AMI Labs, the new AI venture cofounded by Turing Award laureate Yann LeCun, has secured $1.03 billion at a $3.5 billion pre-money valuation to pursue “world models” — systems that learn how the physical world works from rich, real data rather than text alone. The raise signals deep investor conviction that the next leap in AI will come from grounded understanding and predictive reasoning, not larger language models.
Why World Models Matter For Grounded, Predictive AI
World models aim to internalize cause and effect, perception, and planning by training on video, audio, and interactive signals. The approach contrasts with LLMs that excel at language but can falter when asked to reason about dynamics in the real world, where hallucinations carry risks in fields like healthcare and robotics.
LeCun has long argued for self-supervised learning as the path to machine intelligence, proposing the Joint Embedding Predictive Architecture (JEPA) in 2022 to predict missing information in high-dimensional inputs. Similar lines of research underpin notable advances such as DeepMind’s MuZero and the Dreamer family of agents, which demonstrate that learning internal models can improve planning and control.
AMI’s first disclosed partner is Nabla, a digital health company focused on clinical productivity. The early goal is to test model behavior on real workflows and real evaluations, a necessary step if these systems are to support decision-making where accuracy and calibration are nonnegotiable.
A Research-First Plan And Global Footprint
Unlike typical applied AI startups, AMI is positioning as a deep, multi-year research program. The company says it will prioritize quality over headcount and hire selectively across four hubs: Paris as headquarters, New York alongside LeCun’s academic base at NYU, Montreal with a strong learning theory community, and Singapore for both talent and proximity to Asian customers.
The scientific bench is stacked: former Meta leaders and prominent academics including Laurent Solly as COO, Saining Xie as chief science officer, Pascale Fung as chief research and innovation officer, and Michael Rabbat as VP of world models. The roadmap centers on large-scale self-supervised training over video and multimodal streams, followed by rigorous real-world evaluation outside the lab.
Two cost centers dominate — compute and talent. Training high-capacity predictive models over long video horizons can demand massive GPU clusters and careful data curation. Industry estimates suggest state-of-the-art multimodal training can run to tens of thousands of accelerator-days, making strategic access to hardware as pivotal as algorithmic breakthroughs.
A Cap Table Built For Compute And Scale
The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions. A long list of strategic and individual backers joined, including Tim and Rosemary Berners-Lee, Jim Breyer, Mark Cuban, Mark Leslie, Xavier Niel and Eric Schmidt.
Corporate and industry-linked investors include NVIDIA, Samsung, Sea, Temasek and Toyota Ventures, alongside French groups Association Familiale Mulliez, Groupe Industriel Marcel Dassault and Publicis Groupe. Aglaé Lab, Alpha Intelligence Capital, Artémis, Bpifrance Digital Venture, New Legacy Ventures, SBVA and ZEBOX Ventures also participated.
Beyond capital, this coalition is designed to de-risk core bottlenecks: access to accelerators and systems engineering, routes to data partnerships, and downstream channels in sectors such as mobility, consumer devices and healthcare. For a company betting on long-horizon R&D, those advantages matter as much as the headline number.
Rivals And Momentum In World Modeling And Grounded AI
AMI is not alone in this bet. Fei-Fei Li’s World Labs recently assembled a $1 billion war chest for similar work, and European startup SpAItial drew attention with an unusually large $13 million seed. Earlier reports suggested AMI was targeting roughly €500 million; the final close near €890 million underscores how quickly interest has accelerated around grounded AI.
The resurgence of model-based learning reflects a broader shift: as text-only scaling shows diminishing returns on reasoning and planning tasks, researchers are turning to predictive objectives on video and interaction. Robotics benchmarks, autonomous systems, and clinical decision support stand out as early proving grounds where structured generalization and uncertainty estimation are essential.
Open Science As Strategy For Faster, Higher-Quality AI
AMI plans to publish research and release significant portions of its code as open source, echoing the culture LeCun helped shape at FAIR and the communities that formed around PyTorch. While many frontier labs have shifted to closed releases, AMI is wagering that openness accelerates progress, sharpens peer review and builds a stronger recruiting magnet.
That stance could also speed standardization for evaluating world models — for example, long-horizon video prediction, causal reasoning under intervention, and controllable simulation fidelity — areas where public benchmarks remain immature. Expect AMI to push for transparent metrics that move beyond next-token accuracy.
What To Watch Next As AMI Builds Large-Scale World Models
Early signals of traction will likely include state-of-the-art results on video prediction and planning tasks, evidence of robust calibration under distribution shift, and pilot deployments with partners like Nabla. Another key milestone will be hybrid models that fuse language with learned world dynamics, marrying instruction-following with grounded prediction.
With $1.03 billion in fresh funding, a deep bench of researchers and a cap table engineered for compute, AMI Labs has the resources to test whether world models can deliver the next step-change in AI capability. The question now is not just if the theory holds, but how quickly it can be turned into dependable systems that navigate the messy constraints of the real world.
