Public Money, Private Machines: Europe's €30bn Compute Wager

Analysis as of 13 August 2026.

On 30 July 2026 the EuroHPC Joint Undertaking opened an open procedure, reference EUROHPC/2026/OP/0008, to select consortia for as many as seven AI Gigafactories across Europe. Brussels attached a striking number to the launch: more than €30 billion of investment expected to be mobilised. Yet that figure is an extrapolation rather than a budget line. It multiplies an indicative unit cost of €4–5 billion per facility by the number of projects the call permits. What has actually been committed is narrower, later and more conditional than the announcement suggests, and the distance between the two is where the hard questions about Europe's computational sovereignty begin.

The arithmetic is deliberately capped. The Union will cover at most 17 per cent of a supported project's eligible computing-infrastructure capital cost, with participating states matched under an identical ceiling — a combined public share of roughly 34 per cent. Of the €5 billion Union maximum, only €1 billion sits inside the current financial framework: €100 million for each of four smaller projects and €200 million apiece for three larger ones. The remaining €4 billion cannot be awarded before 2028 and depends on a budget, a successor entity and appropriations that do not yet exist. Buildings, utilities and every euro of operating cost fall on private sponsors.

National announcements add texture but not consolidation. France has promised to order €100 million of computing capacity for its public administrations; Poland has adopted a cabinet resolution backing purchases of the same value between 2028 and 2033, while expecting private capital to carry a project it prices above PLN 10 billion. Spain went further, disclosing a €720 million public equity stake alongside a separate €300 million contribution for services. These are three different instruments — offtake, shareholding, contribution — and treating them as one pot double-counts money and misreads risk. Eighteen states signed the joint procurement, but the sums behind each signature remain unpublished.

Ownership disclosures are thinner still. The AETHER grouping presented twelve industrial members in July, yet a regulated update of 5 August — issued after an exchange with the French market authority — revealed that the project vehicle has a single shareholder: the holding company of the chairman of one member firm. In Spain, the advertised private majority amounts to 51.01 per cent against 48.99 per cent in public hands, a margin of scarcely one percentage point resting partly on a 4 per cent stake. Consortium membership, it turns out, is not ownership, and announced sites are not energised megawatts.

The deeper issue is what public money actually purchases. The model buys guaranteed access time on privately run machines, not the machines themselves. Anchor customers enjoy a carve-out from sovereignty and data-residency conditions for capacity they resell; classified workloads sit outside scope; the only crisis mechanism is a best-efforts undertaking. Grid connections in parts of Europe queue for two to ten years, while the contract allows eighteen months from signature to readiness. Bids close on 12 November 2026, with awards expected in early 2027. Before then, several questions deserve to be asked out loud:

— Whose entitlement prevails when compute becomes scarce — the lender's, the commercial anchor customer's, or the public purchaser's?

— Can a state genuinely rely on capacity it neither owns nor administers if the operator slides into insolvency?

— Will the executed hosting agreements convert nominal accelerator counts into guaranteed, schedulable hours with enforceable remedies?

— Does a one-point private majority, or a single-shareholder vehicle behind a twelve-member coalition, satisfy the ownership-control scrutiny the tender itself demands?

— And by what yardstick should the outcome be judged — is the six-part sovereignty grid the analysis applies, from jurisdictional through financial control, and its falsification test set on the first specific contract, the right measure of whether Europe is buying control or merely renting it?