Infrastructure is the nightmare of anyone betting on AI, but in Europe it risks becoming the gatekeeper of failure. A fresh survey from Onnec – a company long embedded in the invisible motorways of connectivity – paints a picture no one, especially in Brussels, will swallow easily: the digital sovereignty machine is running faster than the data centers meant to host it. And the blow is double: on one side, the US administration tightens access to its most capable models; on the other, the European Commission pushes hard for an autonomous path. Caught in the middle: the missing bits, kilowatts, and square meters.

The Onnec report isn’t just a reminder that data centers don’t sprout from the ground. Building capacity for AI workloads, especially for Large Language Models (LLMs), is a three-move game: electric power, thermal management, and GPU supply chains. On the Old Continent, all three are limping. Ageing power grids and multi-year permitting processes turn every new campus into an ordeal. While traditional data centers could get by with a few megawatts, training nodes for the most talked-about models today suck up hundreds, with liquid cooling no longer a luxury but a necessity.

A short circuit is plain: AI sovereignty is first of all physical control over where and how models run. A regulatory framework or investment fund is not enough. If the material fabric can’t hold, the rest is storytelling. Europe is trying to impose data residency and security rules, but without an ecosystem of high-density data centers designed for on-premise or hybrid inference and fine-tuning, the risk is building legal cathedrals in a kilowatt desert.

Those smiling are the large non-EU cloud providers, who already host the bulk of the continent’s AI workloads. With local capacity scarce, companies and public administrations are pushed to consume services outside European control, undermining the very sovereignty they aim to win. American providers, already ahead with immersion cooling and direct deals with GPU manufacturers, pocket yet another competitive edge. For European AI startups, instead, the message is a red flag: they might develop models, but running them close to sensitive data will become increasingly harder and costlier.

It’s not just about money; it’s about systemic incentives. If European governments don’t unlock extraordinary electrification and administrative simplification plans, “sovereign” AI risks becoming a flag waved on someone else’s data centers. The Total Cost of Ownership for those evaluating on-premise deployment is already shaped by VRAM availability and energy prices, but bottlenecks will make it prohibitive for many. Meanwhile, Asia advances with compute gigafactories supported by unabashed power policies. The game isn’t merely technological: it’s a matter of material sovereignty, and the Onnec report has the merit of dragging this out of meeting rooms and throwing it in the face of those designing the future.