How Europe can save a trillion euros and win the AI race

How Europe can save a trillion euros and win the AI race

On June 12, the US Department of Commerce informed Anthropic that it had to secure an export license for any foreign person, inside or outside the US, to access its two most advanced artificial intelligence (AI) models. This left the firm with no choice but to disable Fable 5 and Mythos 5 for everyone, with no warning.

While the US lifted the order on June 30, apparently satisfied that Anthropic had addressed the security risk, European businesses, governments and researchers had by then spent 18 days using older models, with no knowledge of when they might regain access.

As Europe processed the realization that its access to frontier AI could be revoked at any moment, another development sent a very different message. The Chinese lab Moonshot AI released Kimi K3 — a 2.8-trillion-parameter model whose weights are now free to download — pointing to the possibility that near-frontier capability might soon cost almost nothing.

Given the impossibility of knowing which of these conflicting signals would prove prophetic, European policymakers must build an AI strategy suitable for both scenarios.

Discussions of Europe’s AI strategy largely rest on the premise that policymakers must choose one possibility and go all in.

Some want to bet on the Chinese signal: AI capabilities would plateau, the distance between the frontier models and their pursuers would narrow, and AI would become a commodity that can be purchased cheaply, much like electricity. This camp advocates a kind of hopeful resignation: accept that Europe lacks the compute, the capital and the talent to win the AI race, but do not worry too much about it, because attempting to reach the AI frontier now would be a waste of money.

On the other side are those who want Europe to sprint for that frontier, with a publicly funded initiative that would dwarf any European industrial program since postwar reconstruction. This camp is betting that AI capabilities would keep compounding, and the gap between frontier models and the rest would continue to widen. In this scenario, resignation is tantamount to economic suicide.

Each position is coherent within the context of its assumed future, but nobody can predict what would actually happen, and neither side has a plan if the future is “wrong.”

Decision theory holds that when two radically different scenarios are plausible, and you cannot reliably assign probabilities to them, the rational move is to choose the course of action that avoids catastrophe in either case — to make a partial investment and buy an option, rather than make an irreversible commitment.

For Europe today, an effective AI strategy would recognize that the components of AI sovereignty run on different timelines. Chips can be acquired in 18 months once capital is secured. Decarbonized power, data centers, model-training expertise and decisionmaking institutions take five to 10 years to build. Frontloading spending on compute amounts to locking in the biggest bet precisely where uncertainty is greatest. A better approach would begin with what takes longest. The resulting assets would have value in whichever future materializes.

To be sure, the risks raised by dependency are real. Frontier AI runs on rationed compute. US labs serve US government contracts first, enterprise customers second and the rest of the world last. When capacity comes under strain, foreign users would be the first to lose access, and geopolitically motivated access restrictions are another risk.

A Europe without domestic inference capacity or ability to train large models is a Europe without a fallback plan.

However, if the risk is losing access to frontier models, the solution is not necessarily to try to surpass the labs that created them. Instead, it is to build the capacity to make surviving that loss possible.

This means securing power and grid-connected land; erecting data centers on European soil, operated by European firms; and building training capabilities, with multiyear procurement contracts ensuring that local expertise keeps pace with technological progress.

None of this requires the budgetary moonshot that some advocate. On the contrary, the combined cost would run to a few billion euros per year, and it could be financed largely by private capital against productive assets, rather than by public commitments based on an uncertain technological premise.

If model performance levels off and AI is commoditized, this foundation would enable Europe to compete on diffusion and integration — arenas where its enterprise-software firms hold genuine advantages. If capabilities keep compounding and access is constrained, Europe would be well-positioned to launch a sprint for the frontier, at far lower cost than if it had to start from scratch, and on a timeline Europe controls.

For a continent that largely stood on the sidelines while Silicon Valley built the digital economy, the inclination to take bold and fast action this time around is understandable, but sequence matters more than scale.

The French nuclear deterrent, often invoked by proponents of an AI sprint, did not begin with a warhead; it took 15 years and proceeded in stages. The same logic applies to AI.

Building an industrial foundation that forecloses nothing, does not break the bank and requires no permission from the US is not a timid approach.

On the contrary, it is the only strategy that does not depend on anticipating the trajectory of a technology whose very creators admit they cannot predict.

Eric Hazan is founding partner of Ardabelle Capital, a co-founder of Plateforme Progressiste and a lecturer at HEC Paris and Sciences Po. Lenny Benbara and Baptiste Lefort are founders of the Global Trade Policy Observatory.