There's a window of two or three years, and Europe needs to take it!

There's a window, maybe two or three years, where it's still feasible to start a frontier lab from scratch in Europe and have it matter. Every year that passes, more of the architectural and policy decisions that will shape the next decade of AI are made in San Francisco and Hangzhou, with no European seat at the table. The current arrangement, where Europe largely regulates AI it didn't build, isn't sustainable strategically; it produces neither real sovereignty nor confident regulation. SPRIND is the rare European program backing the actual building of frontier capability with serious money and a long-horizon mandate. It's not that 2026 is a special moment but public will for programs of this scale is rare, and I'd encourage serious AI builders to take advantage of it while it's here.

The funding matters, and €125M with a follow-on path toward a billion is a serious commitment by any measure. What I think teams should value alongside it is the time horizon. Two years to do foundational research before being asked to commercialize is unusual anywhere in the world right now, and the program has been deliberately structured to protect that runway. That patience is what allows teams to attempt the hard problems in the first place.

The second piece is institutional weight. SPRIND being the funder changes how a team is read by investors, by senior researchers considering joining, and by the policymakers shaping the environment they'll operate in. It signals that this is serious, publicly backed work, which makes it materially easier to recruit, to raise follow-on capital, and to operate in a public conversation that's often confused about what frontier AI labs actually do. Combined with the compute commitment, the infrastructure, and the cohort of other ambitious teams working in parallel, the program brings together close to the full set of conditions a frontier lab needs to function. The staged design, with increasing commitment as teams prove out their thesis, is one of the more thoughtful structures I've seen for getting a program like this off the ground.

I don't have a clean answer to this, and I'd be wary of anyone who does. What I'd hope to see out of these labs is a result that forces the major labs in the US and China to update their assumptions about what frontier AI has to look like. Either a more sample- or compute-efficient path to current capability, which would be a real economic and strategic shift, or substantive progress on something the field is honestly stuck on; training methods that don't depend on ever-larger data piles, alignment properties that can be verified rather than inferred, or AI for science that genuinely shifts the pace of discovery; a direction where Europe's scientific depth gives it a real chance to lead rather than chase. What I'd be less excited to see is a slightly later, slightly less capable version of work that's already been published elsewhere.

Foundation models look more like infrastructure than consumer products, and the teams building them inherit obligations closer to those of a pharmaceutical company or an airline than a software vendor. That includes serious evaluation work the lab didn't itself design, red-teaming by reviewers outside the company's incentive structure, honest disclosure about training data and what the system is genuinely capable of, and an interpretability investment that scales with capability rather than being added later in response to incidents. There's also a responsibility that sits upstream of any specific model, such as preserving the conditions that make course-correction possible later, in case something does go wrong. That means protecting open scientific exchange, keeping power distributed enough that someone outside the lab can credibly object, and not letting the information environment get flooded with synthetic content past the point of recovery. The riskiest combination to ship into the world is high capability paired with low transparency, and good intentions inside the lab

The hardest part is breaking free of the mainstream's gravity. Generative LLMs have become a kind of black hole. Energy, compute, data, talent, money, it all bends toward the same paradigm, and the innovative alternatives get pulled in and crushed before they ever have a chance to grow.

Europe's challenge is reaching escape velocity: enough funding, persistent patience, and the conviction to chase the architectures the incumbents have no reason to touch.