ada könig
← all pieces
Urtext · 2026.10.04

What the EU Is Actually Buying With Its AI Gigafactories

The EU call for seven AI "gigafactories" closes on 12 November. Of the €30 billion announced, the public money committed today is one billion. The chips come from three US companies, and Berlin has put in writing that it will not own its own.

On 30 July EuroHPC, the joint undertaking through which the European Union buys and runs its supercomputers, opened the call for up to seven "AI gigafactories" in at least seven member states. Each is meant to house around a hundred thousand advanced processors and cost between four and five billion euros. The press release speaks of "more than €30 billion" in investment. Bids close on 12 November, selection is expected in early 2027, and operations should start within eighteen months of that. Henna Virkkunen, the Commission's executive vice-president for tech sovereignty, calls it "a strategic necessity."

The thirty billion needs taking apart. Ten is public money, split between the Union and the member states; twenty is private capital that is expected but not yet signed. Of the five billion that falls to the Commission, one is committed today, from the current budget. The other four depend on the next multiannual financial framework, the EU's seven-year budget, which is still under negotiation. A Commission official put it frankly: "We cannot pre-empt the decisions about the next MFF." The public share also has a ceiling: public buyers can commit to at most 17% of the infrastructure cost.

What that 17% buys is machine time. The public sector comes in as an "anchor customer": it commits in advance to buying computing hours and in return gets guaranteed access in proportion to what it put in. The rules on control exist, and they are serious. The coordinator must be established and effectively controlled in the Union, the facility must be located here, and core operations cannot be subcontracted outside it. The Commission sums it up as "sovereign compute".

Inside, though, the chips will come from three US companies, Nvidia, AMD and Qualcomm, all of which have signed letters of intent with the Commission. Eighteen MEPs have written that this leaves the dependency untouched. This is where the European call crosses paths with Washington. The government that licenses the sale of those chips has already asked Nvidia for 15% of its China revenue in exchange for permission, and it is now weighing stakes in the AI labs. The same processor, in Munich or in Madrid, stays subject to a permission decided somewhere else.

Zuzanna Warso of Open Future has pointed to the other knot. Filling facilities on this scale takes enormous demand. The customers who have it are mostly the big US cloud providers, the very dependency the sovereignty agenda is meant to reduce. As tenants, she notes, they may keep their operational autonomy and stay outside the sovereignty conditions that bind the facility. The gigafactory would be European by address, operator and governing law, while the work running inside it could answer to someone else.

Here in Berlin the story has taken an instructive turn. The federal government set aside €805 million, raised to one billion on 7 August. In April, answering a parliamentary question from Die Linke, it stated that it will take part neither as owner nor as consortium partner. The national consortium of Telekom, Ionos, SAP, Siemens and Schwarz, the group that owns Lidl, broke up over who should lead and where to build.

There is a logic to Berlin's choice, and it is easier to appreciate with Washington in view. A state that holds no shares never finds itself regulating a company whose dividends it collects; it is the old problem of the referee who owns a team. Still, the call leaves one question open: if the state is not the owner and the chips answer to another capital, where does public command sit? It can only sit in the access conditions, which the call states without numbers. Start-ups, researchers and public administrations are named, but no share of capacity is reserved for them.

Then there is timing. A facility selected in early 2027 switches on no earlier than mid-2028. Meanwhile computing is shifting from training to inference, meaning the use of models already trained, which calls for different architectures. And Mistral, which ought to be the natural customer, is building its own data centres in the meantime.

Until 12 November, three things are worth watching:

who signs as anchor customer alongside the public bodies;

whether the access quotas for research and start-ups are given a number;

whether the next multiannual budget confirms the four billion that today is an estimate.

Those three answers will tell us what Europe is buying: a building, machine time, or the right to decide what it is for.