Yann LeCun, the Turing Award-winning scientist who built Meta’s AI research powerhouse, didn’t wait long after leaving the company at the end of 2025 to make his next move. On March 9, his new Paris-based startup, Advanced Machine Intelligence (AMI Labs) as reported by TechCrunch, it had raised $1.03 billion in what is being called Europe’s largest seed round ever, at a $3.5 billion pre-money valuation.
The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with participation from Nvidia, Temasek, Samsung, Sea, Toyota Ventures, SBVA, Daphni, and high-profile angels including Mark Cuban, Eric Schmidt, Xavier Niel, Jim Breyer, and even Tim and Rosemary Berners-Lee. That kind of firepower, just weeks after the company’s quiet launch, signals something bigger than another AI wrapper: a serious bet that the dominant large language model paradigm is hitting a wall.
LeCun has been saying as much for years. In papers, talks, and pointed X threads, he’s argued that today’s systems, built on next-token or next-pixel prediction, are clever pattern matchers but fundamentally lack the reasoning, planning, and common-sense understanding of the physical world needed for real intelligence. In an interview with Reuters around the funding announcement, he put it bluntly: “Current AI approaches based on predicting the next word or pixel will not produce broadly capable intelligent agents by themselves.”
AMI Labs’ answer: “world models.” These are AI systems designed to learn directly from reality, building internal representations of how the world works, maintaining persistent memory, reasoning about cause and effect, and planning actions in complex, uncertain environments. The near-term focus is enterprise customers running high-stakes operations: manufacturers, automakers, aerospace, biomedical, and pharmaceutical companies where mistakes are expensive and hallucinations are unacceptable.
Longer term, the company sees consumer applications, including domestic robots that actually understand physics and safety. There’s even early talk of potential integration with Meta’s Ray-Ban smart glasses a nod to LeCun’s old employer without any confirmed deal.
AMI isn’t starting from zero. Co-founder and CEO Alexandre LeBrun (previously at Meta and Mistral) has already lined up a partnership with French health AI startup Nabla, whose clinical documentation tools could benefit from more grounded, less error-prone reasoning. The company currently has just 12 employees but is moving fast: fundamental research first, then targeted products.
This isn’t the only big check being written for world models. Last month, Fei-Fei Li’s World Labs pulled in another massive round to bring similar ideas into 3D workflows. SpAItial, another European player, raised a hefty seed last year. But AMI’s scale and LeCun’s stature make it the clearest test yet of whether his long-held critique of pure scaling laws will actually pay off commercially.
The timing is telling. The AI industry is pouring hundreds of billions into ever-larger LLMs and agentic layers on top of them, yet real-world deployment remains plagued by brittleness, safety issues, and the inability to handle novel physical situations. Investors pouring into AMI, including chip giant Nvidia and automaker-linked Toyota Ventures, appear to be hedging that bet. They’re not abandoning transformers; they’re buying insurance on what comes next.
LeCun left Meta after more than a decade running FAIR (Facebook AI Research), which he founded in 2013. During his tenure, Meta became one of the most open and productive AI research organizations in the world. His departure was widely seen as the end of an era, and the start of something more independent and contrarian.
In a field increasingly defined by closed models, hyperscale compute, and rapid productization, AMI Labs is doubling down on open-ended science. The company’s website (amilabs.xyz) is sparse, emphasizing “a new breed of AI systems that understand the world, have persistent memory, can reason and plan, and are controllable and safe.” No flashy demo videos. No promises of AGI tomorrow. Just a quiet confidence that the next leap won’t come from bigger language models alone.
CEO Alexandre LeBrun captured the potential ripple effect with a knowing smile in his TechCrunch interview
My prediction is that ‘world models’ will be the next buzzword. In six months, every company will call itself a world model to raise funding
Alexandre LeBrun
Whether $1.03 billion buys enough runway to prove LeCun right remains the open question. History is littered with promising alternative AI architectures that never scaled. But with Nvidia on the cap table, a $3.5 billion valuation from day one, and one of the field’s most respected voices at the helm, the industry now has a very public experiment running in parallel to the LLM arms race.
If world models deliver even a fraction of what LeCun envisions, robots that don’t need constant human correction, industrial systems that plan safely, agents that actually understand physics, the implications stretch far beyond chatbots. Manufacturing, healthcare, transportation, and consumer robotics could all shift. And the current crop of AI unicorns built purely on language might suddenly look a lot more fragile.
For now, the money has spoken. The question is whether reality, the physical kind, will agree.
