Europe's bitter lesson

There is no doubt that Europe has fallen behind in the AI race, with the US constantly releasing new frontier models and China nipping at their heels. And yet, most talk around AI sovereignty seems focused on running open models under European jurisdiction and avoiding the feared kill switch. This should be the least of our concerns. We only have access to these models because it’s currently in China’s best interest to distribute them openly and counterpose themselves against the US frontier labs.

It’s clear that once the frontier models grow strong enough to become economically significant, they’ll be restricted to serve their own economic interests. We will be left with the models of yesterday as the race continues on without us, and the US and Chinese economies enter new phases of economic growth that will forever be out of our reach.

Europe’s only horses in the AI race are the handful of neolabs we have produced. But with their relatively meagre levels of funding, they will have to bet on novel approaches or niche applications ¹, as opposed to raw compute and scaling. Although these bets are worth taking, they run against the bitter lesson ². Compute at scale is the most likely way that AI will continue to improve, but Europe is not betting on it.

If Europe wants to compete, there is really only one way to do so: by putting an unprecedented amount of capital behind a single, focused project. Hundreds of billions to start ³, all put towards only two things: hiring the best engineers and researchers; and acquiring the compute to train models at the frontier. At this point, it can probably only be financed on the scale of nation states, but making it a government affair risks dooming the project from the start. To avoid this, the project must be carefully designed in the most non-European way possible.

A for-profit entity should be formed with participating countries putting up large amounts of capital in exchange for equity. There can be absolutely no loans, no guaranteed employment opportunities, or any other political dealings. The only thing they get is a cut of future, potentially enormous, dividends. If successful, this entity will effectively become a sovereign wealth fund and the best way for ordinary people to gain from AI disrupting our economies.

With nations putting up such massive amounts of funding, politicians will want to run the show, likely through numerous committees and appointments with thinly spread accountability. This will not work. A single person must be chosen to lead the project and given full freedom to hire and manage it as they see fit. This person cannot be a career politician or a tenured executive; they must be an engineer or a scientist, just like the rest of the organisation they’ll build.

What politicians can and should do is wield their power to clear the path for the project to succeed. That might mean giving up land for data centers; expediting reviews and approvals; prioritising grid connections; and perhaps even forcing key companies like ASML to reprioritise their backlog ⁴.

This project would even have a clear advantage compared to the other frontier labs. It could avoid the distractions of inference, pricing, and commercial viability, and focus on the one thing that will matter long term: training the most intelligent models.