Yann LeCun’s AMI Labs Raises $1 Billion for World Models in Bet Against LLMs
Yann LeCun, recipient of the 2018 Turing Award, is one of the leading figures in AI research—and is now doing his own thing. Following a strategic disagreement with his former boss Mark Zuckerberg (LeCun sees World Models as more promising than the constant, expensive improvement of LLMs through massive computing power), he has now established AMI Labs with headquarters in Paris.
Advanced Machine Intelligence Labs (AMI) has now raised $1.03 billion (approximately €890 million) in a seed financing round. This is the largest seed financing in Europe and the second-largest worldwide after the US startup Thinking Machines Lab, which raised two billion dollars in June 2025. Just recently, World Labs, founded by AI researcher Fei-Fei Li, raised $1 billion to expand its world models.
The new company is valued at a pre-money valuation of $3.5 billion.
Leadership Team and Organization
AMI Labs is led by Alexandre LeBrun, the former CEO of French startup Nabla. Yann LeCun, Meta’s former Chief AI Scientist and Turing Award recipient, takes on the position of Executive Chair. Laurent Solly, former Vice President of Meta for Europe, becomes Chief Operating Officer.
The company starts with a team of approximately twelve employees and researchers, distributed across four locations from the outset: Paris, New York, Montreal, and Singapore. AMI Labs sees itself as a global company that deliberately wants to tap into talent outside Silicon Valley.
Investors and Strategic Partners
The financing round is jointly led by:
- Cathay Innovation (France)
- Greycroft
- Hiro Capital
- HV Capital
- Bezos Expeditions (Jeff Bezos)
Strategic investors and long-term supporters include Toyota Ventures, Temasek (Singapore), SBVA (Seoul), NVIDIA, Mark Cuban, Sea, and Alpha Intelligence Capital. Additional significant investments come from Eric Schmidt, Samsung, Bpifrance Digital Venture, Jim Breyer, Tim and Rosemary Berners-Lee, and Mark Leslie.
Meta itself is not an investor, but will enter into a partnership with AMI Labs that grants the technology company access to the developed technology for commercialization. The details of this collaboration are still being worked out.
Yann LeCun’s Vision and Role
Yann LeCun, a French-American scientist and one of the leading minds in AI research, has repeatedly argued that systems trained primarily on text will struggle to achieve human-like reasoning. His new role at AMI Labs enables him to put this conviction into practice.
“We share one belief: real intelligence does not start in language. It starts in the world.”
The startup builds on LeCun’s research work at Meta, where he worked on new AI architectures that can learn from videos and spatial data rather than just language. AMI Labs is therefore focused on developing so-called World Models instead of Large Language Models, and alongside World Labs and Waymo is one of the leading companies in the field.
Applications and Partnerships
AMI Labs plans to deploy its technology in areas where reliability, controllability, and safety are particularly important. These include industrial process control, automation, wearable devices, robotics, and healthcare.
LeBrun’s former company Nabla will be AMI Labs’ first partner and will apply the new models in the healthcare industry. CEO LeBrun emphasizes that the company needs at least a year of research before introducing the first real-world applications.
World Models versus Large Language Models
The central difference between AMI Labs’ approach and current AI systems lies in the fundamental architecture. While Large Language Models (LLMs) like ChatGPT are trained primarily on text data and learn by predicting the next word, AMI develops so-called World Models.
World Models learn abstract representations of sensor data from the real world, ignore unpredictable details, and make predictions in the representation space. These models are designed to:
- Understand the physical environment
- Possess persistent memory
- Plan and reason
- Predict the consequences of actions
LeBrun explains the approach clearly: “For anything that requires understanding the real world, we believe that Large Language Models and generative AI in general are not the right solution.”
The problem with generative approaches for real sensor data lies in their unpredictability. While these methods have been extraordinarily successful with language, they work less well with continuous, high-dimensional, and noisy data from cameras or other sensors. World Models circumvent this problem by operating at a more abstract level and enabling action-conditioned predictions that are crucial for autonomous systems, robotics, and transportation applications.
