Europe can hedge its AI future

Europe’s AI strategy faces two incompatible signals. On June 12, the US Department of Commerce told Anthropic it needed an export license for foreign access to its two most advanced models, forcing the company to disable Fable 5 and Mythos 5 for affected users. The order was lifted on June 30, but European businesses, governments and researchers spent 18 days using older models without knowing when access would return. At the same time, Moonshot AI released Kimi K3, a 2.8-trillion-parameter model with weights free to download, suggesting powerful models could become far cheaper.

A durable European approach would avoid betting fully on either a commoditized AI future or an accelerating frontier race. Chips can be acquired in 18 months once funding is in place, while decarbonized power, data centers, model-training expertise and institutions take five to 10 years to build. Prioritizing those longer-term assets would preserve flexibility without locking Europe into the costliest uncertainty.

The proposed path centers on domestic inference capacity, European-operated data centers, secured power and land, training capabilities, and multiyear procurement to develop local expertise. The expected cost would be a few billion euros per year, supported largely by private capital. If AI plateaus, Europe could compete on diffusion and integration; if frontier access tightens, it could accelerate from an existing industrial base rather than start from scratch.