Summary: Council Regulation (EU) 2026/150, adopted on January 16, 2026, amends the EuroHPC Joint Undertaking so AI gigafactories can be built across the EU and takes effect on January 20. It adds a quantum pillar and uses procurement in which the EU contribution is capped at 17% of each facility’s capital expenditure. A work program is due within six months, with an initial call planned for Q1 2026.
Council Regulation (EU) 2026/150 is now on the books: it was adopted by the Council of the EU on January 16, 2026, rewriting the mandate of the European High Performance Computing Joint Undertaking (EuroHPC JU) so that AI gigafactories can be built across the Union. What the amendment reworks is Regulation (EU) 2021/1173, the 2021 act that created the Joint Undertaking, and it took effect on January 20.
Within the Union’s compute-sovereignty strategy, this is the step that carries the most weight. Training at the frontier costs more every cycle, and training-grade compute sits in the hands of a few US providers; the EU’s answer is to legislate, commit public money, and buy jointly, so the capacity to train next-generation large models stays inside Europe.
Defining the AI Gigafactory
Procurement and operation of the EU’s supercomputers is EuroHPC JU’s existing remit, and a 2024 amendment added “AI Factories” to it. What this round changes is the scale of the ambition, pushing it to hyperscale: in the European Commission’s definition, an AI gigafactory is a very large supercomputing facility optimized for AI, built to back training and deployment work on the most advanced large AI models. The regulation text published on EUR-Lex spells the target out even more plainly — these facilities are to host development work on models with trillions of parameters.
Another change rides along in the same amendment: EuroHPC JU gains a dedicated quantum computing pillar. HPC, AI, and quantum now live under one joint undertaking, and the Commission casts this as strengthening Europe’s “strategic autonomy and global competitiveness” across high-performance computing, AI, and quantum technologies.
Six Months from Proposal to Law
The legislative route began with a Commission proposal in July 2025 and ran through the consultation procedure: Parliament voiced its opinion first, and the Council made the final call. Adoption came on January 16, 2026, and the rules took effect on January 20 — a run of roughly six months, quick work by the standards of EU lawmaking.
The regulation also carries deadlines of its own: within six months of entry into force, EuroHPC JU’s governing board has to adopt a work programme. And per Kilian Gross, head of AI policy at the Commission’s DG CNECT, the official call to establish AI gigafactories is slated for the first quarter of 2026. Before long, siting, operator selection, and the procurement structure should move from principle to concrete tenders.
How the Money Works: InvestAI and PPI
These facilities will not be financed through classic grants. Procurement runs through Public Procurement of Innovative Solutions (PPI): winning consortia pay to build and operate the sites, the EU’s own contribution is capped at 17 percent of each facility’s capital expenditure, and consortia backed by member states carry whatever remains. The setup lets a limited EU budget pull in far larger private and national funding, and delivery risk lands on the operators.
Everything sits within the Commission’s InvestAI framework: out of the €200 billion overall AI investment push, €20 billion is set aside for AI gigafactories. Measured against single US campuses running to tens of billions of dollars, the sum is no moonshot. The aim lies elsewhere — public training compute becomes a standing institution instead of a one-off stimulus.
What Developers — and Europe’s AI Ecosystem — Should Take From It
Three things stand out. One, the pace beats expectations: a first call in Q1 and a work programme owed within six months should leave 2026 with concrete facility picks and the consortia set to run them, instead of yet another white paper. Two, startups and research organizations across Europe gain a fresh lane to training-grade compute — no longer tied wholly to US cloud vendors’ quotas and pricing — and model training gains a sovereign option in how it is scheduled and budgeted. Three, quantum’s placement inside the same undertaking signals that the EU reads post-Moore computing as a single integrated layout rather than two systems running in parallel.
Model cadence keeps accelerating, and open-source strategies and terms keep pulling apart as 2026 opens (see our 2026 opening outlook). The EU is not responding by chasing a particular model; it is putting the infrastructure underneath in order. The payoff on this route is slow — but once built, it changes what training costs for an entire generation of models.