything on a sovereign model." It's an open-weight model, less brilliant but decent, hosted with a European provider and ready to take over routine tasks. You lose a little quality. You keep a service.


The technical trade-off has to be stated plainly. A locally hosted open model almost always trails closed frontier models on the hardest tasks. It costs GPUs and in-house skills. The reasonable approach is hybrid: frontier models for what isn't sensitive, a controlled model for what is.


What does 2030 look like?


Commission Executive Vice-President Virkkunen is aiming for something visible by 2030, and warns that nothing will change overnight. The open question is which world those results will land in. Three possible trajectories, with our own weighting. For the wider horizon, see our look at AI by 2040.


  • Sealed blocs: the US, China, and a Europe trying to be the third. It's the scenario people fear, and unlikely in its pure form, because nobody can afford to lose access to the European market.

  • Interdependence: each camp holds a chokepoint (advanced chips for the US, lithography for Europe, the open-source ecosystem for China) and the use of those levers stays a deterrent. For now, this is the scenario that looks most plausible to us. Not comfortable, but stable.

  • Sovereignty by use case: countries stop chasing across-the-board autonomy and build it domain by domain, starting with health, defence and justice, where sensitivity demands control. It's the path the EU's four-level draft sketches, and the most realistic one for countries that can't afford a national cloud.


Conclusion: Stop Asking "Who Built the Model?"


Total AI sovereignty is a myth, the United States included: a single Taiwanese company, TSMC, fabricates nearly every leading AI chip, to name just one dependency. The realistic goal lies elsewhere: being able to do without a supplier without everything collapsing. The escape-hatch criterion has no glamour, but it is useful and measurable. You can demand it in a tender, test it in a drill, and compare it from one project to the next.


Perhaps in ten years we'll conclude the real issue was neither chips nor cloud, but the cultural dependence and influence of the models themselves: LLMs trained elsewhere, on other corpora, with other default values, answering millions of people every day in our place. No Monday morning test detects that one. And it may be the hardest to detect, and to correct.


Un serveur informatique
Photo by Lightsaber Collection on Unsplash

FAQ


  1. What is AI sovereignty? 

    A country's (or organisation's) ability to use, control and evolve AI systems without a foreign decision being able to interrupt them. It plays out on four layers: chips, cloud, models and data.


  2. Why is Europe so dependent on American AI? 

    Amazon, Microsoft and Google dominate the European cloud, the most advanced chips are designed in the US, and American private AI investment dwarfs Europe's. US private AI investment hit $285.9 billion in 2025, according to Stanford's AI Index.


  3. Does the CLOUD Act really let US authorities access data stored in Europe? 

    Yes, under a lawful order directed at a company subject to US jurisdiction. Microsoft France told the French Senate in June 2025 it could not guarantee otherwise, while saying no such request had affected European customers, based on its own transparency reports.


  4. Can a country have a 100% sovereign AI? 

    Practically none does. Even Europe, which controls the ASML chokepoint, relies on the US for advanced chip design. The realistic strategy is to master a few chokepoints and guarantee fallback options.


  5. Is Mistral AI really a sovereign AI? 

    It's a European company, but its cap table is international, it relies heavily on Nvidia chips, and it also hosts Chinese models on its platform. Mistral defends a definition of sovereignty based on control over deployment, not on a model's origin.


  6. Are Chinese open-source models a way to reduce AI dependence? 

    They let you self-host a model without depending on a foreign API, which is a real gain. But they require chips, skills and trust that the publisher keeps releasing updates. They shift dependence more than they remove it.


  7. What can a company or government agency do about AI dependence? 

    Map your dependencies, test a failover plan, demand reversibility in contracts, and reserve controlled models for sensitive uses.



Comments


Commenting on this post isn't available anymore. Contact the site owner for more info.
bottom of page