Yann LeCun Launches $1.03 Billion AI Startup AMI Labs

On March 10, 2026, Advanced Machine Intelligence Labs—better known as AMI Labs—announced what may be the most symbolically loaded funding round in the history of artificial intelligence: $1.03 billion in seed capital, raised in less than four months, to build AI systems that the company's founder believes will make today's dominant technology obsolete [1][2].

The founder is Yann LeCun, the 65-year-old French-American computer scientist who won the Turing Award in 2018 for his pioneering work on convolutional neural networks. After 12 years at Meta—first as founding director of Facebook AI Research (FAIR) and then as the company's chief AI scientist—LeCun walked away in November 2025, convinced that the large language models at the center of the AI industry's $100-billion-a-year strategy are, as he has repeatedly put it, "a dead end" [3][4].

Now, with nearly $900 million in euros (€890 million) committed at a $3.5 billion pre-money valuation, LeCun has the war chest to prove it—or to stage one of the most expensive scientific experiments in startup history [5].

The Departure That Shook Meta

LeCun's exit from Meta had been rumored for weeks before he confirmed it on November 18, 2025, via a LinkedIn post that read like a manifesto [6]. The departure did not happen in a vacuum. Meta's AI unit had been dramatically overhauled over the prior year after the company's Llama 4 model disappointed developers and executives alike. CEO Mark Zuckerberg pivoted away from the long-term foundational research that LeCun's lab had pursued since 2013, reorganizing the division under Superintelligence Labs, led by new hire Alexandr Wang, the former CEO of Scale AI [7].

Meta then cut 600 employees from its Superintelligence Labs division in October 2025, including researchers from FAIR—the lab LeCun had built from nothing into one of the world's premier AI research institutions. Former Meta employees described FAIR as "dying a slow death" [8]. For LeCun, who had spent years arguing internally that LLMs were insufficient to achieve genuine machine intelligence, the strategic direction was untenable.

"You certainly don't tell a researcher like me what to do," LeCun told The Decoder, explaining his decision to leave [9].

While Meta will not invest in AMI Labs, the two companies have agreed to maintain a partnership, allowing some continued research collaboration [3].

A Contrarian Thesis With Academic Roots

LeCun's central argument is both simple and radical: large language models are a "statistical illusion"—impressive in output, but fundamentally incapable of understanding the world [10][11]. No matter how much data they consume or how many parameters they contain, LLMs operate by predicting the next token in a sequence. They do not reason. They do not plan. They do not understand cause and effect. They have, as LeCun has colorfully put it, less understanding of the world than a housecat [12].

The alternative he proposes is what he calls "world models"—AI systems that build internal representations of how environments function and use those representations to simulate outcomes, reason about consequences, and plan complex actions. The technical foundation is an architecture LeCun first proposed in a 2022 paper called JEPA, or Joint Embedding Predictive Architecture [13].

Unlike generative models that predict raw outputs like pixels or words, JEPA works in an abstract embedding space. It learns to predict the representation of a future state rather than every sensory detail. As LeCun has explained: "The key is to learn an abstract representation of the world and make predictions in that abstract space, ignoring the details you can't predict" [14].

Architecturally, JEPA uses encoder-only transformers with a context pathway encoding observed information, a target pathway encoding masked information, and a predictor operating entirely in latent space. The approach is not generative—it does not produce images, text, or sound—but rather builds an internal model of causal relationships [14].

AMI Labs plans to train its systems not on text corpora but on video, audio, and sensor data of all kinds—from robot arm positions to lidar readings to audio streams. The company is developing a proprietary model called AMI Video as one of its first major research initiatives [5][14].

The Team: A French-American Brain Trust

LeCun serves as executive chairman, while the CEO role belongs to Alexandre LeBrun, a serial French entrepreneur who previously co-founded and led Nabla, a healthcare AI startup. LeBrun's path to AMI Labs has a certain narrative logic: at Nabla, he had reached the same conclusion as LeCun about the limitations of LLMs, particularly in healthcare where hallucinations can have life-threatening consequences [15][16].

Before Nabla, LeBrun co-founded Wit.ai, a natural language processing startup that Facebook acquired in 2015, which led him to work directly under LeCun at FAIR [16].

The broader leadership team reads like a who's who of AI research [5][17]:

• Laurent Solly, Meta's former VP for Europe, serves as COO
• Saining Xie, a former Google DeepMind researcher, is chief science officer
• Pascale Fung, formerly a senior AI director at Meta, is chief research and innovation officer
• Michael Rabbat, former research science director at Meta, is VP of world models

The company is headquartered in Paris, with additional offices planned in New York, Montreal, and Singapore—a deliberately global footprint that reflects both the international pedigree of its founders and the ambition to attract talent across continents [1][5].

The Money: Europe's Largest Seed Round

The $1.03 billion round is not just AMI Labs' coming-out party—it is Europe's largest seed round ever, and one of the largest globally, trailing only Thinking Machines Lab's $2 billion raise led by Andreessen Horowitz [18].

Source: Crunchbase / TechCrunch

Data as of Mar 10, 2026CSV

The round was co-led by five firms: Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions—Jeff Bezos's personal investment vehicle [1][2]. The investor list extends well beyond venture capital into strategic territory:

• Nvidia and Samsung—hardware giants with obvious interest in an AI paradigm that could reshape compute requirements
• Toyota Ventures—reflecting AMI's ambitions in robotics and manufacturing
• Temasek—Singapore's sovereign wealth fund
• Bpifrance Digital Venture—the French state investment bank
• Sea and SBVA—Asian strategic investors

The individual investor roster is equally notable: Eric Schmidt (former Google CEO), Mark Cuban, Jim Breyer (early Facebook investor), Xavier Niel (French telecom billionaire), Tim and Rosemary Berners-Lee (the inventor of the World Wide Web and his wife), and Groupe industriel Marcel Dassault [2][5].

LeCun had initially sought €500 million, but according to a leaked pitch deck reported by Sifted, demand substantially exceeded that target [5].

Daphni's Pierre-Éric Leibovici captured the ambition of European backers: "AMI Labs could be the first European company to reach GAFAM scale" [5].

The Landscape: AI Funding in the Age of Mega-Seeds

AMI Labs arrives at a moment when the definition of "seed round" has become almost unrecognizable. In 2025, investors backed nearly 700 seed-stage rounds of $10 million or more. The total value of seed rounds exceeding $100 million hit $10 billion. Over 42% of all global seed funding went to AI-focused companies, with just over $15 billion flowing to AI-focused seed rounds—a 50% increase from 2024 [18][19].

Source: Crunchbase

Data as of Mar 10, 2026CSV

The trend has only accelerated into 2026, with over 40% of seed and Series A investment going to rounds of $100 million or more [19]. Whether this represents rational capital allocation or a bubble waiting to deflate is an open question—one that AMI Labs, with its pre-revenue $3.5 billion valuation, embodies in sharp relief.

The Skeptics' Case

For all the excitement, AMI Labs faces formidable challenges that its backers are well aware of.

First, the company has no product, no revenue, and no near-term commercial prospects. LeCun has acknowledged that AMI Labs will spend its entire first year purely on research [10]. He projects corporate partnerships within one to two years and "fairly universal intelligent systems" deployable across domains within three to five years [10].

Second, the JEPA framework, while theoretically elegant, remains unproven at scale. Meta's own V-JEPA and V-JEPA 2 projects showed promising results in video understanding benchmarks, but the leap from research demonstrations to commercially viable systems that outperform LLMs in real-world applications is vast [14].

Third, the competitive landscape is daunting. OpenAI, Anthropic, Google DeepMind, and dozens of well-funded startups are not standing still. Many are incorporating multimodal capabilities and reasoning systems into their LLM architectures, potentially narrowing the gap that LeCun claims is unbridgeable.

As The Next Web observed in its analysis: "The question is whether being right about the problem is the same as being right about the solution" [10].

Even Ilya Sutskever, OpenAI's former chief scientist who shares some of LeCun's skepticism about pure scaling, has struck a more measured tone, stating that "the era of 'Just Add GPUs' is over" without endorsing a wholesale abandonment of the LLM paradigm [11].

The Target Markets: From Robots to Wearables

AMI Labs has identified three primary domains for its world model technology: robotics, manufacturing, and wearables [5].

The logic is intuitive. Robots navigating physical environments need to understand physics, predict consequences of actions, and plan multi-step sequences—precisely the capabilities that world models are designed to enable. Manufacturing environments demand the same kind of causal reasoning, while wearable devices could benefit from AI that understands context and anticipates user needs based on environmental cues rather than text prompts.

The company has already established an exclusive partnership with Nabla, LeBrun's former company, to develop agentic healthcare AI powered by world models—a first concrete indication of commercial direction [16].

What's at Stake

The significance of AMI Labs extends well beyond a single company's fundraise. LeCun's departure from Meta and his billion-dollar bet on world models represent a potential inflection point in the AI industry's dominant narrative.

For over three years, the prevailing wisdom has been that scaling LLMs—making them bigger, training them on more data, throwing more compute at inference—would continue to yield improvements that eventually approach general intelligence. This thesis has attracted hundreds of billions of dollars in investment and reshaped the global technology industry.

LeCun is arguing, with his reputation, his career, and now $1.03 billion, that this thesis is wrong. Not incrementally wrong, but architecturally wrong. He believes the entire paradigm needs to change.

If he is right, AMI Labs could pioneer a new generation of AI systems that are safer (because they understand consequences), more efficient (because they operate in abstract representation space rather than brute-forcing predictions), and more broadly intelligent (because they model reality rather than language). The implications for robotics, autonomous vehicles, scientific discovery, and virtually every industry that AI touches would be profound.

If he is wrong—or if he is right in theory but the engineering proves intractable—then AMI Labs will become a cautionary tale about the dangers of contrarian conviction backed by too much capital.

Either way, the world's most credentialed AI skeptic now has the resources to test his hypothesis at scale. The experiment has begun.

Source: GDELT Project

Data as of Mar 10, 2026CSV