The AI Factory Problem

The EU is spending roughly €10 billion to build 19 AI factories across 16 member states by 2027. The strategic instinct behind this is sound. AI capability is now geopolitical weight, and a continent that relies entirely on American and Chinese models for its critical infrastructure has already lost something important. The architecture chosen to act on that instinct, however, is wrong in ways that will take years to fully reckon with.

Participation trophies at scale

Nineteen factories distributed across sixteen member states is a political solution wearing the costume of an industrial one. The governing constraint on this distribution was obvious and predictable: every member state with a vote and a commissioner has a claim on participation, and the architecture had to satisfy that claim. The result is clusters sized to national political requirements, not competitive ones.

Frontier AI models are trained on clusters with tens of thousands of GPUs running continuously for months. The facilities Europe is building are research-grade. They will do real work - academic computing, industry experimentation, model fine-tuning. None of them will produce what OpenAI, Google DeepMind, or Anthropic produce. Distributed across 16 countries, the aggregate compute still falls short of what a single serious frontier lab deploys. You can add up nineteen medium facilities and still not get one large one, because the bottleneck in training large models is not the number of facilities but the communication latency between processors. Geography defeats arithmetic here.

Compare the competitive context. US frontier AI was built through massive private investments in single coherent clusters - decisions made in weeks, deployed in months, iterated on continuously. There are no procurement rules requiring geographic distribution across states with competing political interests. In China, state direction concentrates resources where competition requires them, without the overhead of justifying the decision to twenty-six other governments. Both models are built around a simple principle: if you want to win a race, you concentrate resources on winning it.

The governance mismatch

The deeper problem is who is designing these facilities and under what mandate.

European AI infrastructure is being built by people who are excellent at regulation, procurement, institutional accountability, and public tendering. These are real skills. Europe has them in abundance and genuinely needs them. They are the wrong skills for building competitive AI infrastructure.

Training runs fail at 3am. When they do, the people who fix them are not reading procurement guidelines - they are debugging CUDA errors and hardware failures in real time, under pressure, with months of compute investment at stake. Architecture decisions at the frontier require people who understand the tradeoffs between transformer variants, memory bandwidth constraints, and interconnect topology. Partnership networks that move faster than research breakthroughs cannot be governed through public tendering cycles that take eighteen months from specification to contract award.

This is the core mismatch. The EU has designed its AI infrastructure program as though the primary challenge is accountability. The primary challenge is competition. Accountability matters - for public money, for safety, for governance of powerful systems. But it has to be layered onto a structure built to compete, rather than competing structures having to force themselves through accountability pipelines designed for bridges and highway contracts. Right now, the accountability requirements come first and shape everything else. That order produces facilities that can be audited but not the outcomes that justify building them.

The window is closing

There is a time dimension here that the program’s architecture ignores.

The gap between European compute capacity and the compute required to train frontier models is already significant. In absolute terms: the largest European supercomputing resources are measured in petaflops. Frontier training runs consume exaflops, sometimes sustained over months. This is not a gap that closes gradually through incremental investment in 19 distributed facilities. It closes through concentrated capital deployment of a different order of magnitude.

The factories being built now will set the trajectory for the next decade. Infrastructure of this kind shapes what research can be done, what talent it attracts, what industry clusters form around it. Build research-grade distributed infrastructure and you get a research-grade distributed ecosystem. Build frontier-scale concentrated infrastructure and you get a chance at something that matters strategically.

In three years, if these factories are completed as designed, Europe will have 19 facilities that are genuinely useful for secondary applications and completely irrelevant to frontier AI development. At that point the gap becomes structural - a physical fact about European AI capacity that no policy document can paper over.

What would actually work

Fewer, larger clusters. Two or three facilities at genuinely European scale - sized for frontier training ambitions, located where power infrastructure, cooling, and technical talent already exist, and governed by research operations structures that are separated from regulatory oversight.

Governance that distinguishes between building and auditing. The people running training operations should have operational authority over operational decisions. Regulatory bodies should audit outcomes and enforce safety standards. These should be different institutions with different mandates, staffed by people with different backgrounds. Right now the instinct is to fold both functions into the same bureaucratic structure, which produces an institution that is good at neither.

Procurement designed for AI’s actual pace. Competitive AI research does not operate on 18-month tendering cycles. The technical requirements change faster than European procurement law permits. This is solvable - there are defense procurement models that handle rapid iteration in high-stakes technical programs. The EU has simply not applied them here.

The case for European AI capability is not complicated. AI will shape military logistics, intelligence analysis, energy grids, financial infrastructure, and healthcare systems. Outsourcing all of that to foreign models - even allied ones - is strategic dependency at exactly the moment Europe is learning what strategic dependency costs. Building the capability is the right response. The current program is building the infrastructure that lets Europe say it tried, without building the capacity that would let it succeed.

That difference won’t show up for three years. When it does, everyone will see it at once.