Despite the European Union pouring billions into state-backed supercomputers through the EuroHPC Joint Undertaking (EuroHPC JU), the continent’s most heavily funded AI startups are abandoning domestic hardware. A review of cloud partnership data reveals that over 80% of Europe’s top-funded AI startups rely entirely on American commercial hyperscalers to train and host their models.
This paradox exposes a critical gap in Europe’s pursuit of technological sovereignty: public machines offer raw silicon power, but lack the proprietary orchestration software and rapid elasticity that venture-backed AI companies demand.
The Illusion of Compute Sovereignty
Top-tier European AI companies overwhelmingly depend on U.S. infrastructure rather than domestic supercomputing projects like the upcoming €387.8 million LUMI-AI system in Finland. High-profile national champions, including France’s Mistral AI, rely on Microsoft Azure and Amazon Web Services (AWS) for their core enterprise deployments.
While these models strictly adhere to local regulatory mandates—including EU AI Act compliance requirements—their core training compute remains tethered to American infrastructure. The hardware may physically reside in European data centers to satisfy data residency laws, but the underlying software stack is firmly owned and operated by U.S. corporations. As a result, European venture capital meant for infrastructure simply flows back to Silicon Valley.
The Software Friction of Bare Metal
Securing access to thousands of graphics processing units (GPUs) is only the first step in commercial AI development. State-funded supercomputers typically provide “bare metal” computing power. However, training frontier models requires advanced enterprise orchestration tools, proprietary InfiniBand networking optimizations, and robust containerized environments.
Machine learning engineers repeatedly cite a lack of commercial-grade orchestration as the primary bottleneck. EuroHPC systems traditionally utilize academic job schedulers, such as the Slurm Workload Manager, which are purpose-built for static scientific batches. Venture-backed startups operating on aggressive timelines require dynamic Kubernetes orchestration and modern developer toolchains. Converting a rigid, bare-metal academic cluster into a fluid AI training environment demands extensive engineering hours that fast-moving startups cannot spare.
Allocation Timelines: Months vs. Minutes
In the commercial AI sector, development cycles are measured in weeks. Conversely, accessing state-funded compute can take months. EuroHPC’s standard allocation guidelines enforce rigid application windows and academic peer-review processes that severely delay hardware provisioning.
Commercial cloud providers offer instant scalability. A startup can provision thousands of GPUs, execute a training run, and spin the cluster down within days via simple APIs. If a training run crashes on a public supercomputer, engineers are often forced back to the bottom of a lengthy allocation queue. On commercial platforms, jobs are restarted instantly.
A Pivot Toward AI Factories
Currently, the EuroHPC network effectively serves its original target demographic: universities, climate modelers, and state research laboratories. While these systems execute their mandate for public scientific research, they are structurally incompatible with high-velocity commercial AI.
The European Commission recently acknowledged this structural deficit. To bridge the gap, officials launched the AI Factories initiative, a legislative amendment to EuroHPC regulations aimed at deploying dedicated supercomputing infrastructure with fast-tracked, commercial-grade access for startups. Until these factories materialize and match the software fluidity of AWS or Azure, true infrastructure sovereignty will remain out of reach.
