Yann LeCun's AMI Labs Raises $1.03B to Build Beyond LLMs

AMI Labs, the Paris-based AI company founded by Yann LeCun, closed a $1.03 billion seed round in March 2026 at a $3.5 billion pre-money valuation, making it the largest prelaunch AI funding in European history, according to TechCrunch. The raise arrived four months after LeCun left Meta, where he had served as chief AI scientist since 2013.

The Exit from Meta

LeCun’s departure was confirmed on November 19, 2025, reported by CNBC. He had joined Facebook (later Meta) in 2013 to build what became Meta FAIR (Fundamental AI Research), one of the largest academic-style industrial AI labs in the world. His tenure covered foundational work on convolutional neural networks, open-source model releases through the Llama family, and the origination of the JEPA research program.

The split reflected a growing divergence over architectural bets. By late 2025, Meta had reorganised FAIR into the Meta Superintelligence Labs, completed a $15 billion acquisition of Scale AI, and oriented its product strategy firmly around generative models and Llama. LeCun’s public criticism of LLM limitations had sharpened in step. Speaking to The Decoder on his departure, he was direct: “You certainly don’t tell a researcher like me what to do.”

AMI Labs (Advanced Machine Intelligence) is headquartered in Paris. LeCun holds the title of executive chairman; he appointed Alexandre LeBrun, founder of French health-tech firm Nabla, as CEO. According to Fortune, Meta will not invest in AMI Labs, though the two organisations plan to maintain a research partnership allowing LeCun to continue JEPA-related collaboration with former FAIR colleagues.

A Record European Raise and Who Backed It

The $1.03 billion round was co-led by five institutional investors: Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. Angel investors named in the TechCrunch disclosure include Tim Berners-Lee and Rosemary Berners-Lee, Jim Breyer, Mark Cuban, Mark Leslie, Xavier Niel, and Eric Schmidt.

For comparison, Mistral AI’s €113 million ($124 million) seed round in 2023 had been the record for a European AI startup at seed stage. AMI Labs’ $1.03 billion surpasses it by a factor of roughly 8.3. The $3.5 billion pre-money valuation implies a post-money figure of approximately $4.53 billion before the company had shipped any commercial product.

The investor composition is notable across several dimensions. Bezos Expeditions places Jeff Bezos alongside LeCun; Xavier Niel and Cathay Innovation signal confidence from the French tech ecosystem in AMI’s Paris anchor. Eric Schmidt’s participation continues a pattern of backing post-lab-exit AI founders, having previously invested in Mistral AI and taken positions in several former Google DeepMind researchers’ companies. The Berners-Lee participation is unusual for a seed round at this scale and reflects a long-standing alignment between LeCun’s open research philosophy and Berners-Lee’s public-interest stance on technology.

What V-JEPA 2 Is Actually Doing

AMI Labs inherits a concrete research lineage. Meta FAIR published the first Image-JEPA paper in 2022, then Video-JEPA in 2024. V-JEPA 2, published on arXiv in June 2025, extended the architecture with a 1.2-billion-parameter model trained on more than one million hours of internet video and one million images, then fine-tuned on a small set of robot trajectory data.

JEPA’s mechanism is architecturally distinct from the generative models dominant in commercial AI. Rather than predicting the next token or reconstructing masked pixels, a JEPA model ingests two related views (for example, consecutive video frames) and tries to predict the abstract latent representation of one from the other. It learns what the world does without being forced to hallucinate what every pixel looks like. LeCun described the logic in a January 2026 interview with MIT Technology Review: “We are going to have AI systems that have humanlike and human-level intelligence, but they’re not going to be built on LLMs, and it’s not going to happen next year or two years from now. It’s going to take a while. There are major conceptual breakthroughs that have to happen.” On JEPA itself: “JEPA is not generative AI. It is a system that learns to represent videos really well.”

V-JEPA 2’s published results include 77.3% top-1 accuracy on Something-Something v2, a benchmark focused on physical motion understanding; 39.7 recall-at-5 on Epic-Kitchens-100 for human action anticipation; and an 80% success rate on a robotic cup-moving task using an action-conditioned variant of the model. V-JEPA 2.1, with sharper temporally consistent dense features, shipped in March 2026 alongside the funding close. For a broader view of the world-models research space, several other labs are pursuing related approaches, though none with the same funding base or the same direct lineage to the original JEPA papers.

What It Means

AMI Labs is the best-funded direct challenge to the LLM architectural consensus in AI today. LeCun’s record at Meta FAIR lends technical credibility that most seed-stage founders cannot claim; V-JEPA 2’s robotics benchmarks provide a first concrete demonstration that JEPA-based systems can reach application-layer performance. Peers such as Demis Hassabis at DeepMind are pursuing AI architectures that go beyond current generative models as well, but through hybrid scientific-discovery pipelines rather than through a direct challenge to the token-prediction paradigm. Whether AMI Labs compounds on V-JEPA 2.1 or pivots toward a hybrid approach is the central research question; LeCun has consistently maintained that the path forward requires architectural breakthroughs that do not yet exist.

Sources

• TechCrunch: Yann LeCun’s AMI Labs raises $1.03B to build world models
• CNBC: Meta chief AI scientist Yann LeCun is leaving the company
• Fortune: Yann LeCun is targeting a $3.5 billion valuation for his new startup
• MIT Technology Review: Yann LeCun’s new venture is a contrarian bet against large language models
• The Decoder: “You certainly don’t tell a researcher like me what to do” — LeCun on exiting Meta
• arXiv: V-JEPA 2 — Self-Supervised Video Models Enable Understanding, Prediction and Planning

Editorial standards: every claim is sourced. Tips: [email protected]