AI Gigafactories: Details of the Call for Proposals

Make sure to maintain, over time, at least 90% of the performance measured at the installation.

The European Commission is requesting this from the consortia responding to the call for proposals for “AI gigafactories” (AI gigafactories).

Proposals may be submitted until 12 November 2026. There are two lots, for medium-scale and large-scale AI gigafactories. The medium-scale facilities must eventually reach at least 75,000 accelerators in H100-equivalent terms. The large-scale facilities must reach at least 100,000.

In both cases, there will be an initial deployment phase followed by an extension phase.

The first phase, intended to last a maximum of 18 months, must enable the deployment of at least 25,000 accelerators in the medium-scale gigafactories and 40,000 in the large-scale ones. Brussels envisions a five-year horizon to complete the second phase.

EuroHPC and 18 member states – including France – will jointly acquire computing capacities, with specific contracts for each phase.

Two scales of gigafactories… and funding

For Lot 1 (medium-scale gigafactories), EuroHPC will select up to 4 projects. In Phase 1, each will receive up to €100 million in European funding. Three levers will be mobilized:

• Interconnection Mechanism in Europe
• Programme for a Digital Europe
• Horizon Europe Framework Programme for research and innovation

The member states must provide financing at least equivalent.

The same parity logic applies to Phase 2, with the maximum envelope per project nevertheless raised to €400 million… subject to the availability of funds through the EU’s next multiannual financial framework (2028-2034).

For Lot 2 (large-scale gigafactories), EuroHPC will select up to 3 projects. The EU will provide up to €200 million per project in Phase 1; €800 million in Phase 2. In all cases, its share will be limited to 17% of the IT capex of the projects.

France commits to the medium scale

For medium-scale gigafactories, the initial requirement is a 50-50 mix between IaaS/PaaS services and MaaS (allocation adjustable, up to a 10-point swing). Ultimately, a computing power of at least 120 MW is expected.

Eight countries have positioned themselves to host, with France among them. France commits to investing up to €100 million (with a €10 million top-up from Ireland).

Finland is also in the running. It has the support of Estonia, Latvia and Sweden. All three are applying to host not a gigafactory itself, but a satellite site to the one that would be established in Finland.

The EU allows funding for these satellite sites, which must fulfill an essential function such as backup, load balancing or the provision of specialized nodes.

Member states will need to have finalized their investment commitments no later than six weeks before the call closes. Here is a summary of their current commitments:

Five countries lined up for large-scale gigafactories

For large-scale gigafactories, the initial requirement is a mix of 60% IaaS/PaaS and 40% MaaS. The expected computing power in the long term is at least 150 MW.

Five countries have positioned themselves to host a gigafactory, all with €200 million investment in Phase 1:

• Germany (€800 M earmarked for Phase 2)
• Spain (promises to invest at least €50 M in Phase 2)
• Greece (€150 M earmarked for Phase 2)
• Italy (€200 M earmarked for Phase 2)
• Portugal (in the running for possibly an annex site)

All projects combined, the theoretical total envelope amounts to €10 billion in public funding. The European Commission estimates that the initiative should mobilize at least twice that amount in private funding.

Europe sets its SLAs

The application form lists a number of expectations, such as integration with sectoral data spaces and the establishment of a roadmap to implement a European software stack (cloud + AI) within four years after the initial deployment. Applicants are also invited to plan for European hardware to be used at least starting from Phase 2.

Beyond the requirement to maintain at least 90% of initial performance, Brussels asks them for:

• 99.5% availability for training and 99.9% for inference (excluding planned maintenance)
• A Recovery Time Objective (RTO) of 4 hours and a Recovery Point Objective (RPO) of one hour
• A target 100% availability for external connectivity, with no more than 3 days of maintenance per year
• In inference, a P95 latency ≤ 300 ms for standard workloads

For standard support, hours will be Monday to Friday, 8:00–18:00 local time, with a response within 30 minutes. For critical services, 24/7 with a response within 10 minutes. In case of a critical incident:

Notice must be given at least 7 days for planned maintenance and 4 days for ad hoc maintenance.

A two-stage evaluation

EuroHPC will entrust the evaluation of bids to a panel of five to seven independent experts. Contracts will be awarded in order of ranking, until the EU and member states’ combined budget is exhausted. Proposals not selected may go onto a waiting list for up to two years and, in the meantime, may receive funding if the EU unlocks funds.

The evaluation will be conducted in two parts. The first focuses on the technical quality of the proposals (40% of the score), their strategic impact (20%), and their financial viability (40%).

The mathematics of the scale criterion

The score obtained in the first part will account for 50% of the total score. The remaining 50% will be split equally between “Scale and service quality” and “Cost-benefit ratio.”

The scale criterion depends on the volume of accelerators the project plans to reach.

An equivalence table between accelerators

The service quality scoring is calculated in equal parts between the IaaS/PaaS and MaaS layers.

The cost-benefit evaluation is based on calculating a normalized price per accelerator for each service layer.

The H100 equivalence follows these indices: