Yann LeCun’s AMI Labs Raises $1.03B: The Biggest Bet Against LLMs in AI History

$1.03 billion. That’s what investors just put behind the single most contrarian argument in modern AI: that large language models are fundamentally the wrong approach, and the field has been building on a broken foundation. The person making that argument is Yann LeCun — and with Yann LeCun AMI Labs now funded and running, we’re about to find out if he’s right.

What Is AMI Labs?

Advanced Machine Intelligence Labs — AMI, pronounced like the French word for “friend” — was founded in late 2025 by Turing Award winner Yann LeCun after he departed Meta, where he had spent a decade as chief AI scientist. On March 10, 2026, AMI Labs announced a $1.03 billion seed round at a $3.5 billion pre-money valuation — Europe’s largest-ever seed round, completed just four months after founding.

LeCun himself serves as executive chairman. Alexandre LeBrun (founder of Nabla, the medical AI startup) is CEO. The leadership bench includes Saining Xie as chief science officer, Pascale Fung as chief research and innovation officer, and Michael Rabbat as VP of world models — collectively representing some of the most credentialed names in academic AI research.

The Core Thesis: LLMs Have a Ceiling

LeCun has been making this argument publicly for years: large language models — the architecture behind ChatGPT, Claude, and Gemini — are fundamentally limited. They operate in the domain of language, predicting the next token based on statistical patterns. They don’t understand causality. They can’t model the physical world. They hallucinate because they’re not grounded in reality.

His alternative: world models. AI that learns from reality in the way animals and humans do — building internal models of how the world works, not just how language describes it. A world model can predict what will happen if you push an object off a table. An LLM can describe it.

JEPA: The Technical Architecture Behind AMI

AMI’s technical approach centers on JEPA — the Joint Embedding Predictive Architecture, which LeCun first proposed in 2022 while still at Meta. JEPA works differently from transformer-based language models: instead of generating tokens autoregressively, it learns by predicting abstract representations of future states in embedding space.

The key difference in practice: JEPA learns more efficiently from less data, can handle multi-modal inputs (video, audio, sensor data, not just text), and is designed to be inherently more robust and predictable than models that generate outputs token by token. It’s a fundamental architectural bet — not just a scaling bet.

The Investor List: Who’s Backing the LLM Skeptic

The $1.03B round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. Strategic investors include NVIDIA, Samsung, Sea, Temasek, and Toyota Ventures. Notable individual investors: Eric Schmidt, Mark Cuban, Xavier Niel, and Tim Berners-Lee.

The presence of NVIDIA is particularly notable. Jensen Huang’s company has a financial interest in the scaling approach (more compute = more GPU sales), yet NVIDIA still backed AMI. That’s either a hedge or a genuine belief that world models represent a real technical direction — possibly both.

Paris as Headquarters: A Deliberate European Counterplay

AMI Labs’ Paris headquarters isn’t incidental — it’s central to LeCun’s pitch. “We are one of the few frontier AI labs that are neither Chinese nor American,” he has said publicly. With additional offices planned in New York, Montreal, and Singapore, AMI is positioning itself as genuinely global while being European-rooted.

French investors including Association Familiale Mulliez, Groupe Industriel Marcel Dassault, and Publicis Groupe participated. This alignment with French industrial capital suggests AMI is targeting real-world industrial applications, not just AI research papers — consistent with their stated focus on manufacturing, healthcare, and robotics.

What AMI Is Building: The Target Applications

AMI Labs’ mission statement targets specific high-stakes domains where “reliability, controllability, and safety really matter”:

• Industrial process control — factory automation requiring precise physical world modeling
• Robotics — robots that understand and navigate physical environments
• Healthcare — clinical AI requiring causal reasoning, not pattern matching
• Wearable devices — continuous AI inference on low-power hardware
• Automation — enterprise workflow automation grounded in physical and logical constraints

These are not chatbot use cases. They’re domains where a model that generates plausible-sounding text but hallucinates facts is actively dangerous. World models, if they work as LeCun argues, are inherently better suited for these applications than LLMs.

Timeline and Open Source Plans

LeCun has been direct about timeline expectations: AMI Labs will spend its first year focused entirely on R&D. This is a long-term scientific project, not a product that ships in 90 days. LeCun told Wired that AMI plans to release its first models “quickly” — which in world model terms likely means 2026-2027 for initial research releases. AMI plans to open-source some of its technology and publish academic papers, following the open science tradition LeCun championed at Meta’s FAIR lab.

Why This Matters: The Stakes of the LeCun Bet

If LeCun is right about LLMs having a fundamental ceiling, then AMI Labs could be the research lab that charts the post-LLM path in AI. If he’s wrong — if scaling continues to unlock reasoning capabilities that overcome the limitations he identifies — then AMI Labs becomes an important but niche research organization working in parallel to the LLM mainstream.

Either way, $1.03B in funding ensures AMI Labs will produce rigorous research for years. In a field dominated by two Chinese-American AI labs racing each other on the scaling path, a well-funded European lab with a genuinely different technical bet is exactly the kind of scientific diversity the field needs. Follow AMI Labs’ publications via TechCrunch’s coverage and the MIT Technology Review deep dive.