AI News

AMI Labs: Yann LeCun leaves Meta and raises 1 billion for world models

Turing Award winner Yann LeCun leaves Meta to found AMI Labs in Paris. His $1.03 billion raise funds world models, a radical alternative to LLMs. What it changes for SMBs.

5 min read
World ModelsAMI LabsYann LeCunJEPARobotiqueIA Europe
âš¡ The news in 30 seconds

AMI Labs: Yann LeCun leaves Meta and raises 1 billion to reinvent AI

On March 10, 2026, Yann LeCun — 2018 Turing Award winner and Meta's chief AI scientist for 12 years — announced the creation of AMI Labs, a startup based in Paris. In just four months, AMI Labs raised $1.03 billion from Nvidia, Jeff Bezos, Eric Schmidt and the Singaporean fund Temasek, at a valuation of $3.5 billion. The bet: world models, AI systems that learn from physical reality rather than text, will replace large language models for critical applications.

The father of deep learning is betting that LLMs are a dead end for AI that must understand the real world. His Paris startup is attracting the largest seed round in European history — a strong signal for the continent's AI sovereignty.

What it changes for you

✦ The opportunity

AMI Labs represents a strategic turning point for the European AI ecosystem. For the first time, a world-class European startup is positioning itself as a credible alternative to the American and Chinese giants — with a fundamentally different technological approach.

🇫🇷

Cutting-edge AI made in France

AMI Labs is based in Paris, with offices in New York, Montreal and Singapore. It is the first time a European AI startup has reached a $3.5 billion valuation at seed stage. For French SMBs, this means a strengthening local AI ecosystem, with talent, partnerships and solutions developing on their own territory.

🤖

Robotics and physical automation

World models are designed to understand the physical environment — object trajectories, the geometry of spaces, causal relationships. For industrial, logistics or craft SMBs, this is the promise of robots and AI assistants able to adapt to real environments, not just spreadsheets and emails.

🔬

Beyond hallucinations

LeCun openly criticizes LLMs for their hallucinations — invented information that can be catastrophic in health, law or finance. World models, by learning the structure of reality rather than textual statistics, aim for a more robust and reliable understanding. If the promise holds, it is a game-changer for regulated sectors.

âš  The risk

⚠️

Still-experimental technology

So far, world models have only demonstrated their effectiveness in controlled robotic environments. AMI Labs has no commercial product and won't have one for at least 12 to 18 months. CEO Alexandre LeBrun says it himself: "This is not a classic AI startup that ships a product in three months." Betting your entire AI strategy on this approach would be premature.

🔒

The "world models" bubble effect

In two months, $1.3 billion was invested in the "world models" category. The CEO of AMI Labs predicts that "in six months, every startup will call itself a world model to raise funds." This kind of frenzy is reminiscent of previous hype cycles. SMBs must stay vigilant and evaluate these technologies on concrete results, not on promises.

Our recommendation

AMI Labs is a strategic signal to watch, not a call for immediate action. Here is how to position yourself intelligently:

1

Keep using LLMs for your current needs

Large language models remain the best solution for text use cases: summarizing, writing, document analysis, chatbots. Don't stall your ongoing AI projects waiting for world models.

2

Identify your "physical" use cases

If your business involves robotics, logistics, visual quality control or predictive maintenance, world models could become relevant. Take an inventory of your processes that involve understanding the physical environment.

3

Follow the ecosystem and keep your options open

AMI Labs, Fei-Fei Li's World Labs, and other players will accelerate in 2026-2027. Adopt a modular AI architecture that will let you integrate new models without rebuilding everything. Flexibility is your best investment.

In summary

Opportunity
World-class European AI startup in Paris, a radically different approach from LLMs, a strengthened French AI ecosystem
Risk
Experimental technology, no product for 12-18 months, bubble risk around "world models"
Recommended action
Keep running LLM projects, identify physical use cases, monitor the world models ecosystem
Horizon
First AMI Labs products expected late 2027 — medium-term impact for SMBs

Frequently asked questions

What is a world model?

A world model is an AI system that learns to understand physical reality from video, audio and sensor data, instead of predicting the next word in text like LLMs. It builds an internal representation of physics, geometry and cause-and-effect relationships — a bit like a child learning by observing the world around them.

Will AMI Labs replace ChatGPT?

Not in the short term. World models and LLMs address different needs. LLMs excel at language processing. World models aim at understanding the physical environment. Eventually, the two approaches could converge, but AMI Labs will not have a commercial product for at least 12 to 18 months.

Why does it matter for European SMBs?

AMI Labs is based in Paris, with European investors and a vision of technological sovereignty. If world models live up to their promise, European SMBs will have access to cutting-edge AI developed on their own soil. It is also a strong signal for the European AI ecosystem.

Should you invest in world models right now?

No. The technology is still in fundamental research. Keep leveraging LLMs for text use cases and prepare to integrate these new approaches when they reach commercial maturity, probably during 2027.

For technical profiles

JEPA architecture and positioning

JEPA

Joint Embedding Predictive Architecture

Unlike generative models that predict every pixel or every word, JEPA predicts future evolution in an abstract representation space. The model learns the underlying physical rules while ignoring random noise. Trained on millions of hours of video, audio and sensor data. V-JEPA 2 demonstrated zero-shot robotic control at Meta before LeCun's departure.

AMI Video

First planned commercial model

AMI Labs is developing AMI Video, a world model trained on large-scale video. It targets industrial robotics applications, autonomous vehicles and contextual assistants on smart glasses. Infrastructure: massive GPU clusters to process millions of hours of video. First results expected late 2026.

Comparison of AI approaches

Classic LLMsWorld models
Training dataText, web corporaVideo, audio, sensors
MethodNext-word predictionPrediction in abstract space
StrengthLanguage, writing, dialoguePhysical understanding, robotics
WeaknessHallucinations, no causal understandingImmature, no commercial product
PlayersOpenAI, Anthropic, Google, MetaAMI Labs, World Labs, DeepMind
MaturityProductionResearch

Related articles