Yann LeCun's JEPA Cadence: Four Models in Three Years Across Two Labs

Since leaving Meta in November 2025, Yann LeCun has released V-JEPA 2.1, delivered five keynotes, co-authored two arXiv papers, and closed a $1.03 billion seed round. Here is the full output rate.

6 min read
Yann LeCun, JEPA release cadence and AMI Labs founding, 2026
Yann LeCun at a meeting with Henna Virkkunen, European Commission Executive Vice-President, 2026.· Photo by Nicolas Kovarik / European Union 2026 / EC Audiovisual Service, via Wikimedia Commons (CC BY 4.0)

Since January 2024, the team behind the Joint Embedding Predictive Architecture (JEPA) has shipped a major new version roughly every twelve months, a pace that continued without interruption even as Yann LeCun left Meta in November 2025 to found Advanced Machine Intelligence Labs. (arXiv, TechCrunch)

A JEPA release roughly every twelve months since 2023

The Joint Embedding Predictive Architecture traces back to I-JEPA in 2023, LeCun's first public demonstration that a model could learn useful image representations by predicting missing patches in a latent space, without pixel reconstruction or contrastive learning. V-JEPA, released in January 2024, extended the same architecture to video, producing a model that could anticipate what happens next in a scene rather than describe it in words.

V-JEPA 2, released by Meta FAIR in June 2025 and detailed in arXiv 2506.09985, marked a significant capability step. The paper reported 77.3% top-1 accuracy on the Something-Something v2 motion-understanding benchmark and 39.7 recall-at-5 on Epic-Kitchens-100, a standard test for human action anticipation where the authors reported state-of-the-art performance. V-JEPA 2-AC, the action-conditioned variant, could solve robot manipulation tasks without environment-specific training data, an early sign that the architecture was moving toward physical-world control.

V-JEPA 2.1, released in March 2026 alongside the AMI Labs seed announcement, extended the lineage into LeCun's new institutional home. The gap between V-JEPA and V-JEPA 2 was seventeen months; between V-JEPA 2 and V-JEPA 2.1, nine months. AMI Labs describes the architecture as the foundation for AI systems that "understand and interact with the physical world" rather than generate text or images. (TechCrunch, March 2026)

Five keynotes and two papers in eight months

Between January and August 2026, LeCun delivered at least five major keynotes: the Davos "Imagination in Action" event in January, the Brown University Lemley Family Leadership Lecture in April, an ETH Zürich "Frontiers of Embodied AI" talk in June, the AI Impact Summit, and the Machines Can Think conference in Abu Dhabi. At the AI Impact Summit, he described AI as "an amplifier for human intelligence," a position consistent with his long-standing view that current AI systems will extend human capability rather than replace cognitive labor wholesale in the near term. (Storyboard18, Brown University)

On the research side, LeCun's group posted at least two arXiv preprints in late May and June 2026 addressing when JEPA-based world models can provably learn a faithful representation of an environment. One, arXiv 2606.27014, titled "A Generalization Theory for JEPA-Based World Models," establishes formal conditions under which the architecture converges and identifies where current implementations fall short. The pairing of formal theory with empirical model releases has been characteristic of LeCun's FAIR output, and the same pattern is visible in the first eight months at AMI Labs.

StartupHub.ai data shows 96 startups in the physical and embodied AI space actively building on world-model approaches, spanning industrial automation, humanoid robotics, autonomous vehicles, and healthcare. The category has attracted capital well beyond LeCun's own company: firms like Wayve and AMI Labs represent the larger end of a cohort that increasingly frames its architecture around the prediction-in-latent-space ideas LeCun has championed in public since 2022.

AMI Labs: the $1.03 billion institutional backbone

AMI Labs closed its seed round on March 10, 2026, raising $1.03 billion at a $3.5 billion pre-money valuation, the largest seed round in European history at the time. (Sifted, TechCrunch) The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. Corporate strategics, including NVIDIA, Temasek, Samsung, Toyota Ventures, and Bpifrance, also participated alongside individual investors Mark Cuban, Eric Schmidt, and Tim Berners-Lee.

The investor mix reflects the breadth of the commercial thesis. NVIDIA's participation signals that AMI Labs' training workloads are viewed as viable at GPU scale. Toyota Ventures' involvement suggests automotive and robotics deployment timelines that do not require waiting for AGI. Bpifrance's backing gives the company access to French and EU industrial partnerships, consistent with AMI Labs' Paris base. LeCun serves as executive chairman and continues as Silver Professor at NYU's Center for Data Science, a dual-role he held during his Meta tenure as well.

The company's near-term focus is on deployments in industrial, robotic, and healthcare settings where LLM limitations, particularly the inability to model sequences of physical cause and effect, are commercially consequential. AMI Labs' profile on StartupHub reflects the $1.03 billion seed as its primary verified funding figure.

What it means

The cadence across JEPA model versions, arXiv preprints, and conference appearances does not reflect a researcher decelerating during a career transition. LeCun's output rate from November 2025 through August 2026, measured across model releases, formal papers, and keynotes, resembles the pace he maintained at Meta FAIR. The primary difference is institutional: the JEPA roadmap now runs under a company with $1.03 billion in seed capital and a board drawn from enterprise software, automotive, sovereign wealth, and GPU supply-chain investors, rather than inside a consumer internet company's research division. Whether AMI Labs moves from research cadence to commercial cadence on a comparable schedule will define the arc of this next phase.

Sources

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

© 2026 StartupHub.ai. All rights reserved. Do not enter, scrape, copy, reproduce, or republish this article in whole or in part. Use as input to AI training, fine-tuning, retrieval-augmented generation, or any machine-learning system is prohibited without written license. Substantially-similar derivative works will be pursued to the fullest extent of applicable copyright, database, and computer-misuse laws. See our terms.