The Brief

Yann LeCun, the Turing Award-winning AI researcher who left Meta last November, has raised $1.03 billion in a seed round for his Paris-based startup Advanced Machine Intelligence Labs — Europe’s largest seed deal ever, valuing the company at $3.5 billion before the investment. AMI Labs will develop world models based on LeCun’s Joint Embedding Predictive Architecture as a declared alternative to the large language models that dominate the current AI industry.

The Report

Advanced Machine Intelligence Labs, the AI startup co-founded by Yann LeCun and CEO Alexandre LeBrun, has closed a $1.03 billion seed round — approximately €890 million — at a pre-money valuation of $3.5 billion. The round, which doubled its original €500 million target due to investor demand, represents the largest seed financing in European startup history and the second-largest globally, behind Thinking Machines Lab’s $2 billion raise.

The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. Strategic investors include Nvidia, Samsung, Toyota Ventures, Temasek, and Bpifrance, France’s public investment bank. Individual backers include Jeff Bezos, Eric Schmidt, Mark Cuban, Tim Berners-Lee, and Xavier Niel. The investor base splits roughly equally across American, European, and Asian capital.

AMI Labs will build what it calls world models — AI systems that construct abstract representations of physical environments rather than predicting the next word in a text sequence. The core technology is derived from LeCun’s Joint Embedding Predictive Architecture, which he proposed in 2022 and developed through successive implementations at Meta’s FAIR lab. JEPA works in abstract embedding space, predicting high-level representations of sensory inputs rather than attempting pixel-level or token-level reconstruction. The company has identified robotics, healthcare, industrial process control, and autonomous systems as target verticals.

LeCun, who spent twelve years at Meta overseeing the development of the Llama model series and the PyTorch framework, departed in November 2025 following Meta’s reorganisation of its AI division under Meta Superintelligence Labs. The new structure, led by 28-year-old Alexandr Wang following Meta’s $14.3 billion acquisition of a 49% stake in Scale AI, shifted the company’s emphasis from exploratory research toward commercial deployment — a direction LeCun had publicly resisted. “Large language models are a statistical illusion,” LeCun said. “Impressive, yes. Intelligent, no.”

The founding team draws heavily from Meta-FAIR: Michael Rabbat, formerly a director of research science there, serves as VP of World Models; Pascale Fung, a former senior director of AI research at Meta, is Chief Research and Innovation Officer; and Saining Xie, previously at Google DeepMind, is Chief Science Officer. Laurent Solly, Meta’s former VP for Europe, is COO.

AMI has designated Nabla, LeBrun’s medical AI startup serving 85,000 physicians, as its first strategic partner. The company currently employs around twelve researchers and plans to grow to thirty to fifty within six months. LeBrun has indicated the first year will be devoted entirely to research, with commercially viable products several years away. “My prediction is that ‘world models’ will be the next buzzword,” LeBrun said. “In six months, every company will call itself a world model to raise funding.”

French President Emmanuel Macron described the announcement as representing “the France of researchers, builders and the bold.” AMI Labs operates from Paris with planned offices in New York, Montreal, and Singapore.

The Angle

A billion dollars for twelve researchers, no product, and a thesis that the entire current AI stack is built on the wrong foundation. The investment is not really in AMI Labs. It is in the possibility that LeCun is right — that autoregressive language models are a developmental cul-de-sac sophisticated enough to be mistaken for the main road.

What makes this interesting is not the money or the valuation. It is what the bet reveals about the state of conviction inside the industry. The investors writing these cheques are not contrarians betting against AI. Many of them — Nvidia, Bezos, Schmidt — are the same names backing the LLM paradigm at vastly larger scale. This is hedge capital. It is the market acknowledging, through revealed preference rather than public statement, that the question of whether token prediction leads to genuine machine understanding remains genuinely open. A billion dollars is what that uncertainty looks like when it is priced.

LeBrun’s prediction about “world models” becoming the next fundraising buzzword is the most honest sentence in the entire announcement. He is describing the cycle that will dilute the term before the research has a chance to vindicate or falsify it — the same process that turned “AGI” from a specific technical claim into a marketing category. The question for AMI is whether JEPA-based systems can demonstrate capabilities that LLMs structurally cannot before the commercial gravity of the current stack becomes its own justification. LeCun has the scientific credibility. What he is now racing is not a rival architecture but the compounding network effects of a paradigm that is already deployed, already profitable, and already defining what customers expect AI to be.

The European sovereignty angle — Macron’s congratulations, the Bpifrance participation, the “neither American nor Chinese” positioning — is real but secondary. Sovereign AI ambitions do not survive contact with the market unless the underlying technology works. What AMI has bought with this round is time: a few years of research runway insulated from quarterly revenue expectations, in a field where nearly every competitor is already shipping. Whether that insulation produces a breakthrough or an expensive proof that LLMs were closer to sufficient than their critics believed is the only question the money cannot answer.

The most telling detail is not in the funding. It is in the team composition. Five of six founders came from Meta-FAIR — the lab LeCun built, the lab Meta restructured out from under him. A billion-dollar seed round seeded by the dispersal of the largest AI research operation outside the United States. Meta chose to reorganise its research division around commercial urgency. The researchers chose to leave and bet that the work Meta deprioritised was the work that mattered. Both sides will eventually be graded on the same curve.