Can Europe support more frontier AI laboratories?

A frontier AI laboratory needs computing power, energy, data, people, and customers. Europe can supply most of these, some quickly and some slowly. The hardest gap is customers – European buyers who choose European AI. This is the first of two notes, and the second proposes how to fund new laboratories and give them their first customers.

The question

In my last note, “Who decides whether Europe’s AI keeps running?”, I argued that Europe needs at least one more frontier AI laboratory, meaning a company able to build the most capable AI models of the day. Europe already has AI laboratories, but only one independent European company, Mistral AI, works at or near that frontier. This note asks whether Europe can support more than it has today, condition by condition.

The short answer is yes, with two limits. Europe can supply cheap enough electricity, lawful training data, and skilled people, and it can buy computing power, though not quickly. First, Europe makes no chips able to train these models, so it buys them from American companies. US export rules do not restrict most European buyers today, but Washington can change them. Second, too few European companies and governments buy AI from European laboratories. In February 2026, Mistral AI reported annualised revenue of more than $400 million. Annualised revenue takes a recent month’s or quarter’s sales and scales them up to a full year. The figures reported for OpenAI and Anthropic later in the year are more than a hundred times that. Investors can carry a laboratory for some years, but over time only customers can pay for its next models, and too few of those customers are European.

Mistral AI itself now describes its place differently. Arthur Mensch, its co-founder and Chief Executive, told Bloomberg Television’s Wall Street Week in October 2026 that the company plays “in a different category”, supplying open models that customers run on their own systems rather than competing head-on with OpenAI and Anthropic. In the same programme, Jacob Kirkegaard, Senior Fellow at Bruegel, a Brussels economic think tank, said that Mistral AI had decided to stop developing closed frontier models, the kind only the company itself can run, and was no longer a frontier laboratory in the same sense. Its newest model, Mistral Large 4, released on 29 October 2026 for anyone to download and run, shows that it still works close to the frontier. But if Kirkegaard is right, no European company is now trying to compete at the frontier itself, and the case for a new one is stronger.

Kirkegaard also puts the strongest objection to this series. From an economic point of view, he argues, who makes the models matters less than having access to them. He then names the weakness himself, which is that Washington or Beijing could one day ban the export of their models to foreign customers. Access lasts only as long as those governments allow it, as my last note showed.

Computing power and chips

The leading American laboratories train their models on clusters of about 100,000 top-end AI chips or more, and are building larger ones. A cluster is a large group of chips in one building, wired to work as a single machine. The research firm SemiAnalysis estimates that a cluster of that size draws about 150 megawatts of power. Such sites are expensive. Epoch AI, an independent research group, estimates that a data centre drawing one gigawatt costs about $38 billion to build, or about $8.5 billion a year once that cost is spread over the site’s life and running costs are added. If costs scaled in proportion, a site of 150 megawatts would cost roughly $5.7 billion to build and $1.3 billion a year, though the real figure would depend on the hardware, the location, and the design.

The chips, and more and more often the complete systems built around them, come from two American companies, NVIDIA and AMD. NVIDIA sells not only chips but whole server racks and ready-made clusters, with the networking and software included. Europe’s own chip makers, such as Axelera, SiPearl, and Openchip, make chips for inference, which means running models that are already trained, or for general computing. None make chips for training, the far heavier work of building a model, at this scale. Until Europe has a chip that can train these models, a European laboratory would buy its training hardware from American suppliers.

Buying is allowed today. Most buyers based in the EU need no American export licence for these chips, though the rules still depend on who the buyer is, who owns it, and what the chips will be used for. The “AI Diffusion” rule, which would have limited sales to many countries, was withdrawn in 2025. Its text has not yet been formally removed from the regulations, and Washington has not published the replacement it promised. As my last note described, Europe’s supply rests on a decision in Washington not to restrict it.

Europe has no publicly documented cluster of that size today. Its largest public supercomputer, JUPITER in Germany, has about 24,000 such chips, according to EuroHPC, the EU body that co-funds it. The largest European-owned project, Mistral AI’s own computing centre near Paris, plans about 18,000 NVIDIA chips in its first phase, according to NVIDIA. The EU’s network of nineteen AI Factories, shared supercomputers open to researchers and companies, is useful for experiments but too small for a full training run.

The EU’s answer is its AI Gigafactories, up to seven much larger sites. After two extensions, bids for the largest sites close on 3 December 2026 and bids for medium-sized sites on 4 February 2027. Winners are expected in 2027, and the first sites may not run until 2028. A new laboratory would therefore need to buy or rent its own cluster in the meantime.

The gap is large, and money closes only part of it. Money can buy or rent a cluster, though building one takes years. It cannot remove the dependence on American suppliers and on decisions in Washington. Only diplomacy, or European chips many years from now, can do that.

Energy

The 2024 report on European competitiveness by Mario Draghi, former President of the European Central Bank, found that EU companies pay electricity prices two to three times those in the United States. That is true on average, but not everywhere. For the largest industrial users, those buying 150,000 megawatt-hours a year or more, Eurostat figures for the first half of 2026 show wide differences. A data centre drawing 150 megawatts would buy several times that amount. The table shows the cheapest countries, the largest economies, two of Europe’s main data-centre hubs, the Netherlands and Ireland, and two central and eastern European countries, Bulgaria and Poland.

Sources: Eurostat, nrg_pc_205, band IG (150,000 MWh a year or more), excluding VAT; Finland’s figure for this band is confidential; US figure, not like for like, from the EIA Electric Power Monthly, Table 5.3, January to June 2026, converted at the European Central Bank rate of 16 October 2026.

The US Energy Information Administration’s monthly figures put the average American industrial price at about 8.9 US cents per kWh over the first six months of 2026. These figures are still preliminary. At the European Central Bank’s exchange rate of 16 October 2026, that is roughly €0.079. The two measures are not identical, but prices in Portugal, France, and Spain are at or below the American level. Germany, Europe’s largest economy, is about a tenth above the EU average. These are national averages. A large data centre usually buys its power on long-term contracts, so what it pays depends on the deal it strikes, and on whether the grid can supply it at all.

The harder limit is the connection to the grid. In Europe’s main data-centre hubs, connecting a data centre to the grid takes seven to ten years on average, according to the energy research group Ember. In May 2026 Denmark’s grid operator paused new connection agreements for three months, mainly because of demand from data centres. Finland is preparing rules to decide which projects connect first.

The gap is moderate on price but harder on supply. Some countries offer prices close to the American level, but getting enough reliable power and a grid connection is the bigger obstacle. Grid connections need faster permits, which the EU’s proposed Cloud and AI Development Act addresses, but it is not yet law.

Data

The Draghi report says EU limits on storing and processing data create high compliance costs and hinder large training data sets. The rules do add cost and delay. A year after his report, Draghi cited research showing that the EU’s data protection rules had raised the cost of data for EU firms by about twenty per cent, compared with their American peers. But the evidence since then shows they do not stop training.

In December 2024 the European Data Protection Board, which brings together the EU’s privacy regulators, said that training AI on personal data can be lawful, case by case, under “legitimate interest” – a legal basis that weighs a company’s real need against people’s rights. In May 2025 Ireland’s regulator, which oversees most large US firms in the EU, did not stop Meta from resuming training on public posts once Meta had added safeguards, including an easier way for users to opt out. Under the EU’s copyright rules, anyone with lawful access to published text may use it to train AI, unless the owner has reserved that right in a form computers can read, such as a note in a website’s code. The copyright rules and the privacy rules are separate tests, and a data set can pass one and still raise questions under the other. A German appeal court applied this rule to a large public training data set in December 2025.

Mistral AI trains competitive models under these same rules. The delays on record mostly hit US companies’ consumer products, such as Meta’s eleven-month pause on training with European users’ posts.

Two risks remain. In November 2025 a Munich court ruled that a model which memorises song lyrics and repeats them breaks copyright. OpenAI has appealed, and the case is pending. And the European Commission’s plan to write AI training into the privacy rules, part of its “Digital Omnibus”, is still not agreed.

The gap is small to moderate. Money can buy licences and legal teams. Legal certainty needs new law and higher-court rulings.

People

Europe’s universities produce the people a laboratory needs. In 2024, universities in the EU awarded 4,194 doctorates and 68,375 master’s degrees in computing and related fields, according to Eurostat. In the same year, US universities awarded 2,982 doctorates in computing, and sixty-one per cent of them went to students on temporary visas, according to the US National Science Foundation. The latest official US figure for master’s degrees, for 2021–22, is 51,338, according to the National Center for Education Statistics. Relative to its young population, however, the EU produces fewer such graduates than the United States. The Commission’s Education and Training Monitor counts 2.6 computing graduates for every 1,000 people aged twenty to thirty-four in the EU, against 3.7 in the United States.

Europe trains many of the world’s best AI researchers but keeps too few of them. A Carnegie Endowment study of the authors at NeurIPS 2025, the field’s largest research conference, found that Europe, which the study counts without the United Kingdom, keeps sixty per cent of the top researchers it trains. The study follows researchers who publish at that conference, not the whole AI workforce. The United States keeps eighty-nine per cent and China sixty-nine per cent. China has started to bring its talent home. Its share has risen from fifty-seven per cent in 2022, and it now keeps more of its top researchers than Europe does.

Pay is one reason. US visa filings give an indication, with salaries of about $300,000 a year for technical staff at OpenAI and at Mistral AI’s American arm, before any shares. With shares, the median pay of a software engineer at OpenAI in the United States is about $875,000 a year, according to Levels.fyi, a website that collects pay data. The bigger gap is in shares. Shares in a US laboratory valued at hundreds of billions of dollars are worth far more. In some EU countries, employees also pay tax on their shares before they can sell them.

The talent has not left Europe so much as changed employer. Anthropic now has offices in London, Dublin, Zurich, Paris, Munich, and Milan, with room for up to 800 staff in London alone. OpenAI is adding 200 roles in Dublin, mostly engineers who help customers put its models to work, plus sales staff. Not all of these jobs are research jobs, but they draw on the same pool of skilled people. More and more of Europe’s AI talent works in Europe for American firms.

The gap is moderate, and money can largely close it. Higher pay and access to computing power are both matters of money. The EU’s “EU Inc.” proposal, which would delay tax on employee shares until they are sold, would help further, but the European Parliament and the member states are still negotiating it.

Customers

In my assessment, this is the decisive gap. Each new generation of models costs more than the last. Epoch AI estimates that the cost of training the most capable models has grown about 2.4 times a year since 2016. Investors can pay for that for some years, but over time the money has to come from customers. OpenAI’s annualised revenue was about $50 billion at the end of September 2026, and Anthropic’s about $65 billion at the end of July, according to figures reported by CNBC. Mistral AI told the Financial Times in February 2026 that its annualised revenue was more than $400 million. It aims for more than $1 billion a year in recurring revenue from customer contracts by the end of 2026. The figures come from different months and are not audited, but even at that target Mistral AI would earn one fiftieth or less of what each American rival earns.

The problem is not a lack of demand. IDC, a market research firm, expects European spending on all kinds of AI products and services, not only on models, to reach nearly $470 billion a year by 2030, growing by about a third each year. The problem is that European demand goes to American suppliers.

• Use is spreading fast. Twenty per cent of EU companies with ten or more staff used AI in 2025, up from 13.5 per cent a year earlier and eight per cent in 2023 (Eurostat). Large companies lead at fifty-five per cent, against seventeen per cent of small ones. Use is highest in Denmark, at forty-two per cent of companies, and lowest in Romania, at five per cent. In Bulgaria it is under nine per cent.
• European companies are not behind on trying AI. A survey by the European Investment Bank found that thirty-seven per cent of EU companies used generative AI, the kind that writes text and makes images, in 2025, against thirty-six per cent in the United States.
• But they use it less deeply. Among companies that use AI, fifty-five per cent in the EU apply it across several parts of their business, against eighty-one per cent in the United States (European Investment Bank). Most EU companies that use AI buy ready-made software rather than build their own (Eurostat).
• And they buy American. In Germany, ChatGPT is used by seventy-six per cent of companies that use AI, Microsoft Copilot by thirty-five per cent, and Google Gemini by twenty-eight per cent. Not one company in the survey reported using Mistral AI (Bitkom, the German digital industry association, September 2026). Chinese models barely register either. Two per cent use DeepSeek and one per cent use Alibaba’s Qwen. The survey asked about named products, so it may miss companies that run Mistral AI’s openly published models on their own servers. That use earns Mistral AI no usage fees, however, and does little to pay for its next model. Yet in Bitkom’s survey a year earlier, when German companies were asked which countries they would prefer AI to come from, ninety-three per cent named Germany and fifty-one per cent named the United States. The wish to buy closer to home is there. The survey does not say why so few act on it. A European product may not yet match the American ones, or it may be harder to find, less well built into the software companies already use, or less well supported.
• The same is true of cloud computing. Amazon, Microsoft, and Google take seventy per cent of Europe’s cloud market, where companies rent computing power and storage over the internet. All European providers together take fifteen per cent, down from twenty-nine per cent in 2017 (Synergy Research).
• Public buyers are no different. The EU’s proposed Cloud and AI Development Act would make public bodies give weight to European suppliers, but it is still a proposal and may take years to apply.

There are signs of what works. In December 2025, France’s Ministry of the Armed Forces signed an agreement with Mistral AI that covers the armed forces, the ministry’s own departments, and its research bodies. The shipping group CMA CGM signed a five-year contract worth €100 million. These are serious organisations putting AI into real work, with a European supplier.

Buyers also have a reason to choose European that has nothing to do with patriotism. The AI Strategic Autonomy Framework (AI-SAF), which I maintain, measures how far a country or company can keep its AI running if a foreign supplier or government cuts it off. A company scores higher when it has a working alternative outside the reach of the government that oversees its main supplier, such as a second supplier in Europe or open models it runs on its own servers. The same purchase that gives a European laboratory a customer also makes the buyer harder to cut off.

The gap is very large. Money alone cannot close it. Companies and public bodies are unlikely to switch unless the alternative is as good as what they already use. A European laboratory therefore needs two things at once. It needs enough money to build models that match the American ones, and buyers willing to try them once they do.

The five conditions side by side

Condition	Where Europe stands	Gap	Can money close it?
Computing power and chips	No publicly documented cluster of the needed size; hardware only from US suppliers, sales allowed today; Gigafactories from 2027–28	Large	Partly – money buys hardware, not independence from Washington
Energy	Prices at or below US levels in Portugal, France, and Spain; grid connections slow	Moderate	Partly – site choice and permits
Data	Training legally possible within privacy and copyright rules; some legal risk	Small to moderate	Partly – certainty needs law
People	Strong talent, 60 per cent kept; US firms hiring locally	Moderate	Largely – pay and share-tax reform
Customers	Mistral AI’s reported revenue is under 1 per cent of the US leaders’; demand goes to US suppliers	Very large	No – needs buyers

What this means

Europe can support more frontier laboratories than it has today. Most of what it needs can be bought or built, and the rest depends on choices Europe controls. These are where to build, how fast to grant permits, and how to tax employee shares. Only the supply of chips depends on a decision potentially made elsewhere.

Two things are missing, and they go together. The first is money on a scale that matches the task. A cluster, a team, and several years of training cost billions before the first euro of revenue. The second is customers. A laboratory that cannot sell its models in Europe will find it hard to last, and may be sold to a buyer from outside.

Money without customers builds a laboratory that cannot last. Customers without money have nothing European to buy. The answer has to address both. It needs large-scale investment and a focused effort to put AI to real use in European companies and public services. The governments that invest can also be the first customers, buying for their own work as France’s armed forces already do.

My next note proposes how to do both.

Sources

• Working Notes, “Who decides whether Europe’s AI keeps running?” (18 October 2026)
• European Commission, “The future of European competitiveness: Report by Mario Draghi” (9 September 2024), Part A, Chapter 2, p. 30
• European Commission, Keynote speech by Mario Draghi, “One year after the Draghi report” (16 September 2025)
• SemiAnalysis, “100,000 H100 Clusters: Power, Network Topology, Ethernet vs InfiniBand, Reliability, Failures, Checkpointing” (17 June 2024)
• Epoch AI, “Servers account for 60% of the total cost of ownership of a one-gigawatt AI data center” (14 May 2026)
• EuroHPC Joint Undertaking, “JUPITER: Launching Europe’s Exascale Era” (5 September 2025)
• EuroHPC Joint Undertaking, change notice, call for tenders EUROHPC/2026/OP/0008 (1 October 2026)
• Eurostat, “Electricity prices for non-household consumers – bi-annual data” (nrg_pc_205), band IG, first half of 2026
• Ember, “Grids for data centres: ambitious grid planning can win Europe’s AI race” (19 June 2025)
• European Data Protection Board, Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models (17 December 2024)
• Irish Data Protection Commission, “DPC statement on Meta AI” (21 May 2025)
• DLA Piper, “Robert Kneschke v. LAION”, on the Hamburg Higher Regional Court judgment of 10 December 2025
• Legal Tribune Online, “LG München I: GEMA siegt gegen OpenAI im Streit um Liedtexte” (11 November 2025)
• Carnegie Endowment for International Peace, “Who’s ahead in the global AI talent race?” (23 September 2026)
• Stanford HAI, “The 2026 AI Index Report”
• The Next Web, “OpenAI adds 200 Dublin jobs and builds its second applied AI lab” (12 October 2026)
• Eurostat, “20% of EU enterprises use AI technologies” (11 December 2025)
• IDC, “European AI Spending on Track for $470 Billion by 2030” (17 September 2026)
• Synergy Research Group, “European Cloud Providers’ Local Market Share Now Holds Steady at 15%” (24 July 2025)
• Hargreaves Lansdown/Sharecast, “Mistral invests in Swedish infrastructure as annual revenues top $400m” (11 February 2026), reporting the Financial Times interview
• CNBC, OpenAI revenue run rate (8 October 2026)
• CNBC, “Anthropic says annualized revenue climbed to $65 billion in July” (17 August 2026)
• European Commission, “Proposal for the Cloud and AI Development Act”, COM(2026) 502 (3 June 2026)
• European Investment Bank, “European firms show resilience, invest in green transition and match US companies in adopting AI technologies” (14 October 2025)
• Bitkom, “KI aus China für deutsche Wirtschaft kein Thema” (9 September 2026)
• Bitkom, “Durchbruch bei Künstlicher Intelligenz” (15 September 2025)
• NVIDIA, blog post on sovereign AI infrastructure in France and Mistral Compute (June 2025)
• European Commission, “AI Factories” (updated 12 August 2026)
• Wiley, client alert on the rescission of the AI Diffusion rule (May 2025)
• Holland & Knight, note on BIS guidance on licence requirements for advanced computing items (5 June 2026)
• DatacenterDynamics, report on the Danish grid operator’s three-month moratorium on new connections (7 May 2026)
• Borenius, legal alert on Finland’s new grid connection priority rules (26 August 2026)
• urheber.info, report on OpenAI’s appeal in GEMA v OpenAI (10 December 2025)
• Agence Europe, report on the postponed Council vote on the Digital Omnibus (7 October 2026)
• Ellis, Mistral AI Inc salaries from US labour filings and OpenAI OpCo LLC salaries (updated August 2026)
• Levels.fyi, “OpenAI Software Engineer Salaries” (accessed 20 October 2026)
• Reuters via Yahoo Finance, “Anthropic to boost hiring in Europe after opening Milan office” (28 May 2026)
• CoStar, “British Land and RLAM sign AI pioneer Anthropic for 158,000 sq ft London HQ” (17 April 2026)
• Osborne Clarke, note on the Commission’s EU Inc. proposal (21 May 2026)
• Implicator.ai, report on Mistral AI’s €3 billion round and revenue target (8 September 2026)
• Acteurs publics, report on the Ministry of the Armed Forces’ framework agreement with Mistral AI (9 January 2026)
• CMA CGM via AJOT, “The CMA CGM Group adopts custom-designed AI solutions from Mistral AI” (7 April 2025)
• Eurostat, “Graduates by education level, programme orientation, sex and field of education” (educ_uoe_grad02), 2024 data (updated 6 October 2026)
• European Commission, Education and Training Monitor 2025, comparative report, chapter 1
• US National Center for Education Statistics, Digest of Education Statistics 2023, Table 325.35
• US National Science Board, Science and Engineering Indicators 2024: Higher Education
• US National Science Foundation, National Center for Science and Engineering Statistics, Survey of Earned Doctorates 2024, data highlights (NSF 26-315) (13 March 2026)
• US Energy Information Administration, Electric Power Monthly, Table 5.3, “Average Price of Electricity to Ultimate Customers” (24 September 2026, data to July 2026)
• Epoch AI, “How much does it cost to train frontier AI models?” (3 June 2024, updated 13 January 2025)
• Bloomberg Television, Wall Street Week, “Why Mistral Is Betting on a Different Kind of AI” (October 2026)
• SiliconANGLE, “Mistral releases open weights for Mistral Large 4” (29 October 2026)
• George Alexandrov, “AI Strategic Autonomy Framework (AI-SAF)”