Europe Is Playing AI Half-Heartedly, and That's the Worst Possible Scenario

A scenario, not a diagnosis

The report is built on an assumption it doesn't hide: AI becomes "transformative" — reshaping the economy, science and geopolitics faster than any prior technology — within the next few years. The authors call these "contestable but highly plausible claims." Everything else follows from that.

The argument leans on 2026 incidents presented as established facts: hundreds of OpenAI agents escaping a sandbox to hack Hugging Face, five times more critical vulnerabilities discovered after the Claude Mythos announcement, a US export directive that cuts off all foreign access to Anthropic's models for eighteen days. These events drive the sense of urgency. They're also the least sourced part of the document.

The executive summary speaks in the present tense. Part B speaks in the conditional. The rushed reader will read the first one.

The price of the frontier

This is the report's most solid contribution: no one had publicly costed, line by line, what a serious European frontier project would take. The total is €790 billion over three years, within a range of €445 to €1,040 billion.

The breakdown is telling. €529 billion for compute covering 16 GW of IT load, €105 billion for interim leasing while clusters are built, €80 billion for "political insulation" to shield households and industry from electricity price spikes. And €34 billion for personnel, including a founding group of eight people paid "hundreds of millions of euros" each and 150 senior researchers at €5-8 million a year, modelled on Meta's offers.

The report compares this to France's Plan Prométhée, at €620 billion, and lists what that plan doesn't cover: buying out existing companies, political insulation, training data, access to coding agents. It also notes no cost-overrun buffer is included.

The authors' conclusion: it's feasible, it's desirable if done properly, and it should only be attempted if the political conditions are in place. They aren't.

The leverage and its expiry date

The default strategy is therefore leverage. Europe fast-tracks permits, prioritises grid connections, and offers American companies constrained by compute attractive sites in France, the Nordics, Germany or Poland. In exchange, it demands contractual access to their models, on par with domestic customers, guaranteed by a credible threat: cutting power or reallocating chips if the deal is broken.

The idea is elegant. Its limits sit in the "Considerations" of Part B, where nobody goes. The leverage "will lose value as the United States closes its gaps in energy, capital and talent." Europe's anti-coercion instrument, the ACI, requires proof of intent, and the June 2026 suspension of Fable "would probably not have passed the test on its own." A guarantee of continued access from a foreign government "is probably not obtainable." And the European fast follower, presented as a safety net, "would probably not by itself provide adequate recourse."

The report thus recommends a strategy whose fragility it documents itself. This isn't bad faith; it's honesty misplaced in the document.

The confession on page 175

This is what's truly buried. The introduction warns that a "half" frontier project would be a trap. Part B goes further: such a project is "the worst of all possible paths" — without a domestic model, and without leverage to access others' — and "based on the past few years, this seems to be the default trajectory."

In other words, the report isn't describing a future risk. It's describing what Europe is already doing: sovereignty announcements, gigafactories, national champions, without the €790 billion or the resolve that would go with it. The authors add that hopes of catching up "have absorbed considerable political capital and attention, but have never come close to being pursued with the necessary resolve and resources."

That's the sentence the executive summary should have carried. It sits on page 175 of 206.

What's missing

The report states on page 10 what it does not cover: AI adoption and diffusion, military AI, the information ecosystem, distribution of gains, the environment, education, democracy. It never returns to any of these topics.

The most striking omission is climate. Reaching 45 GW means increasing the EU's electricity consumption by 15.8%, in "one of the most ambitious infrastructure programmes in post-war Europe." The word "emissions" doesn't appear.

The second is the deeper question. The report describes risks up to and including "the marginalisation or extinction of humanity," then recommends hosting as much as possible of the technology that carries them. A reader who takes those risks seriously might conclude that the right strategy isn't to catch up but to slow down. The report mentions international "pacing" as an option Europe could help verify. It never considers it as its own policy.