AI Lab Blog

The $1.03 Billion Bet Against the LLM Era

September 9, 2026

Tags: World Models, AI Infrastructure, AI Governance, Healthcare AI, Agentic AI

Yann LeCun spent more than a decade running Meta’s AI research organization and building the deep learning revolution alongside it. Turing Award, published the CNN papers that made modern vision possible, the works. Then he walked out of Meta and put his credibility behind a single contrarian claim: large language models, as an architecture, will not get us to machines that actually understand the world.

This week, the money caught up with the thesis. AMI Labs (Advanced Machine Intelligence), the company LeCun co-founded after leaving Meta, is moving its world model work into healthcare through an exclusive partnership with Nabla, the clinical AI company behind one of the most widely deployed ambient documentation assistants in medicine. The partnership is the latest beat in a bigger story: earlier this year, AMI closed a $1.03 billion seed round at a $3.5 billion pre-money valuation - the largest seed round ever raised by a European company - with backing from NVIDIA, Temasek, Samsung, Toyota Ventures, Bpifrance, and individuals including Jeff Bezos, Eric Schmidt, and Tim Berners-Lee.

That is not a normal seed round. That is the smartest capital in technology hedging against the current paradigm.

What AMI is actually building

LeCun’s argument, which he has been making publicly for years, is that language is a lossy compression of reality. Text is how humans describe the world to each other, not how the world works. An AI that learns only from text is learning the map, never the territory.

AMI’s answer is the world model - systems built on LeCun’s Joint Embedding Predictive Architecture (JEPA). Instead of predicting the next token in a sequence, JEPA learns abstract representations of environments directly from sensor data: video, audio, physiological streams, imaging. In AMI’s own framing, these systems “predict how situations evolve, and how actions lead to consequences, so that they can plan sequences of actions under real-world constraints.”

Translation for the operators: simulation-based reasoning. What-if analysis before execution. The model does not guess the next plausible-sounding sentence - it models the environment, anticipates outcomes, and plans accordingly.

The founding team is essentially Meta FAIR alumni: LeCun as president, Alexandre LeBrun (founder of clinical AI company Nabla) as CEO, Saining Xie as Chief Science Officer, and Pascale Fung as Chief Research and Innovation Officer. Hubs in Paris, New York, Montreal, and Singapore.

Why healthcare is the interesting part

The Nabla partnership is the detail most coverage is sleeping on, and it is the part that should matter most to anyone running regulated workloads.

LLMs in clinical settings are probabilistic text generators. They are genuinely good at documentation and knowledge retrieval - and they fail exactly where clinical work is hardest: deterministic reasoning, multimodal data processing, and long-horizon planning. A model that is 95% accurate on a discharge summary is a productivity tool. A model that is 95% accurate on a medication decision is a patient safety incident.

AMI’s pitch is that world models address that gap structurally rather than by prompt-engineering around it. Because a world model reasons over a simulation of consequences rather than a distribution of plausible text, its decisions are intended to be auditable: here is the predicted evolution of the patient’s state, here is the planned action, here is the anticipated outcome. That maps directly onto what regulators and hospital oversight committees actually require - a decision trail, not a vibe.

Nabla’s roadmap makes the ambition concrete: moving beyond ambient documentation into agentic workflows - coordinating referrals, navigating scheduling systems, assessing insurance constraints, initiating pre-visit lab orders, executed as multistep workflows across the fragmented mess that is the modern EHR estate.

I run cloud and security operations for a medical center. I sit inside this exact problem space every day - EMR migrations, PACS, HIPAA-aligned controls, all of it. I can tell you the appeal of the argument from the inside: the industry has been bolting guardrails onto probabilistic systems and calling it governance. A model class that plans against a simulated environment, with an auditable decision trail built into the architecture, is a genuinely different governance posture. It is the difference between explaining why a system said something and being able to show what it predicted, planned, and did.

The governance catch

The same auditability argument cuts both ways. If the world model is the decision engine, then the world model itself becomes the critical control surface. Who validates that the simulation is right? Who tests the model’s representation of drug interactions, of physiological trajectories, of the edge cases that never made it into training data? We have two decades of hard-won lessons about why black-box probabilistic systems are hard to certify. We have zero decades of lessons about certifying learned world simulators in regulated environments.

NIST’s AI RMF and the emerging regulatory stack for clinical AI are built around documented, testable, governable system behavior. World models do not escape that requirement - they just change what documentation and testing have to cover. “It simulated the outcome first” is not a control until someone defines the assurance case for the simulation itself.

What I would require before it touches a clinical workflow

The Nabla partnership will put world-model reasoning next to real patient data. Here is the short list of controls I would demand as the person who has to sign off on it:

• An assurance case for the simulator itself. Defined boundaries of validity, known failure modes, documented training data lineage. “It plans against a simulation” is only a control if someone can show what the simulation is wrong about.
• Retrospective validation before prospective use. Golden sets of patient trajectories where the model’s predicted evolution is scored against what actually happened. If the prediction record is not measurably better than the current baseline, the audit trail is decoration.
• Change control on the model, not just the software. Drift monitoring, version pinning, and rollback paths for the world model itself - the same discipline we apply to PACS upgrades and EMR patching.
• Human breakpoints at decision boundaries that matter. Agentic scheduling and referral coordination can run with oversight. Anything adjacent to medication, diagnosis, or discharge criteria needs a human signature, and the workflow has to be built so the signature is cheap, not ceremonial.
• Data governance that understands what the training data is. Video, audio, and physiological streams captured in clinical environments are not ordinary telemetry - they carry PHI implications from the moment of capture, and the training pipeline needs the same HIPAA boundaries as the EMR estate.

None of this is exotic. It is the NIST AI RMF mapped onto hospital reality: govern, map, measure, manage. The difference is that with world models, the “measure” function has a concrete artifact to measure - the simulated decision trail - instead of a probabilistic explanation nobody can reproduce.

The infrastructure angle nobody is pricing in

There is also a quieter story here for the infrastructure people, and it is the one I keep coming back to.

World models learn from video and sensor streams, not text. The compute profile is categorically different from LLM training - potentially heavier, because modeling physical environments at scale means processing dimensionalities that make token streams look like telegrams. The entire industry buildout of the last three years - the H100 fleets, the gigawatt data centers, the inference economics - is optimized for a text-first architecture.

If the JEPA bet pays off, some meaningful fraction of that capital is positioned on the wrong workload. Notice who is on AMI’s cap table: NVIDIA. That is not a passive investment. That is the dominant AI infrastructure vendor buying a hedge against the architecture that made it dominant.

The skeptic’s ledger

To be clear about what AMI has actually shown so far: roughly a dozen employees, no product, a first year explicitly dedicated to research, and timelines the CEO himself describes in years, not quarters. The $3.5 billion pre-money valuation is a bet on scientific credibility and option value, and it should be analyzed as such. Fei-Fei Li’s World Labs raised $230 million at a $1 billion valuation and was called aggressive for a pre-product raise. AMI cleared 3.5x that on day one.

Maybe that is a bubble premium on a famous name. Or maybe LeCun is right - right in the way he has been right before, when the CNN papers looked like an interesting aside and became the foundation of computer vision.

The bottom line

I do not know whether world models will beat scaled LLMs. Neither does anyone else, including the people who just wrote the nine-figure checks. But the operational question does not wait for the research to resolve.

The systems are coming either way - agentic AI in clinical workflows, in industrial control, in places where hallucination is not an embarrassment but a hazard. The organizations that will handle that transition well are not the ones that pick the winning architecture. They are the ones that demand an audit trail from whatever architecture shows up, test it like the control surface it is, and refuse to confuse capability with governance.

That part is not speculation. That is just the job, and it is about to get more interesting.