The European Union has opened bidding for up to seven AI Gigafactories, putting a €30 billion public-private target behind the idea of “sovereign AI” and, more importantly, turning it into a procurement program with a deadline. Consortia have until November 12, 2026 to compete for public support and long-term demand for compute capacity; selected projects are expected to be chosen in early 2027, with operations due no later than 18 months after contracts are signed.
The European Commission’s July 30 announcement promises as much as €10 billion in combined EU and national public support, intended to mobilize at least €20 billion in private investment. AP, IT Pro, and Le Monde have independently reported the core numbers and the planned ceiling of seven facilities. The Commission says the sites will serve frontier-model training, fine-tuning, and inference for companies, researchers, public authorities, startups, and SMEs.
But the tender’s most consequential detail is easy to miss beneath the €30 billion headline: Brussels is not simply building and owning seven giant European supercomputers. EuroHPC and participating governments will procure guaranteed compute access time from private-led consortia. The EU is buying capacity, rules, and influence over access rather than declaring immediate independence from the hardware and cloud suppliers that make that capacity possible.
That is a more realistic version of AI sovereignty. It is also a much narrower one.
The call is structured around two lots and two development phases. Up to four projects in the first lot can receive as much as €100 million in EU funding during phase one and a further €400 million in phase two; up to three larger projects in the second lot can receive as much as €200 million initially and up to €800 million later. Host member states are expected to match EU funding, while private partners provide the much larger share of construction and operating capital.
Those figures describe eligibility ceilings, not money already awarded. A seven-site outcome would exhaust the theoretical public maximum, but only if every project advances through both phases and participating states provide their matching contributions. The Commission’s €30 billion figure is therefore a mobilisation target: €10 billion in public support plus a minimum €20 billion in private investment. It is not a €30 billion EU appropriation sitting in a completed budget.
The legal structure backs up that reading. Council Regulation 2026/150, which expanded EuroHPC’s mandate for AI Gigafactories in January, allows the Union contribution to take the form of pre-agreed, guaranteed purchased access time. It also anticipates arrangements in which infrastructure ownership can later transfer to the consortium, with the residual value converted into equivalent access rights for the Union.
In practical terms, the public sector is using its balance sheet to make enormous AI infrastructure projects financeable, then reserving capacity for European users. That can matter a great deal to a university, public research body, startup, or public agency that cannot negotiate multi-year accelerator commitments with a hyperscaler. But it is not the same model as a state-owned compute utility with permanent control over the stack.
Eighteen member states have signed the joint procurement agreement with EuroHPC: Croatia, Czechia, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Latvia, Lithuania, Poland, Portugal, Slovakia, Spain, and Sweden. Their participation creates a shared market for reserved compute, but it also makes the programme dependent on cross-border agreement about who pays, who gets priority, and what workloads qualify.
The actual sovereignty test will be found in the hosting agreements: the amount of public access time, its price after the guaranteed period, rules for startups and public workloads, security conditions, and what happens when commercial demand conflicts with public commitments. The Commission’s announcement identifies the access mechanism, but it does not publicly disclose those operating terms.
That change is not necessarily a retreat, but it does expose how the headline numbers work. The original €20 billion was always presented as money to be mobilised, while the current call makes the split visible: roughly one-third public backing and at least two-thirds private capital. Le Monde reported that the €10 billion public figure consists of €5 billion from the EU budget and a matching €5 billion from host member states, and that a substantial portion of the EU share depends on the next long-term EU budget still being negotiated.
The Commission’s July 30 release does not spell out that budget timing. Nor does it say how much of the national matching money is committed today versus contingent on political approval after a winner is selected. That omission matters because data centers can begin consuming capital long before they deliver useful compute. A consortium can have a site, an industrial partner, and a glossy rendering without a final power contract, chips, debt financing, or the public money required for phase two.
The Commission says it received 77 expressions of interest from 16 member states across 60 potential locations. A prior Commission announcement described 76 submissions, a minor inconsistency that does not change the scale of interest but illustrates how preliminary the earlier exercise was. The tender now forces those speculative pitches into formal financing, technical, energy, and delivery commitments.
This is where the project becomes more than a declaration of intent. Some of Europe’s best-known infrastructure players have already signaled interest: France’s AION consortium includes EDF, Orange, Scaleway, Capgemini, Atos subsidiary Bull, and investment firm Ardian, while Spain has promoted a multi-site proposal involving Telefónica, ACS, Banco Santander, Multiverse Computing, and public entities. The winners will be the groups able to bind together power, land, financing, connectivity, cloud operations, hardware procurement, and customers—not simply the groups with the strongest sovereignty rhetoric.
That is the crucial disclosure. The EU is not claiming that the first wave of Gigafactories will train frontier systems on European-designed, European-manufactured accelerators. It is pursuing sovereignty over location, legal jurisdiction, user access, workload allocation, and some operational control while relying on US chip companies for the core processing hardware.
That approach is more honest than pretending Europe can quickly substitute for Nvidia-class accelerators. It also changes what the initiative can deliver. A facility operating inside the EU, under EU data-protection and security rules, with guaranteed capacity for European public and private users, is strategically useful even if its accelerator racks were designed in California or built through Asian supply chains. It reduces exposure to a sudden loss of capacity from an overseas cloud contract and gives European institutions a direct customer relationship with infrastructure on European soil.
Yet the letters of intent should not be confused with guaranteed chip supply. The Commission has not published their terms, volumes, pricing, delivery schedules, or priority rights. In a tight global accelerator market, the difference between a political assurance of “seamless access” and binding supply commitments is substantial. European bidders will still need to prove they can obtain hardware at a scale and timetable compatible with their construction plans.
The programme is also intended to create predictable demand for European AI processors and, eventually, domestic production. That ambition is long-term industrial policy, not a condition already met by the tender. The first Gigafactories may help European chip designers validate and sell specialized components, networking equipment, storage, cooling systems, and software. They will not erase the present dependence on foreign frontier accelerators.
For an AI facility, power availability is a deployment constraint, not an environmental footnote. A bidder needs enough firm electricity for high utilisation, grid connections that arrive on time, backup power, cooling capacity, water planning where applicable, and a commercial electricity arrangement that does not make training runs unaffordable. Energy-efficiency claims do not solve the problem if a project is stalled in an interconnection queue or faces power prices that make its hardware uneconomic to run.
The tender’s two-phase design is intended to limit this risk. Initial capacity must already equal, or in the larger lot reach up to twice, the level of Europe’s most powerful existing AI Factory. Later phases require expansion to three or four times that reference capacity. The structure gives EuroHPC a chance to avoid paying the full public contribution before a project proves it can deploy and scale.
Still, the benchmark is relative rather than a public chip count, energy target, or delivered-exaflop commitment. The July announcement does not identify the reference system, disclose a minimum number of accelerators per site, or set a public megawatt threshold. That leaves outside observers unable to compare the actual scale of competing bids until tender documentation and eventual awards provide more detail.
For Microsoft-centric organisations, the development is a reminder that European digital sovereignty is increasingly being enforced and financed at the infrastructure layer, not only through software compliance. The winning facilities will need Windows-compatible enterprise environments, identity controls, secure cloud access, high-bandwidth networking, storage, model-management systems, and governance that can support public-sector and regulated workloads. But the tender is fundamentally about reserving scarce GPU capacity, not replacing Azure, AWS, Google Cloud, or the software stacks built around them.
Europe has moved beyond speeches about compute sovereignty. It has now created a mechanism to pay for it. By early 2027, the question will be whether the winning consortia can turn pooled public purchasing power into live, affordable, and reliably supplied AI capacity—or whether Europe has principally subsidized European locations for imported chips and privately controlled clouds.
But the tender’s most consequential detail is easy to miss beneath the €30 billion headline: Brussels is not simply building and owning seven giant European supercomputers. EuroHPC and participating governments will procure guaranteed compute access time from private-led consortia. The EU is buying capacity, rules, and influence over access rather than declaring immediate independence from the hardware and cloud suppliers that make that capacity possible.
That is a more realistic version of AI sovereignty. It is also a much narrower one.
The EU Is Buying Access to Compute, Not Just Financing Buildings
The call is structured around two lots and two development phases. Up to four projects in the first lot can receive as much as €100 million in EU funding during phase one and a further €400 million in phase two; up to three larger projects in the second lot can receive as much as €200 million initially and up to €800 million later. Host member states are expected to match EU funding, while private partners provide the much larger share of construction and operating capital.Those figures describe eligibility ceilings, not money already awarded. A seven-site outcome would exhaust the theoretical public maximum, but only if every project advances through both phases and participating states provide their matching contributions. The Commission’s €30 billion figure is therefore a mobilisation target: €10 billion in public support plus a minimum €20 billion in private investment. It is not a €30 billion EU appropriation sitting in a completed budget.
The legal structure backs up that reading. Council Regulation 2026/150, which expanded EuroHPC’s mandate for AI Gigafactories in January, allows the Union contribution to take the form of pre-agreed, guaranteed purchased access time. It also anticipates arrangements in which infrastructure ownership can later transfer to the consortium, with the residual value converted into equivalent access rights for the Union.
In practical terms, the public sector is using its balance sheet to make enormous AI infrastructure projects financeable, then reserving capacity for European users. That can matter a great deal to a university, public research body, startup, or public agency that cannot negotiate multi-year accelerator commitments with a hyperscaler. But it is not the same model as a state-owned compute utility with permanent control over the stack.
Eighteen member states have signed the joint procurement agreement with EuroHPC: Croatia, Czechia, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Latvia, Lithuania, Poland, Portugal, Slovakia, Spain, and Sweden. Their participation creates a shared market for reserved compute, but it also makes the programme dependent on cross-border agreement about who pays, who gets priority, and what workloads qualify.
The actual sovereignty test will be found in the hosting agreements: the amount of public access time, its price after the guaranteed period, rules for startups and public workloads, security conditions, and what happens when commercial demand conflicts with public commitments. The Commission’s announcement identifies the access mechanism, but it does not publicly disclose those operating terms.
The €30 Billion Headline Is a Recast of the Original Plan
The programme’s public framing has shifted since Commission President Ursula von der Leyen announced InvestAI at the Paris AI Action Summit in February 2025. The earlier pledge described €20 billion mobilised for up to five AI Gigafactories. The current call describes up to seven sites, €10 billion in EU and national public money, and at least €20 billion from private investors.That change is not necessarily a retreat, but it does expose how the headline numbers work. The original €20 billion was always presented as money to be mobilised, while the current call makes the split visible: roughly one-third public backing and at least two-thirds private capital. Le Monde reported that the €10 billion public figure consists of €5 billion from the EU budget and a matching €5 billion from host member states, and that a substantial portion of the EU share depends on the next long-term EU budget still being negotiated.
The Commission’s July 30 release does not spell out that budget timing. Nor does it say how much of the national matching money is committed today versus contingent on political approval after a winner is selected. That omission matters because data centers can begin consuming capital long before they deliver useful compute. A consortium can have a site, an industrial partner, and a glossy rendering without a final power contract, chips, debt financing, or the public money required for phase two.
The Commission says it received 77 expressions of interest from 16 member states across 60 potential locations. A prior Commission announcement described 76 submissions, a minor inconsistency that does not change the scale of interest but illustrates how preliminary the earlier exercise was. The tender now forces those speculative pitches into formal financing, technical, energy, and delivery commitments.
This is where the project becomes more than a declaration of intent. Some of Europe’s best-known infrastructure players have already signaled interest: France’s AION consortium includes EDF, Orange, Scaleway, Capgemini, Atos subsidiary Bull, and investment firm Ardian, while Spain has promoted a multi-site proposal involving Telefónica, ACS, Banco Santander, Multiverse Computing, and public entities. The winners will be the groups able to bind together power, land, financing, connectivity, cloud operations, hardware procurement, and customers—not simply the groups with the strongest sovereignty rhetoric.
Imported Chips Are Built Into the Sovereignty Plan
The Commission’s own release makes the programme’s central dependency explicit. AI Gigafactory consortia may procure hardware from suppliers in Europe or “like-minded countries,” and Brussels says it has signed letters of intent with AMD, Nvidia, and Qualcomm to help ensure access to necessary hardware.That is the crucial disclosure. The EU is not claiming that the first wave of Gigafactories will train frontier systems on European-designed, European-manufactured accelerators. It is pursuing sovereignty over location, legal jurisdiction, user access, workload allocation, and some operational control while relying on US chip companies for the core processing hardware.
That approach is more honest than pretending Europe can quickly substitute for Nvidia-class accelerators. It also changes what the initiative can deliver. A facility operating inside the EU, under EU data-protection and security rules, with guaranteed capacity for European public and private users, is strategically useful even if its accelerator racks were designed in California or built through Asian supply chains. It reduces exposure to a sudden loss of capacity from an overseas cloud contract and gives European institutions a direct customer relationship with infrastructure on European soil.
Yet the letters of intent should not be confused with guaranteed chip supply. The Commission has not published their terms, volumes, pricing, delivery schedules, or priority rights. In a tight global accelerator market, the difference between a political assurance of “seamless access” and binding supply commitments is substantial. European bidders will still need to prove they can obtain hardware at a scale and timetable compatible with their construction plans.
The programme is also intended to create predictable demand for European AI processors and, eventually, domestic production. That ambition is long-term industrial policy, not a condition already met by the tender. The first Gigafactories may help European chip designers validate and sell specialized components, networking equipment, storage, cooling systems, and software. They will not erase the present dependence on foreign frontier accelerators.
Power, Not Permitting, Will Separate Credible Bids From Aspirational Ones
The Commission describes the Gigafactories as energy-efficient facilities with sustainable energy and water infrastructure. Those are necessary selection criteria, but they do not make electricity abundant or cheap. AP reported that a Commission assessment presented to the European Parliament found electricity can cost two to three times as much in the EU as in the United States and China; Le Monde also noted that the environmental requirements remain broad, leaving consortia to propose much of their own performance framework.For an AI facility, power availability is a deployment constraint, not an environmental footnote. A bidder needs enough firm electricity for high utilisation, grid connections that arrive on time, backup power, cooling capacity, water planning where applicable, and a commercial electricity arrangement that does not make training runs unaffordable. Energy-efficiency claims do not solve the problem if a project is stalled in an interconnection queue or faces power prices that make its hardware uneconomic to run.
The tender’s two-phase design is intended to limit this risk. Initial capacity must already equal, or in the larger lot reach up to twice, the level of Europe’s most powerful existing AI Factory. Later phases require expansion to three or four times that reference capacity. The structure gives EuroHPC a chance to avoid paying the full public contribution before a project proves it can deploy and scale.
Still, the benchmark is relative rather than a public chip count, energy target, or delivered-exaflop commitment. The July announcement does not identify the reference system, disclose a minimum number of accelerators per site, or set a public megawatt threshold. That leaves outside observers unable to compare the actual scale of competing bids until tender documentation and eventual awards provide more detail.
November 12 Is the First Real Test of Europe’s Compute Strategy
For European enterprises and public bodies, the most immediate outcome is not a new service to buy. It is a competition to determine which consortia will become long-term gatekeepers for publicly supported AI capacity. The winners could shape where sensitive workloads run, which cloud and orchestration stacks become entrenched, and whether startups receive meaningful access or find themselves behind industrial customers with larger commitments.For Microsoft-centric organisations, the development is a reminder that European digital sovereignty is increasingly being enforced and financed at the infrastructure layer, not only through software compliance. The winning facilities will need Windows-compatible enterprise environments, identity controls, secure cloud access, high-bandwidth networking, storage, model-management systems, and governance that can support public-sector and regulated workloads. But the tender is fundamentally about reserving scarce GPU capacity, not replacing Azure, AWS, Google Cloud, or the software stacks built around them.
Europe has moved beyond speeches about compute sovereignty. It has now created a mechanism to pay for it. By early 2027, the question will be whether the winning consortia can turn pooled public purchasing power into live, affordable, and reliably supplied AI capacity—or whether Europe has principally subsidized European locations for imported chips and privately controlled clouds.