Treatise · The History of AI in Europe
The homeland of Turing, the birthplace of LSTM, the author of the world's first comprehensive AI law — Europe plays a complex role in the history of AI: pioneer's glory, lost opportunity, and rule-maker's ambition all woven into one narrative. It has had the earliest ideas but not the largest market; it has the strictest rules but not the largest companies. And yet remove the European thread from AI history, and many of its key turning points become inexplicable.
I. Britain: The Spiritual Home of AI
The story begins in Cambridge.
In 1936, Alan Turing (1912–1954) in On Computable Numbers introduced the Turing machine, drawing the boundary of what is computable. Beginning in 1943, he was secretly dispatched to Bletchley Park, where in a Victorian country estate he led the Hut 8 team — together with thousands of cryptographers, mathematicians, and linguists — to break the German Enigma and Lorenz ciphers. After the war he returned to academia, and in 1950 published Computing Machinery and Intelligence, proposing the "imitation game" — what later generations call the Turing test, still the first thought experiment through which the public encounters AI. In 1954 Turing took his own life by poison, sixty-eight years before ChatGPT. He never saw the seeds he planted blossom; but when AI as a discipline first stood up, the foundation under its feet was his.
Bletchley Park has itself become a continuing symbol. In November 2023, the United Kingdom chose this very estate to host the first global AI Safety Summit (the Bletchley Summit), where twenty-eight countries plus the EU jointly signed the Bletchley Declaration. Seventy years on, the secret base for breaking Nazi codes had become the conference room where humanity discussed how to live with superhuman intelligence.
Britain's other academic stronghold was Edinburgh. Donald Michie (1923–2007) founded the Department of Machine Intelligence and Perception at Edinburgh in 1965; his student Robert Kowalski (1941–) later carried logic programming across Europe. But the 1973 James Lighthill (1924–1998) report nearly killed British AI — commissioned by the Science Research Council, it judged that AI had "failed to meet expectations" on general problem solving and recommended cutting funding. The first British AI winter followed and lasted into the late 1980s.
Britain's AI fire was rekindled by Demis Hassabis (1976–). A Cambridge computer-science alumnus and chess prodigy turned UCL neuroscience PhD, Hassabis founded DeepMind near London's King's Cross in 2010; in 2014, Google acquired it for around six hundred million dollars — the most symbolic "European AI company taken away" of all, and Britain's largest academic asset of the decade. Ten years later, DeepMind's AlphaFold won Hassabis and John Jumper the 2024 Nobel Prize in Chemistry. The UK AI Safety Institute (AISI), founded in London in 2023, is the world's first sovereign-state body devoted to evaluating frontier-model risk, with a "pro-innovation" line drawn in subtle contrast to Brussels.
II. The Continental Academic Tradition
If Britain was AI's source of ideas, continental Europe was its source of methods.
Switzerland's IDSIA (the Dalle Molle Institute for Artificial Intelligence, founded 1988 in Lugano) is the birthplace of LSTM. Jürgen Schmidhuber (1963–) and his doctoral student Sepp Hochreiter introduced long short-term memory in 1997 — an architecture that powered the speech and translation cores of Google Translate, Apple Siri, and Amazon Alexa for an entire generation of products. Schmidhuber, fiercely independent, has long argued that "almost every key breakthrough of modern deep learning was published earlier in my own work"; his priority disputes with Geoffrey Hinton (1947–), Yann LeCun (1960–), and Yoshua Bengio (1964–) are a famous recurring affair in academic AI.
ETH Zürich and EPFL (the Swiss Federal Institutes of Technology in Zürich and Lausanne) are Europe's twin towers of robotics and computer vision. ETH's Robotics Institute trained the ANYmal four-legged robot and a string of drone teams; EPFL has deep roots in brain-computer interfaces and differentiable programming. Germany's Max Planck Institute for Intelligent Systems (MPI-IS, in Tübingen and Stuttgart) is Europe's flagship for foundational research, with strengths in causal inference, probabilistic inference, and motor control.
The French mathematical tradition has given AI a distinctive theoretical texture. INRIA (the French National Institute for Research in Digital Science and Technology), founded in 1967, has trained generations of AI researchers grounded in applied mathematics; the École Normale Supérieure (ENS) and the École Polytechnique contributed Yann LeCun (Paris-born), Léon Bottou, Stéphane Mallat, and others. Today, the founding team of Mistral AI is almost entirely a product of this French math-engineering pipeline.
CERN (the European Organization for Nuclear Research) is not directly an AI lab but is one of the world's longest-standing handlers of large-scale scientific data. The LHC's event triggering and reconstruction made it an early proving ground for machine-learning methods in physics, training a generation of engineers fluent in massive distributed computation — many of whom went on to DeepMind, Mistral, and various European cloud providers.
III. European Startups: Sprinting in the Cracks Between Giants
Europe's startup ecosystem has long been labeled "small and scattered," but the era of large models has produced a notable cohort.
Mistral AI, in Paris, is the most representative. Founded in April 2023 by former Meta and Google DeepMind researchers Arthur Mensch, Guillaume Lample, and Timothée Lacroix, it closed a 113-million-dollar seed round on a slide deck in just four weeks — among the largest European seed rounds ever. Its Mistral 7B, Mixtral 8x7B, Mistral Large, and Mistral Medium series became the standard-bearer of "European sovereign AI" through 2024–2025 with an "open source plus efficiency" line. In 2024, Mistral struck strategic partnerships with the French government and Microsoft separately, with valuation crossing six billion euros in 2025.
London's Stability AI is the company behind Stable Diffusion (open-sourced August 2022), placing high-quality text-to-image models into ordinary hands for the first time and opening the open-source AI-art wave. From 2024, Stability faced a CEO departure, a major copyright lawsuit (the Getty Images case), and a cash crisis; in 2025 a new board including James Cameron took over.
Heidelberg's Aleph Alpha (founded 2019) was the earliest European bet on "sovereign AI," advocating data localization, auditability, and explainability. After raising five hundred million dollars in a 2023 Series B, it pivoted in 2024, in the face of open-source large-model pressure, toward enterprise RAG and government services.
London's Synthesia (founded 2017) represents the other side of European generative AI — AI digital-human video generation. Its share in enterprise training and multilingual marketing video far exceeds that of North American peers, and its valuation crossed 2.6 billion dollars in 2024. Finland's Silo AI (acquired by AMD in 2024 for 665 million dollars), Italy's iGenius, and Sweden's Sana Labs together form Europe's second tier of AI companies.
IV. The EU AI Act: The Brussels Effect Returns
On August 1, 2024, the European Union's Artificial Intelligence Act (AI Act) entered into force — the world's first comprehensive AI legislation.
It classifies AI systems into four risk levels: unacceptable risk (such as social scoring and subliminal manipulation — banned outright); high risk (biometrics, critical infrastructure, education, employment, and others — subject to compliance assessment); limited risk (chatbots — must disclose AI identity); and minimal risk (spam filters and the like — no mandatory requirements). For general-purpose AI (GPAI), the Act adds three obligations — technical documentation, copyright compliance, and systemic-risk assessment; models trained with more than 10²⁵ FLOPs of compute are presumed to be "systemic-risk" models requiring extra reporting and adversarial testing.
The Act phases in: prohibitions take effect from February 2025; GPAI obligations from August 2025; high-risk provisions from August 2026 in full. Companion measures — the AI Liability Directive and revised Product Liability Directive — shift evidentiary burdens to platforms.
The Brussels Effect surfaces again. To comply with EU rules, OpenAI, Anthropic, and Google have had to retain EU-specific behaviors in their global builds; Chinese vendors entering the EU must meet the same standards. Rules became Europe's most powerful non-technological weapon.
V. European Supercomputing and Sovereign AI Infrastructure
Realizing that compute is the new oil, Europe in 2018 launched EuroHPC JU (the European High Performance Computing Joint Undertaking), funded jointly by the EU and member states, to build a series of top-tier supercomputers.
France's Jean Zay (at the IDRIS data center south of Paris, online since 2019, expanded in 2024 to 125.9 PFlop/s) is the main training site for French academic large models, where the BLOOM multilingual model was trained. Finland's LUMI (in Kajaani, online 2022, HPL 379.7 PFlop/s) was briefly Europe's most powerful supercomputer, sustained 100 percent on hydropower and Nordic cold. Italy's Leonardo (in Bologna, online November 2022, 238 PFlop/s), operated by Cineca, is one of the training homes of Mistral and EuroLLM. Spain's Barcelona Supercomputing Center (BSC) hosts MareNostrum 5 (online January 2024, HPL 175 PFlop/s), known for sustainable wind- and solar-based power.
From 2024, the EU also launched the AI Factories initiative, deploying "AI factories" across thirteen member states to directly link supercomputers to startups and research institutions. Mistral, Aleph Alpha, and Silo AI were early beneficiaries. This is Europe's first attempt at the compute layer to break unilateral dependence on American GPU-cloud providers.
VI. The European School's Quiet Output
Europe is almost absent from the leaderboard of headline AI companies, yet it has long run a surplus in "key persons exported."
Yann LeCun was born in Paris and earned his doctorate from Université Pierre-et-Marie-Curie; he is a representative of the French AI school and now Chief AI Scientist at Meta and a professor at NYU. Geoffrey Hinton studied at Cambridge and earned his PhD at Edinburgh; he moved to North America only in 1987. Yoshua Bengio completed his postdoc in Paris before moving to Montreal. The youthful education of all three deep-learning godfathers was completed in Europe.
ELLIS Society (the European Laboratory for Learning and Intelligent Systems), founded in London in 2018 by Bernhard Schölkopf, Nuria Oliver, Sepp Hochreiter, and others, was conceived as an academic alliance to keep European AI talent in Europe. It has not stopped top PhDs from going to North America, but it has at least organized Europe's remaining research strength into a visible network. Schmidhuber's IDSIA, ETH's robotics department, EPFL's BCI lab, the Tübingen AI Center, Aalto's ML group — these scattered lights, taken together, remain a substrate of European AI that cannot be ignored.
VII. Europe's Predicament and Open Questions
Europe holds some of the world's smartest AI minds, and yet, time and again, it has handed off its largest companies to others. DeepMind to Google; Mobileye to Intel (2017, fifteen-point-three billion dollars); ARM to SoftBank (2016) and nearly to NVIDIA (2020–2022); Synthesia and Stability under continual North American capital pressure.
The reasons are structural: European VC in aggregate is less than half the U.S. total, and public equity markets lack a deep tech-stock pool; GDPR places severe constraints on data training, putting Europe behind on data scale from the first day; multilingual, multi-jurisdictional fragmentation forces startups to internationalize from day one, eroding home moats; and talent flow between academia and industry is far less fluid than in the U.S.
A deeper paradox lies inside regulation itself. The EU is on the one hand the world's pacesetter in AI governance, and on the other it is criticized by its own founders as "rule-making that locks itself out." Mistral's Mensch told the Financial Times in 2024 that the AI Act's compliance costs "do not put European founders on the same starting line as American or Chinese players."
Europe's open question is whether it can find a path of its own — neither replaying Silicon Valley's "winner takes all," nor reducing itself to the dull binary of "American software plus European regulation"; making rules a guardrail for innovation rather than a stumbling block. There is no standard answer. But this is the true main thread of Europe's AI story in 2026.
Historian's Note
I have observed eighty years of European AI with mingled joy and sorrow. The joy: the source of ideas has not run dry — Turing set the will, Schmidhuber kept the school, and LeCun, Hinton, and Bengio all grew tall on European soil. The sorrow: capital and market do not match the talent; DeepMind sold, Mobileye sold, ARM passed hand to hand. To stitch the bridal gown for others has become a European fate. Yet Europe has its own gift: governance by rule, technology disciplined by law. The AI Act, criticized at home, has nonetheless made every global model company write into product documents, "modified to meet EU requirements" — this is the Brussels Effect. Sima Qian wrote, "with bronze for a mirror one straightens one's robes"; for America and China, Europe is precisely a mirror of rules. But a mirror reflects others; it does not stand alone. Mistral's rise is fresh vigor; Jean Zay and LUMI are catching up; sovereign AI is becoming consensus. If posterity asks how Europe might rise again in AI, the answer cannot lie with one company or one law. It lies in rebuilding the triangular circulation of "ideas — capital — market." Otherwise, however refined the rules, they will be no more than an audience's whistle.
Eyewitness Accounts
Call for contributions
If you have participated in AI research, founded companies, or shaped policy in Europe, please contribute on GitHub.
References
- Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433-460.
- Lighthill, J. (1973). Artificial Intelligence: A General Survey. SRC Report.
- Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735-1780.
- European Parliament & Council. (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act).
- UK AI Safety Institute. (2023). Bletchley Declaration. AI Safety Summit, Bletchley Park.
- EuroHPC JU. (2024). EuroHPC Annual Report 2024. Luxembourg.
- Mensch, A. et al. (2023). Mistral 7B Technical Report. arXiv:2310.06825.
- ELLIS Society. (2024). European Laboratory for Learning and Intelligent Systems Annual Report.