For years now, Yann LeCun has been the industry's most prominent heretic. Large language models, he's insisted with mounting frustration, represent a technological cul-de-sac. The chatbot craze? A "misdirected fad," he told TIME back in September 2023. The race to scale transformers into artificial general intelligence? Doomed to "hit a dead end."
On March 9, 2026, LeCun stopped merely talking. His Paris-based startup, AMI Labs, closed a $1.03 billion seed round—among the largest first-time raises in AI's short, frenzied history. The pre-money valuation: $3.5 billion, according to TechCrunch and Dataconomy. Backing him is a roster that reads less like a cap table and more like a manifesto of tech optimism: NVIDIA and Samsung on the strategic side, Cathay Innovation and Greycroft leading venture, and a constellation of individual believers including Tim Berners-Lee, Eric Schmidt, and Mark Cuban. Even Jeff Bezos is in, through Bezos Expeditions.
But AMI isn't building bigger transformers. It isn't chasing the next GPT. Instead, LeCun's company is pursuing what he calls "world models"—systems that learn how the physical universe actually works, not just how language patterns flow. If he's right, today's LLM-powered assistants will soon look like very clever autocomplete. If he's wrong, well. A billion dollars buys time to find out.
The Theoretical Bet
LeCun's argument against LLMs isn't exactly subtle. By April 2025, speaking to Newsweek, he'd put a timeline on it: five years until large language models are "largely obsolete," supplanted by systems capable of reasoning about physical reality.
The logic has a certain elegance. LLMs excel at pattern matching—predicting the next token, answering questions, generating plausible text. What they can't do, LeCun argues, is understand causality. They don't grasp how objects interact, how actions ripple into consequences, how to plan in environments where gravity and momentum and occlusion actually matter. They're astonishingly fluent mimics with no internal model of the world.
World models flip the script. Instead of predicting text tokens, they predict future states of environments based on observations and actions. Feed one enough video data, the theory goes, and it begins to learn the latent structure of physical dynamics. Not just prediction—genuine planning and control.
The technical foundation is something called JEPA, or Joint Embedding Predictive Architecture, which LeCun codified in a June 2023 position paper. Rather than predicting raw pixels, JEPA-based models work in representation space, learning abstract features that capture essential dynamics while filtering out noise. Meta—LeCun's former employer—released V-JEPA 2 in June 2025, trained on over a million hours of video. An action-conditioned variant targeted zero-shot robotic planning.
Now AMI is taking that research lineage and trying to commercialize it. Trying being the operative word.
What They're Actually Building
The technical distinction matters, especially for anyone trying to parse which architectures will power the next wave of products.
World models learn from sensor data—video, images, proprioception, audio—to build internal simulations. They're trained to predict what happens next given current state and potential actions. Done well, this yields systems that can mentally rehearse scenarios, evaluate outcomes, and optimize plans before acting.
AMI's approach, according to a March 10, 2026 company update, emphasizes "action-conditioned world models that learn abstract representations from real-world sensor data." The stated goals: persistent memory, reasoning, planning, controllability, safety. The company has hubs in Paris (headquarters), New York, Montreal, and Singapore—a geographic spread that signals both global ambition and an already fierce talent war.
Recent papers from AMI-affiliated researchers offer glimpses of the work underway. C-JEPA, focused on object-centric world modeling, appeared February 11, 2026. A preprint on "World2Act" demonstrated skill-compositional world models achieving state-of-the-art results on simulated tasks and real-world robotic benchmarks, according to a March 11 paper.
But research excellence and commercial viability are different animals. AMI CEO Alexandre LeBrun—who previously founded Wit.ai (acquired by Meta) and clinical AI company Nabla—estimated in late February 2026 that it would take "about one year to get the first things we can use in the product." TechCrunch, covering the funding announcement, reported that broad commercial applications could take "years."
That's a long runway. The $1.03 billion buys patience, not certainty.
The Crowded Field
AMI isn't alone in this pursuit, not by a long shot. World models have gone from academic curiosity to Valley obsession in under 18 months.
World Labs, founded by AI luminary Fei-Fei Li, landed its own $1 billion round on February 18, 2026—less than a month before AMI's announcement. That round included a $200 million strategic investment from Autodesk, telegraphing immediate commercial intent. World Labs had already shipped Marble, a product generating editable 3D worlds, in November 2025. The Autodesk partnership aims to integrate those capabilities into design and entertainment workflows, giving Li's company a near-term revenue path that AMI doesn't yet possess.
"If AI is to be truly useful, it must understand worlds, not just words," Li said in February. It's an echo of LeCun's thesis, but with code already in production.
Meanwhile, the incumbents aren't standing still. NVIDIA's Cosmos platform, unveiled January 6, 2025 and expanded in March, offers open-weight world foundation models for robotics and autonomous vehicles. Early adopters reportedly include 1X, Agility Robotics, XPENG, and Uber's autonomous division. Google DeepMind's Genie models generate interactive environments; Demis Hassabis flagged world models as a "major frontier" in a December 2025 Axios interview. Even OpenAI frames its Sora video models as steps toward "world simulators," per research posts dating back to February 2024.
Meta itself continues pushing V-JEPA 2 as open-source research, enabling academic work on physics-aware video generation and collision prediction. That feeds the broader ecosystem—but it also compresses AMI's window for differentiation.
Wayve, the London-based autonomous vehicle company, launched GAIA-3 in December 2025, a generative world model for safety evaluation simulating edge-case scenarios. Waabi uses mixed-reality environments for AV training. Skild AI is building generalist robotics foundation models in Pittsburgh. The landscape is crowded, well-funded, increasingly productized.
Healthcare as the Opening Move
AMI's initial vertical is healthcare, via an exclusive partnership with Nabla announced in December 2025 and reiterated in March 2026. Nabla, which focuses on ambient clinical documentation and care workflows, gets first access to AMI's world models. The goal: FDA-certifiable agentic healthcare AI.
It's a shrewd choice, perhaps. Healthcare combines high regulatory barriers (favoring slower, more methodical approaches), massive data moats from clinical workflows, and genuine willingness to pay for systems that improve outcomes or reduce physician burnout. LeBrun, who chairs Nabla while serving as AMI's CEO, presumably understands the product-market fit intimately.
But "FDA-certifiable" is a heavy lift. The FDA released final guidance on Predetermined Change Control Plans for AI-enabled devices in December 2024, with ongoing webinars and comprehensive draft guidance through the 2024-2025 cycle. Those frameworks demand rigorous validation, transparency, post-market monitoring. World models, by their nature, are probabilistic systems making predictions in high-dimensional latent spaces. Translating that into something regulators will approve for clinical decision-making? Uncharted territory.
One person familiar with the partnership, speaking to HLTH in March 2026, called the timeline "optimistic but plausible." An analyst in a Sacra teardown published the same month flagged "data bottlenecks and speed-to-production risks" as material concerns.
The Longer Game—and the Counterfactual
The deeper question isn't whether world models work—the research looks increasingly compelling—but whether they deliver enough value, fast enough, to justify the capital intensity and development timelines.
AMI's leadership bench is formidable. Beyond LeCun (executive chair) and LeBrun (CEO), the roster includes Laurent Solly (former Meta VP Europe) as COO, Saining Xie as chief science officer, Pascale Fung leading research and innovation, and Michael Rabbat overseeing world models. TechCrunch's March 9 interview confirmed AMI intends to "publish papers… make a lot of code open source," staying true to LeCun's academic roots.
That open posture buys credibility and attracts talent. It also gives competitors a clear view of the roadmap.
PwC's March 2026 global industrial manufacturing outlook, based on a July 2025 survey, found that manufacturers expect to more than double automation of key processes by 2030. But 78% see robotics primarily as productivity drivers, not growth engines. That's telling: world models need to solve real problems, not just demonstrate impressive demos.
Regulatory headwinds loom. The EU AI Act begins general application August 2, 2026, with phased obligations for high-risk systems continuing into 2027. Healthcare deployments face evolving FDA frameworks. Industrial robotics relies on fragmented ISO standards rather than unified oversight.
Goldman Sachs research projected the humanoid robot market reaching at least $6 billion within 10-15 years, with robotaxi revenue growing at a 90% CAGR from 2025-2030. Those are long-tail forecasts. AMI has capital to survive the wait, but investors will want proof points sooner.
There's a counterfactual worth considering. Anthropic CEO Dario Amodei, speaking to Axios in late January 2026, predicted Nobel-level AI models within roughly two years and sweeping job impacts—all from scaling current LLM architectures with better agents. If Amodei is even half right, world models might arrive too late to matter.
The cynical read: AMI raised a billion dollars to prove something fundamentally academic that won't ship products until competitors have colonized the market. The optimistic read: LeCun is playing the longest game in AI, betting that physical intelligence is the real prize and that foundation models for the real world will eclipse foundation models for text.
Either way, the next 18 months will be revealing. Watch for AMI's first research releases. Watch for partner demos with Nabla through the second half of 2026. Watch whether NVIDIA Cosmos, Meta's JEPA work, or DeepMind's integrations move faster. Watch the EU AI Act's high-risk obligations as they phase in.
And watch whether LLMs hit the plateau LeCun has long predicted. A billion dollars buys a lot of runway. Whether it buys enough time to be right—that's the bet investors just made.
