Summary: Yann LeCun’s AMI Labs closed a $1.03 billion seed round at a $3.5 billion pre-money valuation, TechCrunch reported on March 9, 2026. The company will pursue world models and JEPA rather than LLMs, with Nvidia, Samsung, Toyota Ventures, and Temasek among strategic investors. It has no near-term revenue plan, and commercialization could be years away.
AMI Labs — the company Yann LeCun co-founded after leaving Meta — closed a $1.03 billion seed round (roughly €890 million) at a $3.5 billion pre-money valuation, TechCrunch reported on March 9, 2026. No seed round in European history has been larger. The capital exists for one purpose: to show that world models, not large language models, are the better route to machine intelligence.
For developers and product teams, the number is not the signal — the roster is. Nvidia, Samsung, Toyota Ventures, and Temasek arrived as strategic investors, meaning chips, manufacturing, and capital markets all placed their bets on a non-LLM route at the same time. A Turing Award winner leaving Big Tech on a fundamental-research agenda usually collects a polite, modest check; this one ran to ten digits.
How a Seed Round Reached $1.03 Billion — a European Record
Co-leading the round: Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. Joining as strategic backers: Nvidia, Samsung, Sea, Temasek, and Toyota Ventures. The French column runs long — Association Familiale Mulliez, Groupe Industriel Marcel Dassault, and Publicis Groupe, together with funds such as Bpifrance Digital Venture, Artémis, Aglaé Lab, and SBVA. On the angel list: Eric Schmidt, Mark Cuban, Jim Breyer, Xavier Niel, and Tim and Rosemary Berners-Lee.
The way the target grew tells its own story. In December 2025, rumor had AMI Labs aiming at just €500 million; what actually landed came close to double. Put that next to Fei-Fei Li’s World Labs, which took in $1 billion in February, and SpAItial, which closed a $13 million seed, and the pattern is clear: money is concentrating on the world-model race fast. Layer in Nscale’s $2 billion Series C — European AI fundraising is running hot across the board — and “European record” has been rewritten several times over just a few weeks.
JEPA and World Models: A Route Around the LLM
At the technical center of AMI Labs sits JEPA, the Joint Embedding Predictive Architecture LeCun put forward in 2022. The idea in one line: AI should learn from reality, not from language alone. The model builds its understanding of the physical world by predicting abstract representations of what it observes; intelligence is not anchored to predicting the next token.
LeCun and CEO Alexandre LeBrun have both argued in public that LLMs carry a defect that refuses to go away: hallucination. In a healthcare setting, a single hallucination can cost a life. Their reading: application-layer patching will not fix this — the disagreement sits in the route’s assumptions, because language data does not stretch to cover the common sense of the physical world.
The cold water belongs in the same breath. AMI Labs puts fundamental research first: for now there is no revenue plan at all, and commercialization could be years out. LeBrun’s own forecast pulls no punches — “world models” is headed for buzzword status, and “in six months, every company will call itself a world model to raise funding.” What separates the real ones, in the end, is an architecture you can verify and experiments you can reproduce.
The People, and Four Cities on the Map
The org chart reads like a manifesto. LeCun takes the chairman seat and keeps teaching at NYU. CEO Alexandre LeBrun chairs digital-health company Nabla and came to this from Meta’s FAIR. COO Laurent Solly previously held the VP-for-Europe post at Meta. The research side rests on chief science officer Saining Xie, chief research and innovation officer Pascale Fung, and VP of world models Michael Rabbat.
Operations span four cities — headquarters in Paris, plus New York, Montreal, and Singapore, the last picked to sit near Asian clients and talent. LeBrun is just as blunt about spending: the company runs on two cost centers, compute and talent, and on hiring it would rather wait than settle.
Healthcare Comes First: Nabla and Open Research
First on the partner list is Nabla, the digital-health company LeBrun founded. Opening in medicine follows straight from the technical thesis: clinical settings are where hallucinations cost the most, which makes healthcare the sharpest test of whether world models really are more dependable than LLMs. No revenue is planned in the near term, but the team still intends to reach prospective customers early, gathering real-world data and evaluation feedback.
The second commitment is openness. Papers will keep coming, and much of the code will be open-sourced — in LeBrun’s words, “we think things move faster when they’re open.” For the research community, that is this round’s most concrete externality: wherever the route ends up, the code and the papers will be public.
How Developers Should Read This
Three things worth watching. One: keep “world model” out of your pitch deck for now — work through the original JEPA papers and the code AMI releases later, and let that calibration sort the imitators from the rest. Two: if you build for healthcare or robotics, put Nabla’s integration progress on your watchlist; that is where world-model reliability gets its first public litmus test. Three: the open-source commitment means engineering assets will arrive in installments — for teams working on evaluation, simulation, and data pipelines, that is a low-cost entry into a new route. LLMs are not going anywhere, but the “learn from reality” hypothesis finally has enough ammunition for a decade.