AMI Labs raises $1.03B across just 12 employees to advance world models as Yann LeCun argues they will ultimately surpass large language models in AI development.

LeCun’s $1.03B Bet: Why World Models Will Outlast LLMs

by
World model AI and JEPA architecture concept — crystalline brain structure representing abstract reasoning beyond language
LeCun’s AMI Labs is betting $1.03B that the entire LLM industry has taken the wrong road — and that world models built on JEPA are the right one — Generated by OpenAI DALL-E 3
💡 Key Takeaways

  • AMI Labs closed a €890M ($1.03B) seed round on March 10, 2026 — Europe’s largest ever — at a $3.5B pre-money valuation, just four months after founding
  • Yann LeCun calls LLMs “a statistical illusion” and frames JEPA-based world models as the only viable path to genuine machine intelligence
  • Nvidia, Toyota, Samsung, Jeff Bezos, and Eric Schmidt are backing the bet — targeting robotics, industrial control, healthcare, and wearables where LLM hallucinations carry life-or-death costs

$1.03 billion. Twelve employees. No product.

Four months old. Those four facts describe AMI Labs — Advanced Machine Intelligence Labs — and the company that just closed Europe’s largest seed round in history. The money came from Nvidia, Toyota, Samsung, Jeff Bezos, Eric Schmidt, and a dozen others who collectively decided that the Turing Award winner behind them is right about something everyone else is getting wrong.

That something is the LLM paradigm. Yann LeCun, who left Meta after years of internal dissent, has a blunt read on where the AI industry stands: “A statistical illusion.

Impressive, yes. Intelligent, no.” While OpenAI closed a $110 billion round in February 2026 to build bigger language models, LeCun is building something structurally different — and betting that world models will outlast the LLM era entirely.

This is not a philosophical debate. It is a capital allocation contest between two incompatible theories of machine intelligence.

Understanding what JEPA is, why the strategic investors chose this over LLM ventures, and what the real risks look like requires working through each layer carefully. Here is what the evidence shows.

Europe’s Largest Seed Round: The Numbers Behind the Bet

On March 10, 2026, AMI Labs announced it had raised €890 million (approximately $1.03 billion) in seed funding at a $3.5 billion pre-money valuation. That implies a roughly 22.7% stake changing hands — a substantial dilution for a company at the idea stage. The round is the largest seed financing ever recorded for a European startup, surpassing the previous record by a significant margin.

The lead investors — Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions — represent a transatlantic coalition of venture capital. The strategic backers expand the picture considerably: Nvidia, Toyota, Samsung, and Temasek joined as corporate investors, alongside a personal investor list that includes Jeff Bezos, Tim Berners-Lee, Jim Breyer, Mark Cuban, and Eric Schmidt. Each of those names signals a different industrial thesis, which matters when reading where AMI’s eventual commercial footprint might land.

The comparable raises in the peer group tell a story of their own. Fei-Fei Li’s World Labs raised $230 million at a $1 billion valuation in August 2024. Mira Murati’s Thinking Machines Lab reached a $12 billion valuation.

AMI sits between them on the valuation spectrum but commands more raw capital than either at seed stage, a reflection of LeCun’s singular scientific credibility in the field. The company has roughly 12 employees, no product, and a research timeline measured in years — yet investors are treating that as a feature, not a flaw.

Machine learning and world models research data visualization
LLMs memorize patterns across trillions of tokens; JEPA learns the underlying rules that govern how the world changes — Photo by Kevin Ku on Unsplash

LeCun’s Case Against LLMs: A Structural Argument, Not a Preference

LeCun’s critique of large language models is neither new nor softened by corporate politeness. He called LLM scaling a “dead end” while still employed at Meta — a position that placed him in open disagreement with the direction of virtually every major AI lab. His core argument is architectural.

LLMs predict the next token in a sequence. They are trained on trillions of words to do this with extraordinary fluency, and fluency gets mistaken for understanding.

The problem, in LeCun’s framing, is that predicting tokens is not the same as modeling the physical world. LLMs have no representation of causality, no model of objects persisting through time, no understanding of how forces act on bodies in space.

They generate text that sounds correct because it matches statistical patterns in training data, not because the model holds a coherent world model in its weights. This is why hallucinations — confident statements of false facts — are not bugs to be patched but structural properties of the architecture.

The counterargument from the LLM camp is not purely theoretical. Anthropic CEO Dario Amodei told an audience at Davos in January 2026 that AI models built on current architecture would replace the work of all software developers within a year and achieve Nobel Prize-level scientific research within two.

LeCun says the road is a dead end. Amodei says the destination is two years away. At least one of them is catastrophically wrong — and $1.03 billion is riding on LeCun being right.

What JEPA Actually Does: Predictions in Abstract Space

JEPA — Joint Embedding Predictive Architecture — is LeCun’s proposed alternative, and its core innovation is where predictions happen. Generative models like LLMs predict future states token by token or pixel by pixel.

This granularity is a liability: much of what happens in the world is fundamentally unpredictable at the surface level. A falling cup’s exact trajectory depends on air currents, surface texture, and dozens of micro-variables that no model can track.

JEPA makes predictions in abstract representation space instead. Rather than reconstructing what the world looks like, it learns what matters about how the world changes.

The architecture is designed to ignore the details it cannot predict and focus on the abstract structure that it can. As Futurum Group’s analysis describes it, this is analogous to how an infant learns about gravity — not by computing exact parabolas, but by building an intuition for how objects behave under physical laws.

The practical implication for hallucinations is significant. An LLM that confabulates a wrong drug interaction does so because it has learned a plausible-sounding pattern, not because it understands pharmacology.

A world model trained on the dynamics of biochemical systems — in theory — would know what it does not know, because its predictions are grounded in the abstract structure of the domain rather than surface-level token statistics. AMI’s target sectors — robotics, industrial process control, healthcare, and wearables — share one characteristic: hallucinations in these domains have physical, legal, or lethal consequences.

💡 Related Posts
$650B AI Spending Spree — Bubble Reset Already Priced In?

The LLM Counterattack: OpenAI’s $110B and the Scaling Faith

OpenAI closed a $110 billion financing round in February 2026. That is 107 times the capital AMI raised, and it is explicitly committed to scaling LLMs larger.

Sam Altman has stated publicly that we have already crossed into AGI territory and are approaching superintelligence. The implication is that LeCun is not identifying a structural limitation — he is misreading an emergent phenomenon in progress.

The scaling camp has empirical backing for its optimism. Each major generation of LLMs — from GPT-3 to GPT-4 to the GPT-5 family — delivered qualitatively different capabilities at scale that smaller models could not replicate.

Researchers call this “emergence”: behaviors that appear discontinuously as model size crosses certain thresholds. If emergence continues, LeCun’s critique may apply to LLMs of 2024 but not to LLMs of 2028.

The honest version of the question is not whether JEPA is theoretically superior but whether it can deliver economically useful systems faster than LLM-centric agents are already commercializing. LLM vendors are not waiting.

Enterprise contracts are being signed, workflows are being restructured around LLM outputs, and switching costs are accumulating. AMI has 3 to 5 years to show something real before the installed base of LLM-native applications makes displacement structurally harder.

The Team and the Timeline: Who LeCun Built Around Himself

AMI Labs is headquartered in Paris, with planned offices in New York, Montreal, and Singapore. The geographic spread is deliberate. New York gives access to U.S. enterprise clients and financial capital.

Montreal has one of the densest concentrations of academic AI talent in North America. Singapore positions the company close to the Asian industrial clients — Toyota’s manufacturing networks, Samsung’s hardware ecosystem — that represent the most commercially plausible near-term partners.

The executive team is built from LeCun’s extended professional network. CEO Alexandre LeBrun co-founded and led Nabla, a Paris-based health AI company. VP of World Models Michael Rabbat, COO Laurent Solly, Chief Research Officer Pascale Fung, and Chief Science Officer Saining Xie all come from Meta or Google DeepMind.

This is not a random assemblage of AI talent. It is LeCun’s deliberate recruitment of the people he considers most aligned with his technical thesis.

The published timeline is structured to manage expectations. AMI’s stated plan is to spend the first one to two years on pure research and initial corporate partnership discussions, with no product releases and no revenue. The three-to-five-year horizon is when LeCun expects “fairly universal intelligent systems” to emerge from the research.

LeBrun added a candid caveat: “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.” It is an unusual warning to issue at the same moment as a record seed round — and it signals that the team understands the hype cycle risk clearly.

World model AI brain architecture concept — embodied reasoning for robotics and physical world understanding
JEPA targets domains where abstract reasoning about physical reality matters more than text fluency — robotics, surgery, and industrial automation — Photo by Natasha Connell on Unsplash

Why Nvidia and Toyota Bet Against Their Own Ecosystem

Nvidia’s investment in AMI is the most striking element of the round. Nvidia sells GPUs primarily to companies training and running LLMs at scale. If LeCun is right and LLMs are a dead end, Nvidia’s core customer base is building on the wrong foundation.

Yet Nvidia invested anyway — which suggests the company is hedging across architectures, not betting on any single paradigm. Robotics is the key: Nvidia’s next-generation computing stack depends on AI systems that can reason about physical environments, and JEPA’s architecture is explicitly designed for that problem.

Toyota’s logic is more transparent. Autonomous vehicles require AI that understands how the physical world works — lane geometry, object permanence, collision physics. LLM-based perception systems have produced embarrassing failures in autonomous contexts precisely because they pattern-match rather than reason.

An AI that hallucinates a pedestrian’s trajectory is not a minor product issue; it is a liability event. Toyota’s investment in AMI reads as a vote for the proposition that JEPA-style architectures will close this gap where LLMs have not.

The structural risks are real and should not be minimized. AMI’s healthcare pathway runs into FDA certification timelines that may misalign severely with a three-to-five-year research agenda. World models face their own generalization challenges in novel environments — JEPA is not exempt from the problem of distribution shift that plagues all learned systems.

Two indicators will determine which direction the balance tips: whether AMI announces a concrete industrial partnership within the next two years, and whether LLM-based agents reduce their physical-domain error rates to acceptable thresholds before world model alternatives arrive. If AMI moves first, the narrative shifts toward LeCun. If LLM agents solve physical-domain reliability independently, the window for disruption narrows sharply.

💡 Related Posts
OpenAI Pentagon Deal Backfires: Hardware Lead Quits, ChatGPT Uninstalls +295%

References

Stay Ahead of the Curve

Get the latest insights on AI, tech, and markets — delivered in both Korean and English.

Explore More Posts