EuroHPC AI Factories

What it funds

The programme opens the AI-optimized supercomputers and the support services of the Factories to European companies: the benefit is machine time plus specialist support, and for SMEs and startups building AI it is free of charge, while other industrial companies enter on paid terms. The network counts 19 Factories and 13 Antennas, so each application targets the Factory hosting the right system. No cash changes hands: anyone looking for money to buy software is looking for another measure.

Requirements and what is left out

• Anyone outside industry: the industrial innovation track is reserved for companies, while universities and public bodies enter through the scientific track, with its own calls and deadlines.
• Anyone who needs to buy software or rebuild the ERP: the benefit here is compute and support, not a cash grant. The European channel for that spend remains the EDIH network.
• The principal investigator must belong to a company; SMEs and startups need a valid VAT number and, where applicable, a PIC code, to be registered on the EU portal if missing.
• One allocation per access mode at a time for each investigator: a second project by the same investigator enters only if time remains in the call.
• In all three modes, resources are requested on a single system per application.
• Whoever receives access publishes the results, except the confidential parts, acknowledges EuroHPC JU and the hosting Factory and delivers a final report within three months of the end.

How to apply

1. Pick the access mode that fits the need: Playground to start small, Fast Lane if the code is already HPC-ready, Large Scale for the largest training runs.
2. Fill in the online form on the access platform at any time: there is no window that opens.
3. For Playground and Fast Lane, pass the eligibility check and the technical assessment, with no cut-off: Playground answers within two working days, Fast Lane within four.
4. For Large Scale, submit by one of the two monthly cut-offs: an application arriving later moves to the next round.

Everything runs through the access platform; modes, systems and conditions in full are on the access modes page.

What it builds

With free machine time I take a model from prototype to production without buying GPUs: training or fine-tuning on the company's data, a repeatable training pipeline and code already tested on a network system. I first ask where the ready data and code are, because without them the allocation sits unused and the right application is another one.

This entry is informational and is not advice: I do not handle grant applications, and before committing to a purchase it is worth reading the official source above or talking to whoever files the application.