Local AI Report #5 — Europe's AI Compute Gap and Sovereignty

Issue #5 · September 14, 2026 · Last updated: September 14, 2026 · Jakub Rusinowski

The EU runs just 2 GW of AI compute to America's 35 GW. Inside Europe's 19 AI Factories, the gigafactory tender, and Luxembourg's 20-exaflop sovereignty bet.

The scoreboard

The United States runs 35 gigawatts of AI compute. China runs 5. The European Union runs 2.

That is roughly 17.5 times more American capacity than European, and it is the number every other number in this report has to be read against. The EU's share of global AI compute is about 5%.

Here is the part that should give you pause. Every announced programme, every billion committed, every ribbon cut between now and 2031 — and the EU's share lands at 5.6%.

Five years of effort, and the gap barely moves.

Global capacity grows more than eightfold over the period. The EU grows with it, from 2 GW to about 21 GW, which is real growth and not nothing. But everyone else grows too, and the US grows from a base seventeen times larger. China does better than Europe on this measure, moving from 11% to around 15%.

Relative position is what matters for the things people actually worry about — whether a European company can train a frontier model at home, whether a European hospital can run inference under European law, whether the continent has enough leverage to be more than a customer.

On that measure, the trajectory is flat.

One caveat worth stating up front, because the rest of this report depends on you trusting the numbers. The 2031 projections are a forecast, and forecasts in this industry have a poor record. The ratio and the share are solid across sources; treat the precise gigawatt figures for 2031 as approximate.

Why Europe is behind isn't the reason you'd guess

76 of the 101 AI compute projects planned in the EU are financed entirely by private firms. Those projects account for 84% of all planned capacity.

Sit with that, because it contradicts the story everyone tells. The usual account of European technology failure is that capital is timid, that there is no risk appetite, that everything needs a subsidy. In AI infrastructure, that is not what the pipeline shows. Capital has turned up.

What has not turned up is permission.

The figure above needs one careful sentence, because it is widely misreported. Those 24 and 42 months are not how long permitting takes. They measure what happens *after* a project has already secured every regulatory permit and confirmed how it will get its power — the time from there to the facility first going online. Germany takes nearly twice as long as the United States at a stage where all the paperwork is supposedly done.

Then there is the grid, which is worse. In Frankfurt, London, Amsterdam, Paris and Dublin, a new large-load connection takes 7 to 10 years. That estimate comes from Ember, cited in Bruegel's analysis.

Seven to ten years. A GPU generation is roughly two. A company that starts queuing for power today is planning around hardware that has not been designed yet.

This reframes the whole policy problem. If the binding constraint were capital, subsidies would fix it, and the EU has subsidies. If the binding constraint is the number of years between deciding to build and being allowed to switch on, then a subsidy is a way of paying people to wait.

Bruegel's own recommendations follow from this: concentrate support where permitting is fastest rather than spreading it evenly, create acceleration zones with a single permitting authority, and put grid investment ahead of compute subsidies in the queue. Reasonable people disagree about the last one, and we come back to that disagreement near the end.

Two different EU programmes, constantly confused

An AI Factory and an AI Gigafactory are not the same thing, and the difference is not pedantic — one is running today and the other has not broken ground.

We are being blunt about this because most coverage of European AI infrastructure treats the two names as interchangeable, sometimes inside a single article. If you leave this report able to tell them apart, you will understand the European compute story better than most of what is written about it.

AI Factories are upgrades. EuroHPC took supercomputers that already existed, or sites that already had the power and the staff, and added AI-optimised hardware. There are 19 of them across 16 countries. They are aimed squarely at startups, SMEs, researchers and public bodies, and the whole network represents around €1.5B of combined EU and national funding.

AI Gigafactories are new construction at a different scale. 7 sites, each specified to house at least 100,000 AI chips. The financing structure is €10B of public money — EU budgets plus national contributions — against a target of at least €20B in private investment, so more than €30B in total if it comes together.

Applications close in November, with decisions expected early 2027 and construction beginning later that year. Selected sites are expected to be operating within 18 months of signing.

Two numbers get conflated here constantly, so keep them apart: InvestAI is a €200B mobilisation target overall, and the gigafactory fund inside it is €20B. They are different scopes.

And a third category belongs in this picture, one that involves no EU money at all: the sovereign cloud providers already selling EU-jurisdiction compute today. Those are companies, not programmes. Section seven covers them.

Europe's 19 AI Factories, in full

19 sites, 16 countries, three selection rounds. Here is the complete roster.

The three rounds matter, because the middle one is the one most published lists leave out — and it is the round that contains Europe's two exascale supercomputers. JUPITER at Jülich and Alice Recoque at Bruyères-le-Châtel are the largest machines on the continent, and both arrived in the tranche selected in March 2025.

First selection — December 2024 (7 sites)

Second selection — March 2025 (6 sites)

Third selection — October 2025 (6 sites)

A few things worth noticing in that table.

Germany, Poland and Spain host two sites each, which is a reminder that these were awarded to consortia rather than parcelled out one per member state. The second Spanish site, at CESGA in Galicia, is health-focused; Poland's second, at Cyfronet in Kraków, works on healthcare, space and language models.

Italy's IT4LIA at CINECA in Bologna is the most technically interesting of the group. It pairs NVIDIA Grace CPUs with Blackwell GPUs, and it carries a dedicated inference partition built on Axelera AI accelerators and SiPearl CPUs — European silicon inside a European sovereign site. That is unusual, and it matters for a reason section seven returns to.

One honest note on this roster: the countries, cities and host systems are well established, but several of the March 2025 factory acronyms are inconsistently reported. We have listed the sites by country and host rather than leaning on names we could not pin down.

The country punching furthest above its weight

Luxembourg has about 660,000 residents. That is fewer people than Frankfurt. It is building a supercomputer rated at over 20 exaflops of AI-precision peak performance.

The system is called MeluXina-AI, and the specification is worth reading slowly: 1,008 NVIDIA GB200 NVL4 GPUs across 252 liquid-cooled nodes, built on the Blackwell architecture.

The contract is €80 million, covering acquisition, delivery, installation and maintenance. EuroHPC funds 50% through the Digital Europe Programme (DEP); the Luxembourg government funds the other 50% from its national budget. That is an unhedged national bet from a country whose entire population would be a mid-sized district in Berlin.

E4 Computer Engineering is the integrator, with Dell Technologies supplying the servers — the same pairing building Italy's IT4LIA. The system is split across two sites, Bissen and Bettembourg, with installation starting in autumn 2026. LuxProvide operates it. The Luxembourg AI Factory has already been running on the existing MeluXina system since spring 2025, so this is a second phase rather than a standing start.

On that 20-exaflop figure: it is AI-precision peak performance, and EuroHPC describes it as an estimate. It is not FP64 Linpack, and the two are not comparable. Anyone setting an AI-precision number beside a classical HPC number without saying so is selling you something.

What makes Luxembourg interesting is not the raw capability. It is the coherence.

The mandate is explicitly built around SME and researcher access, and the strategic focus is finance, space, cybersecurity — which happen to be precisely the three sectors Luxembourg's economy actually runs on. This is not a country buying a supercomputer because supercomputers are prestigious. It is a financial centre buying compute for finance, a space-industry hub buying compute for space, and a jurisdiction that sells trust buying compute for cybersecurity.

Now the part that makes this section worth the space.

Luxembourg is a named customer of Mistral AI, Europe's flagship model developer, under a partnership covering government operations and public research institutions. So the country is building sovereign compute and buying the sovereign model to run on it.

And in the €3 billion round that valued Mistral above €21 billion, one of the investors was the Grand Duchy of Luxembourg.

Builds the compute. Buys the model. Owns a piece of the company.

That is what a sovereignty strategy looks like when a state actually has one, and it is being executed by a country smaller than most European cities. Luxembourg cannot outspend Germany or France, and it is not trying to. It moved faster and it bought a coherent stack — which is exactly the speed-over-scale lesson from section two, demonstrated rather than argued.

What 'data sovereignty' actually means

Data sovereignty is jurisdiction: where your training data, model weights and inference traffic physically sit, and which legal system can compel access to them.

The concrete mechanism most European legal teams worry about is the US CLOUD Act, which lets US authorities compel US-headquartered providers to produce data regardless of which country it is stored in. That is why "we host in an EU region" satisfies some compliance reviews and not others — the region is not the question, the parent company's jurisdiction is. We go through that distinction properly in what sovereign AI actually means, and the GDPR angle covers the data-residency mechanics.

If you need EU-jurisdiction compute today rather than in 2031, these are the providers usually on the shortlist.

Read that last column, because it is the honest part.

Every entry has a caveat, and that is the point rather than a disclaimer. Sovereignty is a spectrum, not a checkbox. The OVHcloud case is the sharpest illustration: the company holds SecNumCloud, France's highest cloud security qualification, and it is genuinely a meaningful credential — but it does not currently extend to the GPU rental products an AI buyer would actually purchase. The certification the provider is known for does not cover the thing you came to buy.

There is a deeper version of the same problem. Almost every one of these providers runs NVIDIA or AMD silicon on largely US-origin software stacks. A European company, in a European data centre, under European law, training on American chips is more sovereign than the alternative — and it is not fully sovereign. That is why Italy's Axelera and SiPearl partition is worth noticing: it is one of the few places where the stack goes European further down.

None of this means sovereign cloud is theatre. It means you should know which specific exposure you are buying down, and check that the certification covers the product rather than the company.

Mistral AI: the company Europe's bet rests on

Three researchers left Google DeepMind and Meta in spring 2023, started a company in Paris, and shipped a 7B open-weight model that autumn that was good enough to make everyone pay attention.

Arthur Mensch (CEO, ex-Google DeepMind), Guillaume Lample (Chief Scientist, ex-Meta), Timothée Lacroix (CTO, ex-Meta) founded Mistral in Paris. The company is now somewhere around 1,500 people — third-party trackers, since Mistral does not publish headcount (checked 2026-09-14).

The shape of that curve is the story. A €240M valuation at the seed round in mid-2023. More than €21 billion a little over three years later.

The most recent round raised €3 billion led by Samsung Electronics, with EQT (Scaleup Europe Fund) and PSG Equity as co-leads, and has been reported as the largest equity fundraising ever completed by a European technology company. Advent International, BlackRock and the Grand Duchy of Luxembourg also took part.

The round before it, in September 2025, raised €1.7 billion at €11.7 billion, led by ASML — the Dutch lithography company without which essentially no advanced chip gets made anywhere on earth. ASML took 11% fully diluted at that round, becoming the largest shareholder. Note the qualifier: that was the stake in September 2025, and the round that followed will have diluted it.

Worth a footnote for anyone who remembers the 2024 headlines: Microsoft's €15M investment, alongside making Mistral's models available on Azure, was announced in February 2024, and it drew an examination from EU competition regulators at the time. You will see that date misreported.

The strategic turn matters more than the numbers. Mistral is no longer only a model developer — it is becoming an infrastructure owner. The company has acquired Koyeb (Paris-based serverless infrastructure) and Emmi AI (Austrian industrial-simulation startup), raised debt earmarked for datacentre capacity near Paris and in Sweden, and has stated a goal of 1 GW of European compute capacity by 2030.

Hold that last figure against section one. The entire European Union operates 2 GW of AI compute today. One private company is targeting 1 GW by 2030 — half of what the whole bloc runs now.

Either that is the single most important fact about European AI infrastructure, or it is a target that will slip. It is worth watching which.

If you want the models rather than the company, Mistral Large 3 is the current flagship and Mistral Small 4 is the one most people can actually run locally.

Who is actually building on it

Airbus, BNP Paribas, Stellantis, SNCF, TotalEnergies, the European Patent Office and the French Ministry of Armed Forces are all named Mistral customers.

That list is the difference between a sovereignty policy and a sovereignty slogan. These are not pilots at digital agencies — they are some of the largest industrial, financial and public institutions in Europe.

Three are worth pulling out.

Stellantis has repeatedly widened its Mistral partnership rather than running a single contained pilot — the pattern of a company that found the first deployment worked. BNP Paribas extended its partnership into a further phase of generative-AI rollout, which for a bank of that size means the compliance function signed off more than once. And Capgemini is a technology partner rather than only a customer, integrating and reselling Mistral inside client engagements, which quietly puts the models into organisations that never evaluated them directly.

ASML appears in two places in this report — as Mistral's largest shareholder after the 2025 round, and as a customer. That is worth a raised eyebrow and not a scandal; strategic investors frequently buy what they back.

These are the customers Mistral names publicly, which is not the same as an audited list. It is still a substantially longer one than any other European model developer can show.

The rest of the field

Mistral is not the only European bet, but it is the only one with the funding, the government customers and the infrastructure position to anchor a full sovereign stack today.

Aleph Alpha in Germany targets enterprise and government buyers, positioning on data sovereignty and explainability for regulated industries. Nebius, covered above as an infrastructure provider, is a model-adjacent business rather than a lab. H Company and Poolside, both French, have frontier ambitions and a fraction of the scale.

The instructive case is Silo AI. Finland's flagship AI company, a genuine European success story — acquired by AMD in 2025, and now part of a US company's stack.

Nothing improper happened. It is simply what "European AI company" means as a category: a label that can change hands. Any argument about sovereign AI capability that rests on a list of company names should be read with that in mind.

The honest disagreement nobody resolves for you

There is a real fight underneath this, between credible people, and it is not the one you would expect. It is not whether Europe should have sovereign AI. It is whether sovereignty rules would make the compute gap better or worse.

The case for strict rules — EU-ownership requirements, data-localisation mandates, limits on foreign hyperscalers — is the direction the EU's proposed Cloud and AI Development Act leans. It reduces CLOUD Act exposure, builds capacity that is genuinely independent, fits the regulatory posture the EU has already adopted elsewhere, and avoids a dependence that could be revoked by another government's decision.

The case against comes from Bruegel, and it is an argument about sequencing rather than about principle. If Europe's binding constraint is speed — permitting queues and grid connections, as section two argued — then strict input-side rules about who owns the rack attack the wrong variable. Worse, they risk deterring the hyperscaler investment that is currently the fastest route to capacity, at the moment Europe most needs capacity built. Bruegel's alternative is output-side rules: "processed in the EU", "data does not leave EU jurisdiction" — the sovereignty outcome without the ownership-structure friction.

A third position has appeared in 2026 analyses: that chasing hyperscale parity is itself the wrong goal, and Europe should aim for enough control at the chokepoints to protect public choice and critical services, rather than trying to match American capacity it will not match.

We are not going to resolve this for you, and you should be suspicious of anyone who does in a single article. Note only that Bruegel is a credible, non-partisan economics institute arguing against the direction of the EU's own draft legislation — which is a genuine disagreement among serious people, not a manufactured controversy. The EU AI Act's on-premise implications are where this stops being abstract for most organisations.

What this means if you're choosing where to run your models

Most of this report is about infrastructure you cannot buy. This section is about the part you can.

If jurisdiction genuinely matters to your use case — regulated data, public-sector work, a contract with a data-residency clause — then the sovereign cloud providers in section seven are available now. Check the certification covers the specific product, not the company. OVHcloud's SecNumCloud is the cautionary example.

If you want the strongest jurisdictional guarantee, the answer has not changed and it is not in this report's headline: run the model yourself. A model on your own hardware raises no question about whose courts can compel whose disclosure. That is the whole premise of running models locally for privacy, and offline deployment takes it further. For organisations, on-premise total cost of ownership is where that decision usually gets made or unmade, and our laptop report covers what the hardware actually costs.

If you are a European public body, Luxembourg is the worked example — and the AI Factories exist specifically to give SMEs and researchers access. That access is the programme's stated purpose, not a side effect. Sovereign AI in government contexts goes into procurement specifics.

What is not usable yet: the gigafactories. Not one exists. Applications are still open, decisions come next year, and the earliest facilities would be operating some eighteen months after contracts are signed. Anyone selling you a 2026 strategy built on gigafactory capacity is selling a 2029 strategy.

The honest summary is that Europe's sovereign compute story is genuinely real at the small end — you can rent EU-jurisdiction GPUs today, and you can run open-weight European models on your own machine this afternoon — and genuinely unbuilt at the frontier end. Our small-model report covers what runs on hardware you already own, which remains the most sovereign option available to anyone.

Frequently Asked Questions

Related

Sovereignty and compliance: What sovereign AI means · GDPR and AI · The EU AI Act on-premise · Sovereign AI for government

Running models yourself: Local LLMs for privacy · Offline AI · On-premise TCO

Earlier reports: Best laptops for local AI · Small LLMs for 8GB and 16GB laptops

Methodology. Estimates are labeled as estimates; verified figures link to their sources. VRAM is weights + KV cache + framework overhead; speed is a three-term latency model. Both formulas are published at /en/methodology.

← All Reports