Europe’s attempt to build sovereign AI computing capacity is running into a less glamorous problem than processor availability: operators still need enough paying customers to make the infrastructure commercially viable.

OVHcloud founder Octave Klaba has questioned the economics behind proposed European AI gigafactories, arguing that the public purchasing commitments attached to the programme are too small to underpin investments built around expensive, rapidly depreciating graphics processors.

Klaba estimates that a gigafactory could require between €300 million and €400 million in annual revenue merely to reach a zero margin. He contrasts that with a proposed €200 million of AI-service purchases from the French state and European Union over five years, equivalent to €40 million annually, while an operator could be required to put at least €600 million into GPUs before accounting for buildings, networking, software, energy, and staff.

On that arithmetic, he argues that the public order would account for only around 10% to 15% of the revenue required, rather than economically de-risking the project to the extent implied by the headline support. “The numbers don’t add up,” Klaba wrote in his assessment of the programme.

The intervention comes while the European High Performance Computing Joint Undertaking is seeking consortia to establish up to seven AI gigafactories in the EU. The facilities are intended to provide large-scale infrastructure for model training, fine-tuning, and inference, with European and national public funding expected to operate partly as an anchor customer capable of drawing in private capital.

EuroHPC expects the wider initiative to mobilise more than €20 billion of private investment. Bids are due by 12 November, with successful projects expected to be selected in early 2027 and begin operations within 18 months.

Compute sovereignty still needs customers

European AI policy has concentrated heavily on the availability of computing capacity because access to advanced accelerators has become an important constraint on model development. Yet capacity and viable utilisation are different problems, and Klaba’s criticism moves the debate from the amount of hardware Europe can install to the amount of sustained demand operators can secure once it is running.

AI infrastructure is particularly exposed to that distinction because its most expensive components can lose economic value quickly as newer processors arrive. A facility that takes several years to permit, finance, build, and fill cannot rely simply on the assumption that scarce GPUs will remain scarce indefinitely or command the same price throughout their useful life.

Klaba also objects to the national and single-site logic he sees in the French implementation. OVHcloud operates data centres across several European countries and argues that the relevant customer market is European rather than domestic, while inference workloads in particular can be distributed between locations instead of being concentrated in one giant facility.

That disagreement goes beyond architecture. Concentrating infrastructure can simplify some very large training workloads, but it also concentrates power requirements, construction risk, grid connections, planning exposure, and the commercial requirement to keep an enormous amount of hardware occupied. Distributed capacity creates different networking and orchestration problems, although it can bring inference resources closer to users and spread operational risk.

The EuroHPC call already permits some cross-border developments, meaning the eventual programme need not fit an entirely national model. Its stated objective is to crowd private capital into European computing infrastructure by combining public financing with contracted access to AI capacity, rather than expecting governments to own the entire stack.

The utilisation problem follows the building boom

Public policy around AI compute has often been framed as a race to secure processors, power, and data-centre sites, although operators ultimately make money by selling usable computing capacity. That requires a customer base capable of sustaining expensive infrastructure after the initial procurement contracts expire.

Europe has several potential sources of demand, including model developers, industrial AI projects, regulated businesses, researchers, government agencies, and software companies. Data-residency and sovereignty requirements may also favour infrastructure operated within European jurisdictions, but the unresolved question is whether those customers will buy enough premium European capacity, at workable prices, to support several facilities of gigafactory scale.

Even where governments can establish an initial floor under utilisation, sovereignty is not a complete product in itself. Customers will compare accelerator availability, software tooling, networking performance, model services, migration costs, and price against the established hyperscale cloud market, while public procurement cannot permanently substitute for competitive demand.

OVHcloud has its own commercial interest in how the scheme is structured. Klaba says the company’s AI strategy already spans GPUs, orchestration, token-based model access, platform services, and applications, while around 250MW of capacity in its existing data centres is available for GPU deployment. A programme favouring giant national facilities could therefore fit its operating model less comfortably than one built around distributed European capacity.

That commercial interest does not invalidate the criticism, but it makes the debate more concrete. Europe’s AI infrastructure programme is moving from declarations about strategic autonomy into decisions about revenue, utilisation, depreciation, and customer acquisition, and those numbers will determine whether sovereign compute becomes a durable market or a collection of heavily subsidised machines.