Summary
- The European Commission opened a tender on 30 July for up to seven AI gigafactories; seven projects are a ceiling, not a guaranteed outcome.
- Up to €10 billion of EU and national funding is intended to attract at least €20 billion in private investment, which is still an expectation rather than committed capital.
- EuroHPC and 18 Member States will jointly procure compute access time, creating an anchor-customer mechanism for selected operators.
- Two lots and two development phases tie larger funding ceilings to progressively larger processor capacity.
- The Commission has letters of intent with AMD, NVIDIA and Qualcomm, but these are not chip orders or delivery guarantees.
- Awards, construction and operation are forward milestones; sites, electricity, cooling, prices, winners and usable capacity remain undisclosed.
The public buyer is the missing financial layer
Most industrial-policy announcements begin with a capital number. This one is more revealing when read as a revenue mechanism. EuroHPC and the participating states will not merely help pay for buildings. They will jointly procure access time from the machines that are selected.
That distinction matters to financing. A developer contemplating a very large AI facility faces several uncertainties at once: accelerator prices, grid availability, utilisation, model efficiency and the willingness of European customers to pay for frontier-scale compute. A grant reduces construction cost. An anchor order can reduce demand risk by placing a contracted buyer on the other side of part of the capacity.
The tender therefore asks public procurement to perform a job that hyperscalers normally do internally. A large cloud operator builds against its own customer book. A new consortium needs a credible path from installed processors to billable workload. Joint purchasing by 18 states may provide the first block of demand against which lenders and equity investors can model revenue.
It does not remove commercial risk. The Commission has not disclosed the volume, duration or price of compute access that public buyers will purchase from each winner. Nor has it said how much capacity must remain available to private customers. Until the contracts are awarded, “anchor demand” is a mechanism rather than a bankable cash-flow schedule.
Two lots turn scale into a staged obligation
The call divides the prospective facilities into two lots and two phases. Lot 1 can support up to four projects, with an EU contribution of as much as €100 million in phase one and a further €400 million per project in phase two. Lot 2 can support up to three projects, with ceilings of €200 million and then €800 million per project.
Funding is paired with expansion. In Lot 1, phase-two capacity must reach at least three times the number of advanced processors installed in Europe’s strongest current AI Factory. In Lot 2, it must reach at least four times. AP describes the intended end-state as at least 100,000 advanced AI chips per facility, but no winner has ordered, installed or activated that inventory through this tender.
The phased structure is an attempt to preserve options. Public money can support an initial build before the largest scale is required. Performance, delivery and demand can be observed before a consortium receives the full second-phase support. For operators, however, the later capacity target creates a forward obligation: early design choices for land, substations, cooling and network fabric must accommodate a much denser second stage.
The two lots also acknowledge that not every project should carry the same risk. A distributed cross-border design may face different interconnection and governance costs from one large campus. The call permits a single site, several sites in one country or a distributed development across Member States. It does not yet explain how latency, accounting or failure responsibility will be compared across those architectures.
€30 billion is a lever, not a balance-sheet fact
The headline arithmetic combines up to €10 billion of EU and national public funding with at least €20 billion of private investment that the Commission expects to attract. It is a target for mobilisation. It is not €30 billion already committed, awarded or spent.
That boundary is essential because leverage can be announced long before it is financed. Consortia still have to assemble equity, debt, suppliers, sites and power. Participating states must match EU funding under the joint procurement arrangement. Private investors will price construction risk and future utilisation against the terms of the public contracts.
The public share may still be powerful. A one-third contribution paired with an anchor customer can change the risk profile more than a grant of the same size without revenue support. It can also crowd in specialist infrastructure funds that would not underwrite frontier-AI demand on technology forecasts alone.
But the state can misprice demand as easily as the market can. If public buyers reserve capacity that applications do not use, taxpayers carry idle-compute risk. If procurement prices are too low, operators may struggle to fund energy and hardware refresh. If they are too high, the programme may protect uneconomic capacity. The decisive documents will be the access-price, availability and performance clauses, not the aggregate funding slogan.
Sovereignty still arrives in an American box
The Commission says hardware may be bought from Europe or like-minded countries. It also says it has signed letters of intent with AMD, NVIDIA and Qualcomm following the EU-US trade deal to facilitate access to necessary hardware. Those letters reveal the immediate contradiction in the strategy.
Europe is seeking greater control over where advanced models are trained and served, who can access the systems and which legal rules govern the data. Yet the accelerators and much of the software stack are likely to remain dependent on non-European suppliers. A European building running American processors can improve jurisdictional control and operational continuity without creating semiconductor independence.
That is not a reason to dismiss the project. Sovereignty is layered. Ownership of land, grid connections, facilities, orchestration, customer access and data governance can matter even when chips are imported. The problem arises when the political label is allowed to conceal the remaining supplier concentration.
Letters of intent are particularly weak evidence of supply. They are not purchase orders, allocations, prices or delivery schedules. Seven projects each seeking very large quantities would be exposed to product cycles and global competition for the newest parts. The tender can create demand faster than the supply chain can fulfil it.
Electricity may set the real clearing price
AP cites Commission analysis that electricity in Europe can cost two or three times as much as in the United States and China. The comparison varies by country, contract and load shape, but it points to the expense that public capital cannot wish away.
AI accelerators convert a large, persistent electricity input into compute output. A facility may receive subsidised construction and still produce expensive tokens or training runs if its power contract is uncompetitive. The eventual access price must absorb energy, cooling, network, maintenance, financing and frequent hardware replacement.
Power also constrains timing. The call expects awards in early 2027, construction to start during 2027 and operation within 18 months of contract signature. That schedule may be achievable at sites with secured transmission and permits. It is much harder where a new high-voltage connection, generation contract or major reinforcement has not begun.
The tender does not identify sites or disclose grid capacity, power sources, water use or cooling design. Those omissions are appropriate before selection but prevent a claim that the programme is already energy-efficient. Efficiency is a design objective. It becomes evidence only when operators disclose power utilisation, delivered compute and the cost and carbon profile of supply.
Nineteen factories and seven gigafactories are different tiers
Europe already has a network of 19 AI Factories intended to give companies and researchers access to AI-optimised supercomputers and support. The new tender adds a larger tier aimed at frontier training, inference and fine-tuning. AP says the seven planned facilities would more than double the current network’s computing power.
The distinction should not become a simple replacement story. Smaller factories can lower the access barrier, provide technical support and serve workloads that do not need frontier scale. Gigafactories concentrate large accelerator pools for models and applications that benefit from tightly coupled compute.
The economic question is how users move between the tiers. Start-ups need more than discounted machine time: they need data, software, engineering support and predictable access when an experiment becomes a product. If the gigafactories become reservation-heavy national assets, capacity may exist without a competitive commercial route.
Joint procurement could help by creating common access rules across countries. It could also produce political allocation. Eighteen states are pooling buying power, but not every one will host a site. The governance test is whether compute follows transparent performance and public-value criteria or becomes a geographic return for national co-financing.
The tender is new because it starts the contest
In January, the EU expanded EuroHPC’s mandate to include AI gigafactories. That legal change authorised the instrument. The 30 July event starts the procurement: eligibility is live, funding lots are specified, capacity steps are defined and the deadline is 12 November.
This is a materially different evidence point. A mandate shows that a public body may act. A tender tells private consortia what they must organise, how much support is available and when proposals must arrive. It begins price discovery for sites, hardware, capital and access.
Awards are expected in early 2027. Construction is expected to follow in 2027, and the selected facilities are expected to operate no later than 18 months after signature. Every verb in that sequence is forward-looking. A credible programme can still lose projects during financing, permitting, grid connection or procurement.
The next reporting task is therefore not to count seven pins on a map. It is to track qualified bids, committed private capital, disclosed access purchases, secured power, hardware contracts and milestones between award and live service.
What success should be measured against
The Commission’s first useful scorecard should reconcile four denominators. Public funding awarded must be separated from private capital contractually committed. Processor orders must be separated from equipment delivered. Installed hardware must be separated from compute accepted into service. Reserved access time must be separated from time actually used.
Price will matter as much as scale. European users gain little autonomy if subsidised capacity remains materially more expensive or harder to access than incumbent cloud services. Conversely, a narrowly priced public service that never covers replacement capital could become dependent on recurring subsidy.
Supplier diversity should be measured by more than the nationality of a building. Procurement can disclose accelerator, networking, software and cloud-stack concentration without exposing security-sensitive details. It can also show whether the optional allocation to European start-ups and scale-ups becomes a real purchase rather than a policy phrase.
Europe has now chosen an intelligible industrial mechanism: pool public demand, stage support and ask private consortia to build at frontier scale. That is more concrete than announcing technological sovereignty. Its success will depend on whether the buyer can turn a political objective into durable utilisation while power and chips remain priced by markets Europe does not control.