European AI Talent: Stay, Drift, or Come Back - The Compute Argument That Changes Everything

The conventional policy response has been to treat this as a mobility problem. If researchers leave because US visas are easier or because American campuses offer better salaries, then the fix is to make European residency more attractive. France has invested heavily in its French Tech Visa programme; Germany accelerated its Skilled Immigration Act reforms. Both are genuine improvements. Neither has meaningfully reversed the trend. The problem was never primarily bureaucratic.

Ask any ML researcher who has made the transatlantic move why they did it and the answer, underneath the polite references to opportunity and culture, is almost always the same: compute. A researcher at a US hyperscaler or a well-capitalised US lab can run a training job with thousands of H100s over a weekend. A researcher at a European university, even a well-funded one, waits weeks for allocation on a national HPC cluster configured primarily for climate science or molecular dynamics, not transformer pre-training. The workflows are incompatible. The ambitions cannot be matched.

This is not a complaint about European science. It is a complaint about European infrastructure policy, and the distinction matters enormously when deciding where to direct public money.

What Inria and ETH Alumni Actually Say

Alumni surveys from major European AI research centres reinforce the compute thesis with uncomfortable specificity. Surveys of Inria and ETH Zurich alumni who relocated to the United States consistently place research infrastructure, meaning GPU access and cloud compute budgets, above salary as a stated reason for the move, particularly among researchers who left after 2020, when scale became the dominant paradigm in deep learning. The message is not that European institutions pay badly, though some do, but that they cannot offer the experimental surface area that large-scale compute provides. A researcher who wants to study emergent behaviour in large language models cannot do serious work on a 16-GPU cluster. They need thousands of accelerators, sustained over weeks, with fast interconnects and the engineering support to use them well.

Hugging Face, the AI company headquartered in Paris, is often held up as a counter-example: a European-founded organisation that retained European talent and built a globally competitive research culture. That framing is partially correct and largely misleading. Hugging Face succeeded in part because it built infrastructure access into its business model, partnering with US cloud providers to give its researchers the compute they needed. It is a Paris-headquartered company whose compute runs overwhelmingly on American cloud infrastructure. That is not a criticism of Hugging Face; it is a structural observation about where the leverage actually sits.

The Technical Case for GPU Collectives

The solution that is beginning to gain serious traction in European policy circles is not another visa scheme or another fellowship programme. It is the European GPU collective: a federated, interoperable cluster of AI-grade accelerators pooled across member states, publicly accessible to researchers, and large enough to be genuinely competitive with what US labs can offer their staff.

The technical argument for this model is strong. Training large models is embarrassingly parallel in the sense that it benefits enormously from scale, but it also requires tight coupling: fast interconnects, low-latency storage, and workload schedulers that understand the specific demands of distributed deep learning. A GPU collective that aggregates hardware from national research networks across Germany, France, the Netherlands, and Nordic member states, interconnected via a purpose-built high-bandwidth fabric, could plausibly rival the on-premise clusters that mid-tier US labs maintain for their researchers. It would not match a hyperscaler's peak capacity. It does not need to. It needs to be good enough that a researcher at ETH Zurich or at the University of Amsterdam does not feel they are doing their best work with one hand tied behind their back.

The EuroHPC Joint Undertaking is the existing institutional vehicle for this kind of coordination, and it deserves credit for having moved faster than most European infrastructure programmes. Its pre-exascale systems at sites including the Barcelona Supercomputing Centre are real assets. But EuroHPC was designed around scientific HPC workloads, and its allocation processes, governance structures, and software stacks are not optimised for the iterative, experiment-heavy workflows of ML research. Adapting EuroHPC into a genuine AI compute collective, rather than bolting AI on as an afterthought, is a political and engineering task that the Commission has not yet fully committed to.

The Political Case

The political case is, if anything, more urgent than the technical one. European governments are spending billions on AI strategy documents, national AI plans, and competitiveness initiatives. Commissioner Teresa Ribera and her colleagues on the digital brief have made AI infrastructure a stated priority. The AI Act has consumed enormous political energy. Yet the concrete compute commitments that would actually change researcher behaviour remain modest relative to the stated ambition.

France's national AI strategy, updated under the influence of work by researchers associated with Inria and the broader French AI research community, does include compute commitments; the Jean Zay supercomputer at IDRIS has been upgraded with significant GPU capacity. Germany's AI strategy references compute access. But these are national efforts that remain fragmented. A researcher in Lisbon cannot easily access allocations on Jean Zay. A team at the University of Warsaw cannot spin up a large training run on a German national cluster without navigating bilateral agreements that were not designed for the speed of modern ML research.

The solution is not simply more money. It is a different governance model: one that treats European AI compute as a shared public good, allocates it through a fast, peer-reviewed process analogous to the best elements of how the US NSF ACCESS programme works, and explicitly measures success by researcher retention, not just by teraflops provisioned.

The scale of the challenge and the opportunity becomes clearer when you put concrete figures alongside the narrative. Europe's compute gap is measurable, the talent flows are quantified, and the policy response, though growing, still falls well short of what the AI Index 2026 talent section describes as the investment threshold needed to shift researcher behaviour at population scale.

What a Real Commitment Looks Like

A credible European GPU collective would need several things that current proposals lack. First, it would need dedicated AI-grade interconnect, not shared capacity on networks designed for MPI-based simulation workloads. Second, it would need allocation processes measured in days, not weeks: a researcher with a promising hypothesis should be able to test it at scale within a week of submitting a brief proposal. Third, it would need a software layer that feels like a modern ML platform, with support for popular frameworks, container environments, and the kind of monitoring tooling that researchers at well-resourced US labs take for granted. Fourth, and most importantly politically, it would need a governance commitment that ring-fences a meaningful share of capacity for academic and independent research, rather than allowing industrial users to crowd out the university researchers whose retention is the point of the exercise.

None of this is technically difficult. The difficulty is political will and budget prioritisation. The EU's AI Factories initiative, announced as part of the broader industrial AI push, is a step in the right direction; it explicitly targets AI compute for research and innovation communities. Whether the AI Factories programme delivers at the scale and speed required is the question that will define whether the next generation of ETH and Inria PhD graduates decide that staying in Europe is compatible with doing world-class work.

The talent is here. The research culture is serious. The universities are, in many cases, excellent. What Europe has consistently failed to provide is the material infrastructure that turns excellent training into excellent research careers that stay on the continent. Visa reform tidies the edges. Compute access changes the equation.

European policymakers who are serious about AI competitiveness need to stop treating compute as a technical detail to be handled by infrastructure ministries and start treating it as the central variable in the talent question. The researchers who left know exactly why they left. The researchers who are currently deciding whether to leave know exactly what would make them stay. The answer is not in the immigration code. It is in the data centre.