# Washington just showed Europe how quickly an AI dependency can become a crisis > Source: > Published: 2026-06-20 10:55:37+00:00 *Washington's order against Anthropic did more than interrupt access to two AI models. It showed European companies how quickly a foreign dependency can become someone else's policy lever.* The uncomfortable lesson from the Anthropic shutdown is not that Europe needs a slogan about sovereign AI. It has plenty of those. The lesson is that if your hospital, bank, software team or public agency relies on a model controlled abroad, access can disappear before your procurement team has finished reading the email. According to The Verge, the Trump administration ordered Anthropic in June 2026 to cut off foreign nationals from its newest models, Fable 5 and Mythos 5, citing national security concerns and fears that safeguards around cybersecurity capabilities could be bypassed. Anthropic then blocked access more broadly while it fought to get the models restored. Business Insider reported the same basic sequence: export controls hit the two models, foreign access was barred, and Anthropic disputed the urgency and scope of the government concern. That is current, and it is not a small paperwork story. Axios reported this week that the move has worried AI companies because customers now have to ask a harder question before signing long contracts with American labs: can you guarantee the model will still be there if Washington changes its mind? You don't need to be anti-American to ask that. You only need to run a business that can't have core systems switched off by a government you don't vote for. Europe has been warned before, but this warning came with a working example. The Associated Press reported earlier this month that the European Union had launched a tech sovereignty package covering chips, cloud and AI, with officials arguing that Europe must avoid risky dependence on one supplier, one company or one third country. The package includes a follow-up to the Chips Act, support for cloud and AI infrastructure, and plans to triple EU data center capacity within five to seven years. Fine. But plans don't run models. The gap is visible in the hard numbers. The Guardian, citing Stanford University, reported that US institutions produced 40 notable AI models in 2024, compared with 15 from China and just three from Europe, all of them French. The European Commission has put a bigger number on the answer, announcing its €200 billion InvestAI initiative in February 2025, including a €20 billion push for AI gigafactories. Those figures sound large until you compare them with the scale of American private spending on cloud infrastructure by Microsoft, Amazon, Google and Meta, you name it. The US lead is not just policy. It is steel, land, power contracts and chips already being ordered. ## Europe's problem is power as much as policy Compute is not an abstract word here. The International Energy Agency estimated that data centers consumed about 415 terawatt-hours of electricity in 2024 and projected that the figure could rise to roughly 945 terawatt-hours by 2030. If Europe wants its own frontier AI capacity, it needs grid connections, cooling systems, permits, chips and customers willing to pay more for resilience. Brussels can't regulate those into existence after a press conference. This is where the AI Act debate becomes more serious than the usual argument about whether Europe regulates too much. The question is not whether AI needs rules. It does. The question is whether Europe can enforce rules on systems it does not control, running on infrastructure it did not build, owned by companies whose most important political risk sits in Washington. Frankly, that is not sovereignty. It is compliance with a rented engine. Mistral has been waiting for this moment. Business Insider reported that the Paris-based startup has long pitched European customers on AI sovereignty, with chief executive Arthur Mensch warning against dependence on US model providers. Its open-weight approach gives governments and companies more room to run systems on their own infrastructure. That doesn't make Mistral a full substitute for Anthropic, OpenAI or Google DeepMind today. It does make the argument harder to dismiss. When access can be cut by executive order, local control stops sounding like a nationalist hobby. There is a risk of overcorrecting. Europe doesn't need to pretend it can build every chip, cloud platform and model family inside the bloc. That would be expensive, slow and probably false. But it does need strategic redundancy for the systems that matter: public health, defence, financial plumbing, energy, legal services and the software stacks sitting underneath them. If those systems depend on foreign models, you need a fallback that has been tested before the outage, not a committee formed after it. The Anthropic order may be temporary. The Verge reported that the legal basis remains unclear and that the company has been trying to get its models back online. That uncertainty is exactly the point. If the rules are unclear even to the company being regulated, a European customer has no serious basis for treating access as guaranteed. **Also read:** [Apple's new Siri arrives as a serious agentic AI and the fight for your iPhone's assistant layer just got real](https://startupfortune.com/apples-new-siri-arrives-as-a-serious-agentic-ai-and-the-fight-for-your-iphones-assistant-layer-just-got-real/) • [South Korea's AI chip bonuses are about to become a monetary policy problem](https://startupfortune.com/south-koreas-ai-chip-bonuses-are-about-to-become-a-monetary-policy-problem/) • [Private capital has crossed $13 billion in fusion energy bets and the field is just getting started](https://startupfortune.com/private-capital-has-crossed-13-billion-in-fusion-energy-bets-and-the-field-is-just-getting-started/)