On 16 October 2026 China published draft rules that would let Beijing decide when its most capable AI models may leave the country. In June, Washington made an American company switch off two of its models for foreign users, Europeans included. Neither step has cut Europe off for good. Both show that the best AI Europe can run today already depends on decisions taken in other capitals.

What China’s draft export rules do

On 16 October 2026, China’s Ministry of Commerce and Ministry of Science and Technology published a joint draft, Announcement No. 48 of 2026. It would add forty technologies to China’s official list of technologies that need government permission before they can be sold or transferred abroad. Twenty-three of them concern AI. Anyone may comment until 15 November 2026.

The main item covers large AI models. A model falls under the rules if it meets any one of three tests.

  • It took at least 10²⁶ calculations to train, a scale only the very largest models reach.
  • It has two trillion parameters or more.
  • It ranks in the world’s top ten on public tests of AI performance.

Parameters are the settings a model learns during training. Their number is a rough measure of a model’s size.

Three rules in the draft matter most for Europe.

  • Models already published are not affected. A model is published when its weights, the files that make up the trained model, are free for anyone to download. Europeans can keep using every Chinese model they have already downloaded.
  • Publishing a new model that meets the tests counts as exporting it. The company must register it and receive what the draft calls an “automatic licence”. The draft sets no deadline for granting this licence and no grounds for refusing it.
  • Installing a covered model on a customer’s own servers outside China needs a licence. Using it through an app, a website, or a paid online service does not.

The draft cuts nothing off today. It gives the Chinese government a legal way to stop the next generation of Chinese models from reaching servers anywhere outside China, including in Europe. Beijing has so far only written its rules down. Washington’s own rule on model weights is still on its books but not enforced. It acted instead through a private letter to a single company. That letter cut off exactly the kind of use that China’s draft leaves open, through apps and online services.

Washington has already used this power

In June 2026 the United States stopped every non-American user, Europeans among them, from using two of the most capable AI models in the world. On 12 June the US Secretary of Commerce, Howard Lutnick, wrote to Anthropic, the company that makes them. He told it to stop every foreign national, wherever they were, from using Claude Fable 5 and Claude Mythos 5.

Anthropic had no way to check each user’s nationality, so it switched both models off for all its customers. The order was lifted on 30 June, but the Commerce Department kept the right to reimpose it. Claude Fable 5 came back on 1 July, while Claude Mythos 5 returned only for a group of US organisations that defend critical infrastructure.

Two days after the letter, a Commission spokesperson, Thomas Regnier, said the Commission was “assessing its implications, including for users in the European Union”. The models had been released on 9 June, only three days before the letter. For almost three weeks after that, European customers could not use them and had no say in when they would return.

What Europe can run on its own

The strongest model built in Europe is Mistral Large 4, which the French company Mistral AI released for testing on 6 October 2026. Artificial Analysis, an independent firm, combines many tests of reasoning, knowledge, mathematics, and coding into a single score for each model. A higher score means a more capable model. The table compares Mistral Large 4 with the best American model and the best Chinese model anyone can download. It also includes Kimi K3, the largest Chinese model and big enough to meet China’s size test. The last row is Mistral Large 3, the best European model anyone can download today.

Model Maker Country Score Free to download
Claude Opus 5.5 Anthropic United States 58 No
MiMo-V2.6-Pro Xiaomi China 46 Yes
Kimi K3 Moonshot China 44 Yes
Mistral Large 4 Mistral AI France 38 Promised by 31 October 2026
Mistral Large 3 Mistral AI France 9 Yes

Scores from the Artificial Analysis Intelligence Index, as of 18 October 2026.

Mistral Large 4 reaches about two-thirds of the best American score. Once Mistral AI publishes its weights, it would rank around eighth among models anyone can download. Mistral AI has not yet said under what licence it will publish them, and strict conditions would limit who in Europe can use the model freely.

Until then, the largest European model anyone can download is Mistral Large 3, from December 2025. Its licence, Apache 2.0, lets anyone use, change, and sell it.

Beyond Mistral AI, Europe’s models are much smaller. Kolibri-1, from the German company Aleph Alpha, has seventy-eight billion parameters, and the Swiss AI Initiative’s Apertus has seventy billion, against more than a trillion for Mistral Large 4. OpenEuroLLM, a project funded by the EU, planned its first models for July 2026. I found no sign that they have been released.

What sits underneath the model

Even a European model rests on foreign decisions. Mistral AI trained Mistral Large 4 on about 4,000 chips made by NVIDIA, an American company. These chips fall under American export rules, and I found no European AI laboratory that trains on chips from anywhere else.

Those rules still require a US licence to export top-tier AI chips to any country in the world. Washington has said it will enforce this only for buyers headquartered in China, Macau, and a few other countries. Europe’s supply therefore rests on a decision not to enforce, which Washington can reverse without a new law.

The money follows the same pattern. Mistral AI’s €3 billion funding round, announced on 8 September 2026, was the largest ever for a European technology company, and South Korea’s Samsung Electronics led it. The Scaleup Europe Fund, set up by the European Commission and run by the Swedish investment firm EQT, shared the lead. NVIDIA is also one of Mistral AI’s investors.

Ownership is moving as well. On 16 September 2026, Aleph Alpha agreed to merge with Cohere, a Canadian company. The combined company will carry the Cohere name and be run from Toronto and Berlin.

Europe’s plan for its own very large computing centres has not yet moved past the call for bids. The EuroHPC Joint Undertaking, the EU body that buys supercomputers, asked on 30 July 2026 for bids to build up to seven “AI Gigafactories”. Bids close on 3 December, three weeks later than first planned, and the winners are expected in early 2027.

Why this is a problem today

The United States has at least six AI laboratories at the frontier, meaning those whose models are among the most capable in the world, and China has at least five. Europe has one, Mistral AI.

Funding shows an even bigger gap. Counting only money paid for shares, OpenAI has raised about €162 billion and Anthropic about €122 billion. Mistral AI has raised about €4.7 billion. For every euro Mistral AI has raised, OpenAI has raised about thirty-five and Anthropic about twenty-six.

That puts Mistral AI level with the leading Chinese start-ups rather than the American ones, as DeepSeek, Zhipu, and Moonshot have each reportedly raised between €5.5 billion and €6.9 billion. Google DeepMind and Meta’s AI laboratory raise no money of their own. Their parent companies, Alphabet and Meta, each plan to spend well over $100 billion on data centres and equipment in 2026 alone.

The problem is not a future one. The best AI a European organisation can run today either comes from abroad or rests on American chips and foreign money. Two governments now have the legal means to limit it, and one of them has already used those means. This matters for Europe’s prosperity, not only its security. In his 2024 report on European competitiveness, former European Central Bank President Mario Draghi wrote that “Europe largely missed out on the digital revolution led by the internet and the productivity gains it brought”. With the spread of AI, he argued, “a window has opened for Europe to redress its failings in innovation and productivity”.

The AI Strategic Autonomy Framework (AI-SAF), which I maintain, is built to answer one question: how far a country or company can keep its AI running if a foreign supplier or government cuts it off. The events of this year touch two of the dependencies it tracks: the model itself and the chips it runs on.

How Europe could build a second laboratory

A competitive European model would give European companies a safer reason to build AI into their core work, and there is plenty of room for that. According to the Draghi report, 30% of large EU businesses had adopted AI in 2023, but only 7% of small and medium-sized ones. A supplier that no foreign government can switch off lowers the risk of taking that step, as European users of Claude Fable 5 found in June. With two competitive European laboratories rather than one, that safer option would not depend on a single company’s prices, licence, or survival.

Europe has paid for some attempts at a second laboratory. In June the Commission chose a group led by the Italian company Domyn to train a leading model on EU supercomputers. AMI Labs in Paris raised $1.03 billion for AI that learns from the physical world rather than mainly from text.

Germany’s innovation agency, SPRIND, is funding up to ten teams. None of these efforts has yet released a model that ranks near the top.

No single European fund can pay for a frontier laboratory on its own. The Scaleup Europe Fund started with about €3 billion towards its €5 billion target, €1 billion of it from the EU budget. It must invest on the same terms as a private investor. Public development banks also follow private money. Germany’s KfW, for example, invests directly only when private investors provide at least half of a funding round. Norway’s sovereign wealth fund, Europe’s largest pool of capital, is not allowed to buy shares in unlisted companies at all.

Mistral AI shows how the money can be put together instead. A large company led each of its last two rounds, ASML in 2025 and Samsung Electronics in 2026. Public funds such as the Scaleup Europe Fund and Bpifrance joined them, and banks lent $830 million to buy chips. A second laboratory would need the same mix.

The cost is high, and it does not stop. Epoch AI, an independent research group, estimates that a computing centre that draws one gigawatt of electricity costs about $38 billion to build and about $8.5 billion a year to own and run. On my rough calculation from those figures, Mistral AI’s own target of 200 megawatts by the end of 2027 would cost about $1.7 billion a year.

Money on the scale Mistral AI has raised can get close to the top. Two Chinese laboratories with similar funding, Zhipu and Moonshot, have built models that score above Mistral Large 4, though they still trail the best American ones. Money alone was not the whole story for them, and it would not be for Europe either. Rebecca Arcesati co-leads a research programme at the Mercator Institute for China Studies (MERICS) in Berlin. She notes that Chinese laboratories still lack the computing power to train the largest models, because of US export controls and limits on Huawei’s own chips. Several other things need to be in place.

  • Computing power. This depends on the AI Gigafactories described above.
  • Chips. Supply depends on Washington, as described above.
  • Energy. According to the Draghi report, EU companies pay electricity prices two to three times those in the United States.
  • Data. The same report finds that EU rules on storing and processing data make it hard to build the large data sets needed to train models.
  • People. Laboratories must be able to pay and keep researchers who could work anywhere.
  • Customers. Revenue from European users is what pays for the next model.

In my view, Europe should take four steps.

  1. Back one or two existing European teams rather than start from scratch.
  2. Put the money together the way Mistral AI did, with a large company in the lead and public funds alongside it.
  3. Set aside computing time in the new AI Gigafactories for those teams, so that they do not depend on American cloud providers.
  4. Commit for several years, not for one round. Staying near the frontier costs billions of euros every year, not once.

What to watch

The coming months will show whether Europe’s position changes.

  • Overdue – OpenEuroLLM’s first models were due by 31 July 2026.
  • By 31 October 2026 – Mistral AI is due to publish the weights of Mistral Large 4, and the licence that comes with them.
  • 15 November 2026 – China’s comment period closes. The final text may change the tests or the “automatic licence”.
  • 3 December 2026 – bids close for the EU’s AI Gigafactories.
  • Early 2027 – the Gigafactory winners are named. It will matter whether any computing time is set aside for European model builders.

Europe cannot take back the powers that other governments now hold over the AI Europe uses. A second laboratory that can reach the frontier would give it somewhere to turn if the first is sold, falls behind, or tightens its licence.

Matching the United States and China would take money on a scale Europe has not yet committed.

Sources

Total funding figures are my own sums of announced share rounds.