Yann LeCun's $1B AMI Labs Raise and Paris's AI Boom

A big claim attached to a big name can be inflated, and it can also be true.

A French AI scientist. A billion-dollar number. A Paris startup. In March 2026 those three things came together for real: AMI Labs, the company Yann LeCun started after leaving Meta, raised $1.03 billion at a $3.5 billion pre-money valuation, as TechCrunch reported. An earlier version of this article, published a day later, called the billion-dollar story a myth and said LeCun was still at Meta. That was wrong, and this version corrects it. The other story that version told, about H, the Paris startup that raised a $220 million seed round in 2024, still matters, and read side by side the two raises say a lot about where AI money is going in Paris.

The first version of this article followed that lead with two sentences, kept here as written. I've seen this pattern play out a dozen times and it never stops being frustrating. Someone sees a French AI name, a big number, and a Turing Award winner, mashes them together, and suddenly it's gospel. The pattern is real. This time the mash-up was in that first version, not in the headlines.

Who Is Yann LeCun, and What Is AMI Labs?

Yann LeCun led Meta's AI research for more than a decade. He won the 2018 Turing Award (announced in March 2019) alongside Geoffrey Hinton and Yoshua Bengio for foundational work on deep learning. He's a professor at NYU. He posts prolifically on social media about AI philosophy, open-source models, and why he thinks autoregressive LLMs are a dead end.

He left Meta in 2025 to build a company around that view, as SiliconANGLE reported. That company is AMI Labs. It is headquartered in Paris with offices in New York, Montreal and Singapore, run by CEO Alexandre LeBrun with LeCun as chairman, and its $1.03 billion round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions. AMI is building world models: AI that learns from camera and sensor data about the physical world rather than from text alone, and that deliberately avoids the generative approach behind today's chatbots.

The first version also closed this section with these lines, kept here as written. As someone who's spent over a decade in software engineering and tries to stay on top of AI developments daily, I find these attribution errors genuinely damaging. They drown out the real story, which in this case is way more compelling than "famous scientist does predictable thing." There was no attribution error to complain about: the famous scientist really did raise the money, and that is part of the real story.

What Is H, and Why Did It Raise $220 Million at Seed?

Two years before AMI's round, a Paris-based AI startup called H raised $220 million in a seed funding round in 2024, as reported by Romain Dillet at TechCrunch. That's not a Series A. Not a Series B. A seed round. Most seed rounds for even well-connected AI startups land between $2 million and $20 million. H raised more than ten times the upper end of that range before shipping a single product.

The founding team read like a Google DeepMind reunion: CEO Charles Kantor alongside Laurent Sifre, Karl Tuyls, Daan Wierstra, and Julien Perolat, researchers with deep backgrounds in reinforcement learning, game theory, and large-scale AI systems. The reunion did not last. Tuyls, Wierstra and Perolat left in August 2024, three months after the raise, over what were described as operational differences, and in June 2025 former Palantir France CEO Gautier Cloix replaced Kantor as CEO, according to Wikipedia's history of the company.

The investor list tells the same story. As reported by Sifted, the round included Accel, UiPath, Bpifrance (France's public investment bank), former Google CEO Eric Schmidt, Amazon, Samsung, and French luxury billionaire Bernard Arnault. When Eric Schmidt and Amazon are both writing checks to the same seed-stage company, that's not normal investor enthusiasm. That's a signal that people with deep technical knowledge believe something fundamental is happening.

What Are "Action Models" and Why Should Engineers Care?

Here's where it gets technically interesting. H isn't building another chatbot. Their stated mission is to build what they call "action models" — AI systems that don't just generate text or images but actually do things in the real world. Think of it as the difference between an AI that can write a perfect plan for deploying a Kubernetes cluster and an AI that can actually execute the deployment, handle the errors, and adapt when something goes wrong.

The gap between an AI that can describe how to do something and an AI that can actually do it is the most important unsolved problem in the field right now. I've been building with multi-agent AI systems for the past year, and the biggest limitation I keep hitting is exactly this: current models are brilliant at reasoning and terrible at acting. They can plan a ten-step workflow, then fail catastrophically at step three because they can't handle an unexpected API response.

H's bet is that this gap requires entirely new model architectures, not just better prompting or fancier tool-calling frameworks. Their approach moves from "multimodal" (understanding text, images, audio) to "actional" (taking reliable, grounded actions based on that understanding). If they pull it off, it would be a real leap toward what the industry loosely calls AGI — not in the sci-fi sense, but in the practical sense of AI systems that can be trusted to act autonomously in complex environments.

This is directly relevant to anyone building agentic AI systems. The current generation of AI agents is limited by models that were trained to predict text, not to act. A model purpose-built for action could change the entire stack.

If you’re curious what this looks like in practice, I tried to push today’s agent tooling to an extreme in Paperclip AI Review: I Tried to Build a Zero-Human Company in a Weekend [2026]. It’s a concrete experiment in where “action models” and automation actually work today—and where they still break.

Why Paris Keeps Producing Record AI Raises

H's raise didn't happen in a vacuum. It follows Mistral AI's rapid ascent — another Paris-based startup that went from founding to a $2 billion valuation in about a year. Add AMI Labs' $1.03 billion round in 2026 and France has three of the best-funded AI startups in the world, none of them trying to be an OpenAI clone.

There are structural reasons for this. France has invested heavily in AI research infrastructure. Institutions like INRIA and École Polytechnique produce world-class machine learning talent. The French government has been actively courting AI companies with favorable policies and research grants through Bpifrance. And critically, the cost of hiring a senior ML researcher in Paris is still significantly lower than in San Francisco or New York.

Having worked with engineering teams distributed across North America and Europe, I've noticed something the Silicon Valley-centric tech press keeps missing: the talent density in Paris for fundamental AI research is exceptional. DeepMind has a major Paris office. Meta has a significant AI research lab there — the one LeCun helped establish. Google, Microsoft, and Hugging Face all have substantial Paris-based teams.

Paris isn't trying to be the next Silicon Valley. It's becoming the global center for a specific kind of AI work: foundational model research with a European sensibility toward open science and safety. Deep research talent, government backing, and now massive private capital. That's not hype. That's infrastructure.

How This Article First Got the LeCun Story Wrong

This article is its own cautionary tale. Its first version went up a day after the AMI Labs raise was reported, treated the billion-dollar headlines as a mash-up of separate French AI stories, and declared them false. The headlines were right. The mistake was treating "this sounds like hype" as evidence that it was hype, instead of checking the funding announcement itself. That instinct is usually sound, which is what made it so easy to trust here.

I've shipped enough features and reviewed enough technical claims to know that the distance between a viral headline and reality is usually enormous. In AI specifically, this gap has become dangerous. Investors make decisions based on vibes. Engineers chase trends based on Twitter threads. And the actual innovations — like H's work on action models — get buried under celebrity gossip.

If you're an engineer trying to stay informed about where AI is heading, use this as a calibration check that cuts both ways. A big claim attached to a big name can be inflated, and it can also be true. Go to the primary source: who built it, who raised the money, and what they are building.

The next two years in AI aren't going to be defined by chatbots getting slightly better at writing emails. They're going to be defined by whether companies like H and AMI Labs can crack the action problem — building AI that doesn't just think, but reliably does. That's a harder story to tell in a tweet.

If you're building AI agents today, start paying attention to Paris. The models that power your systems tomorrow might not come from San Francisco at all.