Yann LeCun Just Raised $1 Billion to Prove the Entire AI Industry Is Wrong

A Billion Dollars Says LLMs Aren't the Answer

Yann LeCun, the 65-year-old Turing Award winner who spent twelve years building Meta's AI research operation into one of the most respected in the world, just put his money where his mouth has been for years. His new startup, Advanced Machine Intelligence Labs (AMI), announced a $1.03 billion seed round on March 10, valuing the company at $3.5 billion before the money even hit the account. That makes it the largest seed round ever raised by a European startup, and one of the biggest bets in AI history on the idea that the entire industry is heading in the wrong direction.

The thesis is simple, even if the execution won't be: large language models are a dead end for achieving truly intelligent AI. LeCun has been saying this publicly for years, and now he's building a company around it. AMI is focused on world models, a fundamentally different approach to AI that learns from reality rather than from text. If he's right, the hundreds of billions being poured into scaling LLMs are solving the wrong problem. If he's wrong, it's going to be one of the most expensive contrarian bets in tech history.

Why LeCun Left Meta

The backstory makes the funding even more significant. In November 2025, LeCun walked into Mark Zuckerberg's office and told him he was leaving. After building Meta's Fundamental AI Research (FAIR) lab from scratch, he had grown frustrated with the company's direction. The core disagreement was strategic: Zuckerberg wanted to pour resources into scaling LLMs, while LeCun believed the future lay in world models and robotics.

LeCun didn't mince words about what went wrong. He publicly stated that while FAIR was extremely successful on the research side, Meta was far less successful at turning that research into practical technology and products. He was particularly upset that Meta had disbanded the robotics group at FAIR, which he considered a strategic mistake. For a researcher who believes understanding the physical world is the key to real intelligence, watching his employer abandon physical AI research was apparently the last straw.

World Models vs. Large Language Models

So what exactly are world models, and why does LeCun think they matter more than the LLMs everyone else is obsessed with? The core idea is about how AI should learn to understand reality.

Large language models work by predicting the next word in a sequence. They're trained on massive amounts of text and learn statistical patterns in language. They're remarkably good at generating text, writing code, and even reasoning through problems. But LeCun argues they have a fundamental limitation: they don't actually understand the world. They understand language about the world, which is a very different thing.

World models, by contrast, learn abstract representations of how things actually work. Think about how a baby learns about gravity: not by reading about it, but by watching objects fall and building an internal model of the physics. AMI's approach, built on LeCun's Joint Embedding Predictive Architecture (JEPA), does something similar. Instead of predicting the future in pixel-perfect detail, JEPA learns to predict what will happen in an abstract representation space, ignoring the unpredictable surface details and focusing on the underlying rules.

LeCun's pitch is that this is the foundation for common sense, the thing that current AI systems most conspicuously lack. A world model that truly understands how physical reality works could power robots, autonomous vehicles, and industrial systems in ways that text-trained models simply cannot.

The Money Behind the Mission

The investor list reads like a who's who of tech and venture capital. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions (Jeff Bezos's personal investment vehicle). Nvidia and Samsung also participated, along with a roster of individual investors that includes former Google chairman Eric Schmidt, Mark Cuban, and World Wide Web inventor Tim Berners-Lee.

The involvement of Nvidia is particularly interesting. The company that makes the GPUs powering virtually every LLM training run on earth is also betting on the guy who says LLMs won't get us to real AI. That's either hedging or genuine conviction that the world model approach has merit. Either way, it suggests Nvidia sees a future where its hardware powers more than just transformer training.

AMI raised €890 million (approximately $1.03 billion), nearly double the €500 million it originally targeted. When investors are throwing twice what you asked for at an unproven thesis, it tells you something about how seriously the AI research community takes LeCun's track record.

Paris as an AI Capital

AMI is headquartered in Paris, which is itself a statement. LeCun, who was born in Paris and teaches at NYU, chose to plant his flag in Europe rather than Silicon Valley. The company is building teams across four locations: Paris, New York, Montreal, and Singapore.

The choice of Paris fits a broader trend. France has been aggressively positioning itself as a European AI hub, with President Macron personally courting AI companies and researchers. Mistral AI, another Paris-based startup, has already proven that world-class AI research can happen outside of San Francisco. AMI's arrival, with the largest European seed round ever, cements Paris as a genuine contender in the global AI race.

The funding will go primarily toward two things: compute and talent. Building world models requires enormous computational resources, and AMI needs to recruit top researchers who are willing to bet their careers on an approach that most of the industry hasn't embraced yet. The company's chief research and innovation officer, Pascale Fung, is leading hiring efforts across all four locations.

The Contrarian Case

It's worth being honest about the risk here. The entire world model approach is unproven at scale. JEPA has shown promising results in research papers, but no one has demonstrated that it can produce the kind of breakthrough capabilities that LLMs have delivered over the past three years. ChatGPT, Claude, and their competitors are generating billions in revenue right now. World models are generating research papers.

LeCun's critics point out that he has been predicting the demise of LLMs for years while they've continued to get more capable. The gap between "this approach has fundamental limitations" and "this approach will hit a wall soon" is enormous, and so far the scaling laws for LLMs have held up better than the skeptics expected.

But LeCun's track record commands respect. He's one of the three "godfathers of deep learning" who shared the 2018 Turing Award. His work on convolutional neural networks in the 1990s was dismissed by much of the AI community for over a decade before it became the foundation of modern computer vision. He has been right about big bets before, even when it took the rest of the field a long time to catch up.

What to Watch

The next 12 to 18 months will be critical for AMI. The company needs to show that JEPA can produce results that go beyond academic papers and into working systems. Look for early demonstrations in robotics and autonomous systems, the domains where understanding physical reality matters most and where LLMs struggle the hardest.

Watch the hiring announcements closely. If AMI starts pulling top researchers away from OpenAI, Google DeepMind, and Anthropic, it'll signal that the world model thesis is gaining real credibility inside the AI research community, not just among investors. And pay attention to whether other major labs start hedging their own bets by investing more in world model research alongside their LLM work. Nvidia's investment in AMI might be the first sign that the smart money is starting to diversify away from a pure LLM future.