France Had the Best AI Research in the World in 1990. Then This Happened.

In 1987, Yann LeCun was working in Paris. He had just completed his PhD at Pierre and Marie Curie University, working under researchers at INRIA — France’s national computing research institute — on what would become some of the foundational work in convolutional neural networks. France had, at that moment, a genuine concentration of world-class AI and computer science talent. INRIA was excellent. The grandes écoles produced rigorously trained mathematicians and engineers. The academic lineage ran deep.

By 1991, LeCun was at Bell Labs in New Jersey. By 2003, he was a professor at NYU. By 2013, he was running Facebook’s AI Research lab. He is now one of the most cited researchers in computer science history and a Nobel Prize winner (shared, in 2024, with Hinton and Bengio). He did not do most of this work in France.

LeCun is not an isolated example. He is the most famous instance of a pattern that repeated across generations of French technical talent. The researcher who builds their foundational skills inside the French system, then discovers that the American system offers faster publication cycles, better-funded labs, more direct paths to commercialization, and significantly more money — and leaves. The exit is rational from the individual’s perspective every single time. The accumulation of individually rational exits produces a catastrophic national outcome.

The question of why is not mysterious. The mechanisms are well-documented and were visible in real time to anyone who wanted to look. France decided not to.

The grandes écoles are the best place to start, because they explain both why France had strong technical talent and why that talent consistently left.

The French elite education system — École Polytechnique, École Normale Supérieure, ENSTA, and a handful of others — is, by the metrics that matter for producing strong researchers and engineers, excellent. Admission is brutal. The mathematics training is rigorous in a way that most American universities don’t attempt at the undergraduate level. The intellectual culture values depth and formal precision. These are real advantages.

The problem is structural. The grandes écoles are deeply embedded in a system of public sector prestige. The most coveted career path for a Polytechnicien was not, historically, to start a company or run a research lab — it was to enter the grands corps de l’État, the elite administrative bodies that run ministries, infrastructure, and state-owned enterprises. This was the status game. Industry was secondary. Entrepreneurship was borderline suspicious.

A French mathematician of exceptional ability in 1985 faced a set of incentives that pointed toward ENA, toward the Inspection des Finances, toward prestige administrative positions that offered security, status, and substantial lifetime income. The same person at MIT or Stanford faced incentives that pointed toward research labs, startups, and commercialization. The incentive difference compounds over decades.

INRIA, to its credit, was doing genuinely important work. The Prolog programming language was developed partly by researchers at the University of Aix-Marseille and had influence on AI research globally. French mathematicians and computer scientists contributed meaningfully to algorithms, formal verification, and theoretical computer science in the 1970s and 1980s. This wasn’t a case of a country that never had the talent. It’s a case of a country that had the talent and then organized itself so that the talent left.

The academic bureaucracy was a separate problem. French university research in the 1980s and 1990s was heavily centralized, with career advancement tied to a national qualification system — the “qualification nationale” — that required habilitation through committees that moved slowly and tended toward conservative academic judgment. Researchers who wanted to work on something unconventional, or move quickly, or collaborate closely with industry, found the system actively resistant. The system wasn’t hostile to research — it was hostile to fast-moving, applied, commercializable research. Which is precisely what became important.

The tenure clock in French academia was also significantly longer than in American universities, meaning that a researcher who committed to France was committing to a decade or more of institutional navigation before achieving the job security that would allow genuinely ambitious work. An equivalent American researcher, passing through a PhD and postdoc in five to seven years and landing a tenure-track position at a research university, had faster access to the resources and independence that serious work required. Speed compounded into outcome differences.

Compare this to the Bell Labs model that produced so much of American computing progress in the same period. Bell Labs had, by design, a relatively flat structure for researchers, generous funding, and minimal pressure on publication timelines. LeCun at Bell Labs was free to work on neural networks for handwriting recognition — work that led directly to the system that processed a significant fraction of US bank checks in the 1990s. He could not have done that work at the same pace inside INRIA. Bell Labs, in the same period, produced the transistor, Unix, C, the laser, information theory, and cellular telephony. It was the most productive research institution in computing history, partly because AT&T’s monopoly profits funded it generously and partly because it was structured to give smart people hard problems and then leave them alone.

The venture capital gap is often cited and genuinely mattered, but it’s slightly downstream of the cultural problems. You don’t build a venture ecosystem in isolation — it requires investors willing to back technical founders, founders willing to take risk, legal structures that allow quick equity grants, and universities that accept commercialization as a legitimate activity. France had none of these in sufficient quantity. The Loi Allègre (1999) was France’s belated attempt to allow researchers to commercialize their work; that it needed to be enacted as specific legislation in 1999 — when Silicon Valley had been running for three decades — tells you something about how far behind the cultural baseline was.

The English language is worth naming as a factor too, not because French researchers couldn’t work in English — they could, and did — but because the gravity of American research institutions in the 1990s was enormous and growing. The best AI conferences were predominantly in the US. The best AI labs were predominantly in English-speaking countries. A French researcher who wanted to be at the center of the field had to leave, and most did. This isn’t a solvable problem through policy; it’s a structural feature of how the Anglo-American research ecosystem accumulated advantages.

What France is now attempting is a partial reversal. Macron’s 2018 AI strategy, developed after the Villani report, committed €1.5 billion in AI investment and restructured INRIA to be more commercially oriented. Mistral AI, founded in 2023 by former DeepMind and Meta researchers (mostly French), raised €385 million in its Series A — the largest seed round in European history — and has produced genuinely competitive open-weight language models. Mistral 7B and Mixtral 8x7B are not afterthoughts; they’re being run in production by serious organizations who evaluated them seriously.

The question is whether this represents a sustainable reversal or a capable-but-late catch-up play.

Honest answer: probably the latter. The structural advantages that accrued to US AI research over thirty years — the dense network of labs, talent, capital, and infrastructure; the compute supply chain centered on Nvidia and TSMC; the depth of the consumer and enterprise software markets that AI needs to be useful — cannot be replicated by French policy in a decade. Mistral is doing impressive work inside significant constraints. The constraints include compute access (GPUs are overwhelmingly allocated to US hyperscalers), talent competition (Meta and Google can pay French researchers more than French startups can), and the fact that the most valuable applications of AI are, for now, built on American platforms.

France could have been a very different story. The ingredients were there in 1985. The country chose, through a hundred small institutional decisions, to build an education system optimized for the state rather than for commercial research, an academic system insulated from the market, and a prestige hierarchy that pointed talent away from entrepreneurship. Those decisions compounded.

There’s a lesson here for every country currently developing a “national AI strategy,” and most of them are making the same mistakes France made in the 1980s. They’re investing in research institutions without reforming the incentive structures around those institutions. They’re announcing compute investments without asking whether their best researchers have reasons to stay. They’re creating AI funding programs without creating the equity, tax, and immigration structures that let startups compete for talent.

The UK published its AI strategy in 2021 with significant fanfare and investment commitments. It has watched a steady outflow of DeepMind talent to Google’s US offices, to OpenAI, to Anthropic. Germany has launched multiple AI research institutes and has seen many of its best ML researchers take positions in California. Canada, which has genuinely excellent AI research clusters in Montreal and Toronto built around Bengio and Hinton, watched those researchers launch companies (Mila spinoffs, Vector Institute spinoffs) that immediately faced capital and talent constraints that their American counterparts did not. The pattern is consistent. Announcing strategy is not the same as creating conditions for retention.

France’s advantage in 1985 was real. It evaporated not because American researchers were smarter, but because American research institutions were structured better for the specific kind of fast-moving, applied, capital-intensive work that AI development became. The work followed the structure.

Mistral can build excellent models. Whether France can build a sustainable AI ecosystem rather than a well-funded outpost is a different question. The answer depends on structural changes — in academia, in capital markets, in immigration, in the prestige hierarchy — that a startup’s success, however impressive, does not by itself produce.

One thing France has that the US doesn’t, and that the Mistral founders understand: a genuine motivation to build AI that isn’t controlled by American companies. European enterprises increasingly want an alternative to sending their proprietary data through OpenAI and Google’s APIs. That regulatory and data-sovereignty pressure is a real market force, not just nationalist sentiment. If Mistral can hold together its team and access enough compute to stay competitive, the demand is there. Whether the French ecosystem can sustain that team, or whether the individual-rational-exit pattern continues and the founders eventually take American capital on American terms, is the open question.