Someone Else Holds the Off-Switch

FOR IMMEDIATE RELEASE Brussels, Belgium, 13 June 2026

On 1 June, after months of lobbying, the EU's cybersecurity agency was finally let in to test Anthropic's Mythos. On 12 June, a directive from Washington switched Mythos off for every foreign national on Earth. Europe's access lasted eleven days. A statement from Ingram Technologies on what that should teach Europe, and where its AI money ought to go.

What happened

At 21:21 UTC on 12 June 2026, Anthropic received a US government export-control directive, issued under national-security authority, ordering it to cut off access to two of its models, Fable 5 and Mythos 5, for any foreign national anywhere, inside or outside the United States, including its own foreign-national staff. To comply, the company disabled both models for everyone, within hours. It said so plainly and made clear it disagrees: by Anthropic's account the government's concern was a narrow jailbreak, the same capability sits in other models including GPT-5.5, and it was complying under protest. On the merits, Anthropic is right. The merits turned out to be beside the point.

Eleven days

Now rewind a fortnight. On 1 June, after European officials had spent weeks publicly demanding it, Anthropic invited the EU's cybersecurity agency, ENISA, to test Mythos under Project Glasswing. It was the first time anyone outside the United States and the United Kingdom had been granted access to the model, and Brussels treated it as the win it was (CNBC, Bloomberg).

Then a single memo in Washington took it back. The access Europe had argued for on the grounds of security and sovereignty turned out to be a courtesy that a foreign government could withdraw without warning, without asking either Anthropic or the EU. There is the whole lesson, inside one fortnight. A seat at someone else's table lasts exactly as long as the host decides it should.

The fight Europe keeps picking

Europe has spent two years having the wrong argument about AI. One camp wants access: get us into the American labs, get us the frontier models, get ENISA a login. The other wants a champion: pour billions into a single European lab built to stand toe to toe with OpenAI and Anthropic at the bleeding edge. This month should retire both. Access gets switched off, as we just watched. And the champion plan asks Europe to win a spending war it has already lost: European public AI funding runs to roughly €200 billion across a decade, while the American hyperscalers are spending north of $400 billion in a single year. That gap does not close with one more national champion. Frontier training is a capital bonfire, paid mostly by American investors and burned mostly on American soil, and Europe should stop organising its ambitions around the fantasy of winning it.

What Europe should build instead

The most expensive part of AI, training the frontier, is being paid for by someone else, and the returns leak out. The research gets published. The methods get copied within months. The weights, more and more, get opened to anyone who wants them. China's labs proved what that makes possible: DeepSeek reached within striking distance of the frontier for a fraction of what the American labs spent, by building efficiently on research those labs had already paid to produce. Compute still buys a real lead. What DeepSeek showed is that the lead is far cheaper to chase than to hold, and that being a fast, smart second costs a sliver of what it costs to be first.

Be clear-eyed about the catch, because this month put it on display. That commons is a choice other people make, and they are starting to close it. The chips to train and serve these models still run through American suppliers. Free-riding on someone else's openness has no future, which is exactly the argument for building the parts Europe can actually keep. The defensible ground is the layer one step behind the frontier: applied systems, fine-tuned and domain-specific models, open-weight derivatives held on your own hardware, evaluation and safety, sovereign deployment, and the governance that real industries need before they will touch any of this. Weights you already hold cannot be switched off by memo. Pair that layer with shared European compute, the one piece worth centralising, and you have capability that survives a directive out of Washington.

So the policy worth arguing for is unfashionable in its modesty. Stop hunting for the European OpenAI. Spread the money: fund a hundred small and mid-sized R&D bets across a hundred teams and a dozen member states, expect most to fail, and let the handful that land define the European AI economy. Diffuse capability you own is far harder to coerce than a single champion sitting on a borrowed login. This is the bet Ingram Technologies has made as a lab, and the work we spend our days on.

Call it industrial policy for optionality: many owners, many fallbacks, much of it open, all of it under European jurisdiction. Europe can buy the best American model on Monday and still refuse to be its hostage on Friday. The off-switch was a cheap lesson this time, two models, with the rest of Anthropic's lineup still running and the provider itself fighting the order. The next one will cost more, and it will not arrive with eleven days of warning. Spend the coming decade building, and the next memo out of Washington becomes somebody else's emergency.

Jerome Leclanche
CEO, Ingram Technologies

About Ingram Technologies

Ingram Technologies is a European AI research and development lab, founded in 2022 and based in Brussels. We build applied, agentic AI systems and advise on AI governance and ISO/IEC 42001 compliance, with a focus on responsible, sovereign AI for European organisations.

Media enquiries: press@ingram.tech