Yann LeCun's AMI Labs Raises $1.03B Seed to Challenge LLMs with JEPA Architecture

On March 10, 2026, Turing Award laureate Yann LeCun unveiled the most audacious bet against the AI mainstream in years. His new venture, Advanced Machine Intelligence Labs — AMI, French for "friend" — closed a $1.03 billion seed round at a $3.5 billion pre-money valuation, making it the largest seed-stage investment in European startup history. The thesis is deceptively simple and enormously ambitious: the large language models that dominate today's AI landscape have hit a structural ceiling, and the path to genuine machine intelligence runs through an entirely different architecture called JEPA — Joint Embedding Predictive Architecture.

The announcement marks a pivotal moment for the AI industry. One of the field's most decorated scientists isn't just publishing papers critiquing the status quo — he's raised a billion dollars in four months to prove the entire industry has it wrong.

From Meta's Chief Scientist to Founder

LeCun departed Meta in November 2025 after more than a decade as its Chief AI Scientist. His exit wasn't sudden. For years, he had been increasingly vocal in his criticism of LLMs, arguing that next-token prediction on text data constitutes a fundamental dead end for achieving real-world intelligence. His now-famous analogy encapsulates the argument: "We don't have autonomous cars that can learn to drive in 20 hours of practice, the way a 17-year-old can." Reading driving manuals — the textual approach — will never replicate the embodied learning that comes from actually experiencing the physical world.

By December 2025, Fortune reported LeCun's plans for AMI Labs. The speed of execution was remarkable: within four months of founding, the company had assembled a world-class team, established offices across four continents, and closed one of the largest seed rounds in AI history. The company is headquartered in Paris, with research hubs in New York, Montreal, Singapore, and Zürich. French President Emmanuel Macron has publicly backed the initiative, positioning AMI as a flagship of European AI sovereignty.

The Dream Team

The leadership roster reads like an all-star lineup of AI research. CEO Alexandre LeBrun previously led Nabla, a medical AI startup. Chief Science Officer Saining Xie comes from Google DeepMind. Chief Research and Innovation Officer Pascale Fung and VP of World Models Mike Rabbat are both Meta AI alumni. COO Laurent Solly rounds out the executive team. The research staff skews heavily toward vision researchers and self-supervised learning specialists — a deliberate departure from the text-centric talent that dominates most frontier AI labs.

JEPA: The Architecture That Bets Against Tokens

At the technical core of AMI Labs lies JEPA, a framework LeCun first proposed in a 2022 position paper — notably before the ChatGPT explosion reshaped the industry around LLMs. The architectural difference is fundamental. Where LLMs learn by predicting the next token in a sequence, JEPA predicts future states in a compressed latent space. In LeCun's formulation, LLMs learn "the surface of text" while JEPA learns "the rules of the world."

The distinction matters enormously for physical-world applications. When an LLM "knows" that water boils at 100°C, it has learned statistical correlations between tokens. A JEPA-based world model, trained on video data, observes actual phase transitions — water heating, bubbling, converting to steam — and learns causal relationships directly. The architecture takes pairs of related inputs (such as consecutive video frames), encodes them into abstract representations that capture only essential features, and trains a predictor module to forecast the next frame's representation from the current one. Irrelevant noise and uncertainty are filtered out during encoding, preserving core information while discarding distractions.

Meta developed several JEPA variants before LeCun's departure: I-JEPA for images, V-JEPA 2 trained on over one million hours of video, and VL-JEPA for multimodal integration. V-JEPA 2 is particularly significant — it's the first video-trained world model capable of zero-shot planning and robot control in novel environments. In controlled comparisons against standard vision-language model training using identical data and vision encoders, VL-JEPA achieved stronger performance with 50% fewer trainable parameters.

The efficiency advantage is architectural, not just engineering. Real-world sensor streams contain vast amounts of unpredictable or irrelevant entropy. Rather than trying to reconstruct every pixel or token — a computationally wasteful endeavor — JEPA operates in compressed representation space, focusing computational resources on learning causal structure and physical rules.

The Investor Coalition

The $1.03 billion round was co-led by five firms: Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions — the personal investment vehicle of Amazon founder Jeff Bezos. Strategic investors include Nvidia, Toyota Ventures, Samsung, and Singapore sovereign wealth fund Temasek. French VC firm Daphni and South Korean investor SBVA also participated, alongside a remarkable roster of individual investors: Tim and Rosemary Berners-Lee, venture capitalist Jim Breyer, entrepreneur Mark Cuban, and former Google CEO Eric Schmidt.

The investor composition tells a strategic story. Nvidia's participation signals that the GPU giant is hedging beyond LLM workloads. Toyota Ventures' involvement reflects world models' direct applicability to autonomous driving — predicting how traffic scenes evolve through physics-grounded reasoning rather than tokenized representations. Samsung and SBVA's investment demonstrates Asian tech's appetite for post-LLM paradigms. French industrial groups Dassault and Mulliez bring manufacturing and aerospace use cases to the table.

Product Roadmap: Patience Required

AMI's first product is AMI Video, a proprietary video-based model targeting robotics, manufacturing, and wearables. But CEO LeBrun has set expectations firmly: "This is not a typical AI startup that can ship a product in three months, generate revenue in six, and hit $10M ARR in twelve." Year one is dedicated to fundamental research, talent acquisition, and open-source code publication. Corporate partnership discussions are expected within 6-12 months.

The target market comprises organizations operating complex physical systems where errors carry severe consequences: manufacturers, aerospace companies, biomedical firms, and pharmaceutical groups. In industrial robotics, world models enable robots to mentally simulate grasping strategies for novel objects without reprogramming. For autonomous vehicles, the approach generates edge-case simulations — unpredictable pedestrians, extreme weather — for testing before real-world deployment. In healthcare, where LLM hallucinations carry the highest cost, the vision transitions from clinical documentation to decision support that simulates treatment propagation through biological systems.

The Competitive Landscape

LeCun himself predicted it: "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." The landscape is already heating up. Fei-Fei Li's World Labs raised $1 billion in February 2026 and has shipped Marble, a commercial product generating persistent 3D environments. Google DeepMind's Genie 3 creates photorealistic interactive environments from text prompts. Nvidia's Cosmos platform has surpassed 2 million downloads by early 2026.

Most researchers envision hybrid integration rather than outright replacement — combining LLM language reasoning with world model physical planning. The risk of "world model washing," where standard LLMs get superficially rebranded, is already drawing skepticism from the research community.

The Billion-Dollar Question

AMI Labs represents the most direct institutionalization of the anti-token-maximalist view from a field luminary. LeCun has expressed ambitions for AMI to become Europe's first GAFAM-scale company. But the critical question, as multiple observers have noted, "is not whether world models sound compelling, but whether JEPA-style methods can scale."

Technical validation remains early-stage. No concrete benchmarks for AMI's own systems have been published, meaning the billion-dollar investment is, at this point, a pure bet on LeCun's reputation and vision. The company's commitment to open-source releases and public research validation will be essential credibility markers in its first year.

The next 6-12 months will deliver the first real checkpoints: initial AMI Video demonstrations, industrial partnership announcements, and open-source research publications. Whether JEPA can deliver on its theoretical promise at commercial scale remains the defining question. What's already clear is that the AI industry's most consequential contrarian bet has officially left the whiteboard and entered the arena — backed by a billion dollars and the conviction of one of deep learning's founding figures.

LeCun has spent years arguing that LLMs are a dead end for real intelligence. Now he has the resources to prove it. The world — and the world models — are watching.