Yann LeCun's Bet Against LLMs Just Raised $1 Billion

What if the AI everyone is racing to build is the wrong kind of AI?

That's the quiet provocation behind AMI Labs, the French startup cofounded by Yann LeCun — Turing Prize winner, NYU professor, and until recently Meta's chief AI scientist. This week, AMI Labs announced it had raised $1.03 billion at a $3.5 billion pre-money valuation. The number is striking. The reasoning behind it is more interesting.

The Problem With AI That Only Reads

Large language models — the technology powering ChatGPT, Claude, and most of the AI products you use daily — learn from text. Vast quantities of it. They're extraordinarily good at generating plausible language, but they have a structural weakness: they don't understand the world. They understand descriptions of the world.

For most applications, that gap is a nuisance. For some, it's dangerous. Alexandre LeBrun, AMI Labs' CEO, knows this firsthand. Before joining the startup, he ran Nabla, a digital health company building AI tools for clinicians. He watched LLMs confidently hallucinate drug dosages, invent lab results, and fabricate clinical details — errors that in a medical context aren't embarrassing, they're life-threatening. "Hallucinations could have life-threatening repercussions," LeBrun said. That experience led him to the same conclusion LeCun had reached independently: LLMs have a ceiling, and it matters where that ceiling sits.

The alternative AMI Labs is building is called a world model — AI trained not just on language, but on the underlying structure of physical reality. The technical foundation is LeCun's JEPA (Joint Embedding Predictive Architecture), proposed in 2022. Rather than predicting the next word in a sequence, JEPA-based models learn to predict the future state of a situation — closer to how humans build intuitive understanding of cause and effect.

$1 Billion for a Concept That Doesn't Have a Product Yet

Here's the tension at the heart of this story: AMI Labs has no revenue, no product, and no near-term plan to generate either. LeBrun is explicit about this. "It's not your typical applied AI startup that can release a product in three months, have revenue in six months and make $10 million in ARR in 12 months," he said. Commercial applications could be years away.

And yet the investor list reads like a who's-who of technology capital. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. Strategic investors include NVIDIA, Samsung, Toyota Ventures, and Temasek. Individual backers include Tim Berners-Lee, Jim Breyer, Mark Cuban, and Eric Schmidt.

Sponsored

Advertise with Us

[email protected]

Sponsored

The strategic investors are particularly telling. NVIDIA sells the compute that trains these models. Samsung makes the chips and hardware that could eventually run them at the edge. Toyota has been quietly building one of the most serious robotics programs in the automotive industry. These aren't passive bets on an interesting idea — they're hedges from companies that need to know what comes after LLMs.

The funding also reflects a broader pattern. Fei-Fei Li's World Labs raised $1 billion last month. European startup SpAItial closed a $13 million seed round — unusually large for a European early-stage company. The world model category is attracting capital at a pace that suggests investors believe the LLM era, while far from over, is not the final chapter.

The Buzzword Problem

LeBrun himself flagged the risk with characteristic candor: "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." He said this with a smile — because he thinks AMI Labs is genuinely different — but the warning is real.

The history of AI is littered with terms that started as precise technical concepts and became marketing language. "Deep learning," "neural network," "generative AI" — each went through a cycle where the signal got buried in noise. If world models follow the same path, distinguishing genuine research from rebadged LLM wrappers will become its own industry.

AMI Labs is trying to inoculate itself against this by committing to open research. The startup plans to publish papers and release code as open source — a stance that LeCun maintained at Meta's FAIR lab and that is, as LeBrun noted, "increasingly rare" in an era where frontier AI research is increasingly proprietary. The logic: open research builds community, and community accelerates progress. Whether that philosophy survives contact with commercial pressure is an open question.

What This Means If You're Not a Researcher

For AI practitioners and investors, the relevant question isn't whether world models are theoretically superior to LLMs. It's whether they'll be practically superior in time to matter.

The most plausible near-term applications aren't consumer products — they're industrial and professional ones. Healthcare, where Nabla will be the first deployment partner. Robotics, where physical world understanding is the core unsolved problem. Manufacturing, where LLMs already struggle with spatial and causal reasoning. These are markets where accuracy isn't a nice-to-have.

For startup founders building on top of LLMs today, the honest assessment is: nothing changes in the next 12-18 months. But the architectural question — whether to build on text-prediction foundations or wait for something more grounded — is worth tracking. The companies that got locked into early mobile frameworks sometimes found themselves rebuilding from scratch when the platform shifted.

For investors, the calculus is simpler and harder simultaneously. The LLM market has incumbents with enormous compute advantages. The world model market has none yet. The risk is that world models never achieve commercial viability at scale. The upside is a position in what could become the foundational AI layer for physical industries.