Sign in Subscribe
AI

Yann LeCun Raises $1B for AMI Labs World Models

Yann LeCun Raises $1B for AMI Labs World Models

By Bob Carlson

Yann LeCun, the Turing Award-winning AI researcher, has launched Advanced Machine Intelligence Labs (AMI Labs) with an outsized $1.03 billion seed round at a $3.5 billion pre-money valuation, the largest seed funding ever for a European startup. The Paris-based company aims to develop "world models" — AI systems that learn abstract representations of the physical world from sensor data, rather than relying primarily on next-token prediction like today's dominant large language models.

The announcement, made in early March 2026, underscores a growing debate in the AI community about the limitations of scaling language models alone to achieve more robust, reasoning-focused intelligence. LeCun has long argued that human-like understanding emerges from grounding in the physical world, not just linguistic patterns.

LeCun founded AMI Labs after departing Meta in late 2025, where he had served as chief AI scientist and led the Fundamental AI Research (FAIR) lab. He continues as a professor at New York University. The move came after he concluded that pursuing his vision of world models would be faster and more effective outside Meta's consumer-focused structure, even though Meta supported the underlying research. "I told him I can do this faster, cheaper, and better outside of Meta. I can share the cost of development with other companies," LeCun recalled of his conversation with Mark Zuckerberg.

At the heart of AMI Labs' approach is LeCun's Joint Embedding Predictive Architecture (JEPA), which he proposed in 2022. Unlike LLMs that predict the next word or token in a sequence, JEPA learns by creating embeddings of input data — such as images, video, or sensor readings — and predicting representations in an abstract space while ignoring unpredictable fine-grained details. This allows the system to build persistent internal models of how the world works, supporting better planning, reasoning about cause and effect, and interaction with physical environments.

"The idea that you’re going to extend the capabilities of LLMs to the point that they’re going to have human-level intelligence is complete nonsense," LeCun has stated. He acknowledges LLMs excel at tasks like code generation but believes they fall short for the kind of common-sense reasoning needed in robotics, manufacturing, or autonomous systems. A domestic robot, for instance, requires an understanding of physics, persistence, and causality that goes beyond pattern matching in text.

"World models that seek to understand the world, and you can’t do that locked up in a lab. At some point, we need to put the model in a real-world situation with real data and real evaluations," said AMI CEO Alexandre LeBrun. LeBrun, who previously founded Wit.ai (acquired by Facebook) and served as CEO of the digital health startup Nabla, brings both technical and commercial perspective. Nabla is AMI's first partner, targeting healthcare applications where LLM hallucinations could have serious consequences.

The funding round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. A broad group of participants includes Nvidia, Samsung, Sea, Temasek, Toyota Ventures, and notable individuals such as Eric Schmidt, Mark Cuban, Xavier Niel, Jim Breyer, and Tim and Rosemary Berners-Lee. The oversubscribed round gives AMI significant runway for compute resources and talent acquisition. The company maintains offices in Paris, New York, Montreal, and Singapore to tap global expertise.

The team is stacked with former Meta and top research talent. Saining Xie serves as chief science officer, Pascale Fung as chief research and innovation officer, Michael Rabbat as VP of world models, and Laurent Solly (former Meta VP for Europe) as COO. This group reflects LeCun's decades of experience building research organizations that blend fundamental science with practical ambition.

AMI's focus on world models aligns with a small but growing cohort of efforts in this direction. Other players like World Labs (led by Fei-Fei Li) have also secured substantial funding recently. LeBrun predicted that "world models" could become the next buzzword, with many companies rebranding to chase investor interest. Yet he emphasized that AMI's effort is a long-term scientific endeavor rather than a quick path to product and revenue. "It’s not your typical applied AI startup that can release a product in three months," he noted.

Near-term applications target industries dealing with complex physical systems: manufacturing, aerospace, automotive, biomedical, and pharmaceuticals. A world model of an aircraft engine, for example, could help optimize efficiency, reduce emissions, or predict maintenance needs more reliably than current simulation methods. Over time, the technology could extend to consumer robotics and even augment systems like Meta's smart glasses, with which LeCun said discussions are ongoing.

The company plans to publish papers and open-source significant amounts of code. "We think things move faster when they’re open, and it’s in our best interest to build a community and a research ecosystem around us," LeBrun explained. This openness contrasts with the more closed approaches of some leading AI labs and reflects LeCun's view that powerful AI should not be controlled by any single entity. He has argued that societal decisions about AI use should come through democratic processes rather than individual executives.

This massive seed round at such an early stage — the company is only months old with a small initial team — highlights investor enthusiasm for alternatives to pure LLM scaling. It also reflects confidence in LeCun's track record. As a pioneer in convolutional neural networks and a key figure in modern deep learning, his critique of the current trajectory carries weight. The bet is that grounding AI in physical understanding will unlock capabilities in planning, long-term reasoning, and safe interaction that remain elusive today.

Challenges remain substantial. Scaling JEPA-like architectures to handle the complexity of the real world will require enormous computational resources and innovative training methods. Turning abstract world models into practical, controllable systems that can be deployed reliably in high-stakes environments will take years, as even proponents acknowledge. Questions persist around evaluation metrics — how do you rigorously test an AI's "understanding" of the world? — and around integration with existing robotics hardware.

Yet the timing feels significant. As the limitations of purely statistical language modeling become more apparent in real-world deployments, interest in hybrid or alternative architectures is growing. AMI Labs positions itself not as an immediate competitor to ChatGPT-style products but as infrastructure for the next layer of intelligence — systems that can reason about the physical universe in ways current models cannot.

LeCun envisions eventually building toward a "universal world model" that could serve as a foundation for generally intelligent systems across domains. In the nearer term, the company will focus on targeted deployments with industrial partners to gather real-world data and refine its approaches.

For the broader AI industry, AMI's launch and funding signal that the post-LLM era may already be taking shape. While scaling laws have driven remarkable progress, many researchers believe additional principles — embodiment, prediction in representation space, hierarchical planning — will be necessary for the next leaps. LeCun's willingness to step outside a comfortable role at one of the world's largest tech companies to pursue this vision independently adds credibility to the bet.

AMI Labs is still in its earliest days. Its success will ultimately depend on whether its world models can deliver measurable advances in reasoning and real-world performance that justify the lofty valuation and expectations. But with this caliber of team, investor backing, and intellectual foundation, it has positioned itself as one of the most closely watched experiments in AI today.

The announcement also reinforces Europe's potential in the global AI race. Despite the dominance of U.S. hyperscalers, a Paris-based lab with this level of ambition and capital demonstrates that talent and bold ideas can still attract serious funding on the continent.

As LeCun and his team begin hiring and publishing their initial work, the AI community will be watching closely. The question is no longer whether world models matter, but whether this particular effort can turn long-standing theory into practical, transformative technology. If successful, it could reshape not just how we build AI, but how those systems ultimately interact with and improve the physical world around us.

Sources

(Word count: 982)

Read next