LeCun Raises $1B Seed to Build AI Beyond LLMs

Yann LeCun, the Turing Prize-winning researcher who spent a decade running Meta's AI research, has closed a $1.03 billion seed round for his new company, Advanced Machine Intelligence Labs. The raise values AMI at $3.5 billion pre-money and stands as the largest seed round ever completed by a European company.

The money comes with a thesis that most of the AI industry doesn't share: large language models are a dead end for the hardest problems in artificial intelligence.

TL;DR

AMI Labs raised $1.03 billion (€890 million) at a $3.5 billion pre-money valuation
Co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with Nvidia, Samsung, Temasek, and Toyota Ventures also participating
The company is building "world models" using JEPA architecture - AI that learns from video and physical-world data instead of text
CEO Alexandre LeBrun previously founded Nabla, which becomes AMI's first healthcare partner with access to 150+ U.S. health systems
Offices in Paris (HQ), New York, Montreal, and Singapore; no product expected for at least a year

The Money

Five firms co-led the round: Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. The strategic investor list reads like a hardware supply chain - Nvidia, Samsung, Toyota Ventures, and Singapore's sovereign fund Temasek all participated. French industrial players including Groupe Industriel Marcel Dassault and Publicis Groupe joined with SBVA from South Korea.

The individual backers are equally standout. Jeff Bezos invested through Bezos Expeditions. Eric Schmidt, Mark Cuban, Jim Breyer, and Tim and Rosemary Berners-Lee all wrote checks. LeCun initially targeted €500 million. Demand pushed the final number to €890 million.

This isn't the first billion-dollar seed to come out of AI research. David Silver's Ineffable Intelligence raised $1 billion in February, also betting against the LLM paradigm. But AMI's $1.03 billion edges it out for the European record. Combined with Nscale's $2 billion Series C last week, European AI funding is having its most aggressive quarter on record.

Who's Running It

LeCun serves as executive chairman while maintaining his professorship at New York University. Day-to-day operations fall to CEO Alexandre LeBrun, who previously founded and ran medical AI company Nabla.

The leadership team is stacked with Meta alumni. Michael Rabbat, formerly Meta's director of research science, serves as VP of World Models. Laurent Solly, Meta's former VP for Europe, is COO. Pascale Fung, a former Meta senior director, holds the Chief Research and Innovation Officer title. The outlier is Chief Science Officer Saining Xie, who came from Google DeepMind.

The Bet Against LLMs

LeCun has been vocal for years about what he sees as the limits of large language models. His position hasn't softened. "The breakthroughs are not going to come from scaling up LLMs," he said at Davos in January. He's called the current generation of text-based AI "a statistical illusion" - systems that can arrange words convincingly without understanding the physical reality those words describe.

A visualization of neural network architecture and AI model connections AMI Labs is building world models that learn from video and sensor data rather than text, using LeCun's JEPA architecture developed during his decade at Meta. Source: unsplash.com

AMI's alternative is what the company calls "world models" - AI systems built on LeCun's Joint Embedding Predictive Architecture (JEPA). Where an LLM predicts the next word in a sequence, JEPA predicts the next abstract representation in a video or sensor stream. It doesn't create pixels. It learns to anticipate how physical environments evolve over time.

How JEPA Works

The architecture has two core components. An encoder takes raw video input and produces embeddings that capture the semantic state of the observed world. A predictor then takes those embeddings and forecasts what comes next - not at the pixel level, but in representation space. This forces the model to learn causal structure rather than surface patterns.

Meta published V-JEPA 2 before LeCun's departure, showing zero-shot robot control in new environments. AMI plans to build on that foundation with its own model, called AMI Video, targeting robotics, manufacturing, and wearable devices.

What It Does Not Tell You

The three-to-five year timeline LeCun has outlined for "fairly universal intelligent systems" is ambitious, and the company's first year will be pure R&D with no commercial product. World models have shown promise in constrained robotics tasks, but nothing approaching the generality that LLMs have reached for language. LeCun is basically asking investors to fund a research program, not a product company - at least not yet.

The $3.5 billion valuation for a pre-revenue startup with no launched product also raises questions about whether the AI funding environment has detached from conventional metrics completely.

The Healthcare Play

AMI's first disclosed partnership is with Nabla, the medical AI company LeBrun founded before taking the CEO role at AMI. The connection isn't subtle - LeBrun now serves as Chief AI Scientist and Chairman at Nabla while running AMI.

The Eiffel Tower and Paris skyline, where AMI Labs has established its headquarters AMI Labs is headquartered in Paris with additional offices planned in New York, Montreal, and Singapore. Source: pexels.com

Nabla currently serves more than 150 health systems and provider groups across the U.S., supports 35+ languages, and has raised $120 million in its own funding. The company's ambient AI assistant creates clinical notes in under five seconds and integrates with all major electronic health record systems.

Under the partnership, Nabla gets exclusive early access to AMI's world model technology. The stated goal is to bring "FDA-certifiable agentic AI systems to healthcare" - autonomous agents capable of deterministic, auditable decision-making rather than the probabilistic outputs that current LLM-based clinical tools produce.

"Healthcare AI is entering a new era, one where reliability, determinism, and simulation matter," LeBrun said in Nabla's partnership announcement.

The pitch is compelling on paper. Clinical AI built on world models could theoretically simulate patient outcomes before recommending treatment paths, offering the kind of causal reasoning that text-based models can't provide. But the gap between that vision and a working, FDA-approved product remains vast.

A European Anchor

AMI is headquartered in Paris, with research hubs planned in New York, Montreal, and Singapore. The decision to base the company in France rather than Silicon Valley is deliberate. LeCun holds dual French-American citizenship, and France has been aggressively positioning itself as Europe's AI capital.

"AMI Labs could be the first European company to reach the scale of the GAFAM companies," said Pierre-Éric Leibovici of Daphni, one of the French investors in the round.

That's a bold claim for a company that hasn't shipped anything yet. But the investor roster - spanning five continents, blending sovereign wealth with strategic chipmakers and individual tech billionaires - suggests real conviction behind the contrarian thesis. Whether LeCun can convert a decade of theoretical work into commercial AI systems that beat the LLM paradigm is the $1.03 billion question.

The AI industry has spent the last four years scaling language models to the point where they can pass bar exams and write production code. LeCun is betting that none of that matters for the problems he cares about - robots that can navigate a kitchen, wearables that can interpret a doctor's gestures, manufacturing systems that can predict failure before it happens. He now has a billion dollars and a team of the researchers who built Meta's AI division to prove it.

Sources: