Yann LeCun Raises $1.03B for AMI Labs: World Models, JEPA, and What Comes After Transformers

Yann LeCun has been saying for years that large language models are not the path to artificial general intelligence. Now he's betting $1.03 billion on a different path — his own.

AMI Labs, LeCun's new research company, raised what TechCrunch called the largest seed round in European history in early March 2026. The backers read like a who's who of the AI industry: Jeff Bezos, NVIDIA, Samsung, Eric Schmidt, and Tim Berners-Lee. The mission is to build world models using a new architecture — JEPA — that LeCun believes will get AI to genuine physical and causal reasoning in ways that transformers never will.

This is worth understanding not because AMI Labs is a product you can use today, but because LeCun's bet, and who's funding it, says something important about where the AI field's most serious money and minds think the next decade is heading. For instance, while projects like Gemma 4 are focusing on local language models, AMI Labs is taking a different approach by emphasizing physical simulation.

Who Is Yann LeCun?

Context matters here. LeCun isn't a contrarian startup founder positioning against the establishment — he is the establishment.

• Turing Award winner (2018) — the highest honor in computer science, shared with Geoffrey Hinton and Yoshua Bengio for foundational work on deep learning
• Inventor of convolutional neural networks (CNNs) — the architecture that powered the computer vision revolution; every smartphone face-unlock and self-driving car perception stack descends from his work
• Chief AI Scientist at Meta, a role he held until founding AMI Labs
• NYU Professor in computer science and data science

When LeCun says transformers are a dead end, it's not a hot take from someone selling something. It's a considered position from one of the three people most responsible for modern deep learning — and it's been his consistent position for several years.

What He's Building: World Models and JEPA

AMI Labs is building world models — AI systems that can simulate the physical world, predict outcomes based on those simulations, and understand causality. This approach contrasts with current large language models like transformers, which excel at pattern recognition but lack a deep understanding of the physical world. LeCun's vision aligns with advancements in multi-agent orchestration, where multiple AI agents work together to achieve complex tasks by simulating real-world interactions.

The JEPA architecture, which stands for Joint Embedding and Prediction Architecture, is designed to integrate perception, action, and prediction in a unified framework. This could lead to more robust AI systems that can handle the complexities of the real world, much like how AI Agent Guardrails aim to ensure safe and effective AI operations by validating outputs and maintaining control over agent actions.

When LeCun says transformers are a dead end, it's not a hot take from someone selling something. It's a considered position from one of the three people most responsible for modern deep learning — and it's been his consistent position for several years. This shift in focus highlights the evolving landscape of AI research, where innovations like those at AMI Labs are paving the way for more sophisticated and capable AI systems.

ict consequences of actions, and plan multi-step behavior in physical environments. This is different from what LLMs do.

LLMs are trained to predict the next token in a text sequence. They're extremely good at this. But LeCun argues that predicting tokens doesn't require — and doesn't develop — genuine understanding of causality, physics, or the structure of the physical world. An LLM can tell you what happens when you drop a glass, but it doesn't "understand" gravity; it has seen many descriptions of glasses breaking and pattern-matches from those.

A world model, in LeCun's framing, would actually represent how physical objects interact, enable planning in the physical domain, and generalize to novel situations the way humans do from infancy.

JEPA: Joint Embedding Predictive Architecture

The technical approach AMI Labs is pursuing is JEPA — Joint Embedding Predictive Architecture — which LeCun has been developing and publishing research on for several years.

The core idea of JEPA is different from both transformers and previous generative models:

Standard generative model (e.g., diffusion): Predicts raw observations (pixels, tokens). The model must learn everything about the world from the signal of "what comes next."

Transformer LLM: Predicts next tokens in abstract sequence space. Extremely effective for language, but the representation is tied to the structure of human language rather than physical reality.

JEPA: Predicts abstract *representations* of future states rather than the future states themselves. The model learns to predict what's *relevant* about the future (an embedding) rather than all the details.

Why does this matter? LeCun argues that the JEPA approach:

• Forces the model to learn more abstract, causal representations of the world
• Avoids the problem of having to model irrelevant details (the exact texture of a surface isn't relevant to predicting that an object will fall)
• Scales better to physical/embodied tasks
• Is more computationally tractable for world modeling than generating raw pixel predictions

The architecture has already shown results in video prediction and visual representation learning in Meta research (I-JEPA, V-JEPA papers, 2023-2025). AMI Labs is the commercial vehicle to take these ideas further.

The LLM Dead End Argument

LeCun has made this argument many times, and it consistently generates controversy. It's worth laying out his actual position rather than the strawman version.

LeCun's claim is NOT:

• That LLMs are useless
• That language models won't get better
• That current LLMs don't have impressive capabilities

LeCun's claim IS:

• That next-token prediction on text is an insufficient training signal to develop the kind of causal, physical, and common-sense reasoning that characterizes human intelligence
• That LLMs are "linguistically fluent but physically ignorant" — they can describe the world well but cannot plan in it at a fundamental level
• That the path to genuinely general AI requires architectures and training objectives more grounded in physical world modeling than language modeling

The evidence he points to: LLMs still fail at tasks that require understanding physical causality, spatial reasoning, and multi-step physical planning — problems that a 4-year-old child handles effortlessly.

His critics point out that these limitations have been steadily eroded by scale and architecture improvements (GPT-4 is much better at these tasks than GPT-2). LeCun's response: you can approach the capability curve but the fundamental training objective is wrong and will always require massive amounts of data to overcome that limitation.

This is a genuine scientific disagreement that won't be settled by argument — it will be settled by which approach produces better AI systems over the next decade. AMI Labs is LeCun's entry in that experiment.

The $1.03B: What Does It Fund?

Seed rounds of this size are extremely unusual. The largest tech seed rounds before 2026 were typically in the $10-100M range. A $1.03B seed — raised at company formation, before product or revenue — reflects several things:

The credibility premium on LeCun. Backing the Turing Award winner, CNN inventor, and Meta's chief AI scientist costs a premium. Investors are buying his track record and network, not just the idea.

The compute requirements. Serious world model research requires serious compute. Training at the scale needed to demonstrate JEPA on complex physical domains is expensive. $1B gets you a few years of serious GPU/TPU budget.

The talent war. Top AI researchers cost $1-10M/year at the frontier. Building a lab capable of competing with DeepMind, OpenAI, and Google Research requires paying those rates and covering research infrastructure.

The strategic position. If LeCun is right — if world models via JEPA is the path forward — then being the company that owns that paradigm is worth more than $1B. Investors are buying an option on a paradigm shift.

The Backers and What They Signal

The investor list is not random.

Jeff Bezos: Already invested in Anthropic ($4B+) and has bet on multiple AI frontiers. Adding AMI Labs gives him exposure to a non-transformer bet.

NVIDIA: NVIDIA's business model is selling compute. More AI research spending = more GPU sales. Backing AMI Labs is partly a business development move — ensure that whatever architecture emerges next runs on NVIDIA silicon. But it also signals NVIDIA believes world models are a real research direction, not a dead end.

Samsung: Samsung is building AI into devices and needs capabilities beyond what current edge LLMs offer. World models with physical reasoning are directly relevant to robotics and smart device applications where Samsung wants to compete.

Eric Schmidt: Former Google CEO, prolific AI investor (Scale AI, Istria, others). His portfolio tracks serious technical bets with long time horizons.

Tim Berners-Lee: The inventor of the World Wide Web. His involvement is unusual — he's been largely focused on data privacy and web standardization work. His backing suggests the project has implications he finds important for the future of open information infrastructure.

Why This Matters for the AI Industry

AMI Labs matters for a few reasons beyond the fundraising headline:

Talent signal. LeCun building an independent lab will attract researchers who believe the LLM paradigm is incomplete. This could pull senior talent away from Google DeepMind, Meta, and OpenAI in ways that matter.

Research legitimacy. JEPA and world model research has been a minority position in the field. $1B from credible investors + LeCun's name gives it institutional legitimacy and will accelerate publication and discourse.

Robotics implications. The most direct application of world models is robotics — physical AI that needs to plan in the real world. The robotics sector (Figure AI, Tesla Optimus, 1X, etc.) is all wrestling with the same problem: language models don't generalize well to physical manipulation. If AMI Labs makes progress on physical world modeling, the robotics industry is the most immediate beneficiary.

The post-transformer era. Every major AI lab is quietly working on what comes after transformers. AMI Labs is the most public, best-funded bet on a specific alternative paradigm. Its results will shape the next phase of the field.

What AMI Labs Is NOT

To be clear about the timelines involved:

• AMI Labs does not have a product or service available today
• World models at the capability level LeCun is describing are likely 5-10 years from consumer applications
• This is fundamental research, not an application company
• Nothing in the seed round suggests near-term commercialization

This is a long-horizon bet on the future of AI architecture. Following AMI Labs makes sense if you care about where AI is heading over the decade, not if you're looking for tools to use today.

The Bottom Line

Yann LeCun's AMI Labs is the most significant structural bet against the transformer-and-LLM paradigm made by anyone with the credentials to be taken seriously. $1.03B in seed funding from Bezos, NVIDIA, Samsung, and Schmidt signals that the world's most sophisticated AI investors think the next 10 years of AI development won't look like the last 5.

Whether LeCun is right — whether JEPA and world models are the path forward or a detour — is an empirical question that will take years to answer. But the question itself, and AMI Labs as the institution asking it, is now one of the most important in AI.

Track this: TechCrunch: AMI Labs raises $1.03B

FAQ

What is AMI Labs?

AMI Labs is a research company founded by Yann LeCun (Turing Award winner, former Meta Chief AI Scientist) in early 2026. It raised $1.03B in seed funding — the largest seed round in European history — to build world models using the JEPA architecture. Backers include Jeff Bezos, NVIDIA, Samsung, Eric Schmidt, and Tim Berners-Lee.

What is JEPA?

JEPA (Joint Embedding Predictive Architecture) is an AI architecture developed by LeCun and colleagues that learns to predict abstract representations of future states rather than raw observations. The goal is to develop more causal and physically grounded reasoning than is possible with standard transformer-based language models.

What is a world model in AI?

A world model is an AI system that learns a structured representation of how the physical world works — causality, object permanence, spatial relationships, and physical dynamics — that can be used to plan actions and predict outcomes. LeCun argues this is the missing ingredient for AGI that LLMs lack.

Why does LeCun call LLMs a "dead end"?

LeCun's argument is that next-token prediction on text is an insufficient training signal to develop genuine physical and causal reasoning. LLMs are linguistically fluent but physically ignorant — they can describe physical events but lack a model of physical causality. He believes world models trained with architectures like JEPA are a more direct path to general intelligence.

When will AMI Labs have products available?

AMI Labs is a fundamental research company, not an application company. World models at the capability level they're targeting are likely 5-10 years from consumer applications. This is a long-horizon research bet, not a near-term product announcement.