The European Commission on Thursday opened the call for tenders on its AI Gigafactories programme: up to seven large-scale AI compute sites across the Union, each required to house at least 100,000 advanced AI chips. Brussels is putting up to €10 billion of EU and national money behind it and expects that to pull in at least €20 billion more from private investors — a headline programme value above €30 billion, roughly $34 billion.
The mechanics are unusually concrete for an EU tech initiative. Bids are split into two lots: the smaller one offers up to €100 million in early-stage funding rising to €400 million per project in phase two; the larger offers up to €200 million initially and as much as €800 million later. A site can sit in one country or be distributed across several. The Commission has signed letters of intent with AMD, Nvidia and Qualcomm so winning consortia can actually get hardware. Applications close 12 November 2026, EuroHPC evaluates, award decisions are expected in early 2027, construction starts that year, and selected sites must be operational within a maximum of 18 months from contract signature.
Now the comparison Europe would rather you skip. Microsoft alone reported $41 billion of capital expenditures and finance leases in the June quarter — more than the entire seven-site European programme is projected to raise from public and private sources combined. The five largest US cloud and AI providers have committed to roughly $660–690 billion of capex in calendar 2026. Europe's answer is a tender that closes in November.
Our take: The money is not the real problem — the clock is. Bids close in November, awards come in early 2027, and the last sites switch on as late as mid-2028. That is a two-year gap against competitors who are currently installing capacity in quarterly increments: Microsoft brought 88 data centers online in a single fiscal year and guided quarterly capex above $50 billion. The sovereign compute project that actually shipped — Japan's national AI factory, 27,500 Nvidia GPUs inside — got built because a vendor and an industrial partner were already holding the shovel before the paperwork started. Europe has letters of intent from three chipmakers, which is not nothing, and a procurement calendar, which is the part that historically kills these things.
What the €10 billion is actually buying
Not a frontier lab. The stated users are startups, scale-ups, SMEs, industry, universities and public authorities — training and fine-tuning infrastructure for companies that cannot rent capacity at hyperscaler scale or do not want their models sitting on American cloud. Judged as an access programme rather than a race, 700,000 chips spread across seven sites is a meaningful public utility. Judged as a bid to catch the US and China on frontier training, it is a rounding error with a ribbon-cutting.
The bottleneck is also unlikely to be silicon. It is power, land and construction labour — the same constraint that has pushed grid contractors to record backlogs in the US, and that Europe carries in a harsher form: higher industrial electricity prices, slower permitting, and grid queues measured in years. A consortium can win in early 2027 and still be waiting on an interconnection agreement in 2029.
What to watch
- Who bids. Whether the consortia are European industrial champions or US hyperscalers wearing a local partner. The second outcome delivers compute faster and sovereignty barely at all.
- The €20 billion. Private capital is assumed, not committed. If the leverage ratio comes in below 2:1, the programme shrinks quietly rather than publicly.
- Chip generation at delivery. "At least 100,000 advanced AI chips" contracted in 2027 for 2028 installation is two silicon generations from what ships today. The tender language, not the press release, decides what actually lands.
- Grid connection dates. The single most predictive number in the whole programme, and the one nobody put in a headline.
Europe has now made the correct diagnosis and written a real cheque. The open question is whether a procurement process built for infrastructure that lasts 40 years can move fast enough for hardware that is obsolete in three.