Sovereignty is not a European chatbot

A home-grown European version of Mythos AI? Forget it, write VUB professors Vincent Ginis, Andres Algaba and Brecht Verbeken. That gap cannot be bridged overnight.

Read the full opinion piece in De Standaard (In Dutch).

OpenAI and Anthropic regularly share the topics most frequently raised with their chatbots. Psychological support invariably tops the list. We can only imagine that, over the past weekend, the question most often asked of these chatbots was of a different kind: "How much does it cost to train a European language model like Mythos?" Psychological support probably still ranks second - much needed, no doubt, to process the answer to the first question. The price tag is anything but modest.

On Saturday morning, it became clear that a single order from Washington is enough to cut us off from the most advanced AI. Anthropic took its most powerful model, Fable 5, offline for all customers, following a US government decision to make it available to American citizens only.
That decision will have far-reaching consequences. What is at stake is a broad societal and economic acceleration. Those who deploy these systems first will build a lead across all the domains on which our prosperity depends.

The reaction on social media was predictable: "This is the ultimate wake-up call." Europe must build its own model, independent of the Americans. The impulse is understandable, but however much we might like it, building a model that can compete at the top level is simply impossible.
This is due, first and foremost, to the self-reinforcing improvement of the best models. Those who already have the best models will develop better ones faster. Top systems attract the world's leading talent and, in the meantime, contribute themselves to the development of their successors. This explains why Meta, Microsoft and xAI struggle to keep up with the frontrunners Anthropic and OpenAI.
There is also the geopolitical bottleneck of the scarce chips required to train such models. This is the primary reason why China, despite its vast ambition and resources, has been unable to match the United States. A government that decides only Americans can access the most powerful AI will hardly allow foreign powers to purchase the cutting-edge chips needed to train their own.

So what should we do? Trade full sovereignty for mutual dependence. We need deep transatlantic anchoring around access, security and continuity: guaranteed access, European-based inference, joint safety testing, and emergency protocols that do not abruptly exclude allies. Such an arrangement will only endure if it is also in the Americans' interest not to break it.
This begins with diplomacy. We must strengthen ties with both the labs and the government, be present where decisions are made, and speak with a collective voice so that the US feels that excluding us carries both financial and reputational costs. The American labs themselves have an interest in maintaining access to one of their largest export markets. Turn them into allies.

At the same time, Europe must become more significant within the AI value chain, increasing interdependence. Europe is already present here. The Dutch high-tech company ASML builds the EUV machines used for the most advanced chips; Leuven's Imec possesses key expertise and scarce High-NA equipment (from ASML). These are levers we already have, but they are far from sufficient. The task is to acquire more - in energy, computing power, and across other parts of the AI ecosystem.
Building a European model sounds logical and appealing. Whether it is wise is another matter. This technology cannot simply be replicated, and chip scarcity ties it directly to geopolitical tensions. Turning away from a stronger power one can never catch up with, under the guise of independence, amounts to choosing isolation. No one is an island - and in the coming years, that is precisely what you do not want to be.