Researchers who spent a decade scaling language models spent 2026 on a different bet. The target: systems that learn how physical space behaves1. Yann LeCun left Meta on Nov. 20, 2025. By March he had raised $1.03 billion for AMI Labs at a $3.5 billion pre-money valuation. Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions co-led. Fei-Fei Li's World Labs, Physical Intelligence, and Figure AI chased the same thesis234. So did a venture from four Google veterans7. Billions moved toward world models as 2026's answer to a question the field had started asking quietly: what comes after transformer scaling plateaus?
LeCun raised $1.03 billion for AMI Labs on top of a $3.5 billion pre-money valuation by March 2026
AMI Labs' March 2026 round, billions of dollars: the pre-money valuation and the new money raised
The bar is AMI Labs' value once the round closed: the blue segment is the $3.5 billion pre-money valuation, and the magenta segment is the $1.03 billion investors put in.
Source: TechCrunch [1]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value | Note |
|---|---|---|
| Pre-money valuation | $3.5 billion | |
| New money raised | $1.03 billion | Co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions |
LeCun's Second Act
LeCun spent years arguing in public that large language models sat on a dead-end path, and that genuine machine intelligence would have to come from somewhere else entirely1. His exit from Meta let him test that argument with his own capital and hires. AMI Labs confirmed its existence in December 2025, and investors backed the thesis faster than his former employer ever had. Nvidia, Samsung, Toyota Ventures, and Bpifrance joined the round. Angels included Tim Berners-Lee, Mark Cuban, and Eric Schmidt.
Chief executive Alexandre LeBrun runs day-to-day operations under LeCun's technical direction1. He set the ambition plainly: "AMI Labs is a very ambitious project, because it starts with fundamental research". LeBrun also refuses the field's favorite vocabulary. "My prediction is that 'world models' will be the next buzzword," he said. "In six months, every company will call itself a world model to raise funding". Read that as a man naming the trap he is standing inside.
LeBrun drew another line on openness: "We are developing world models that seek to understand the world, and you can't do that locked up in a lab"1. A promise followed: "We will also make a lot of code open source". The distinction matters strategically. Selling world understanding as a narrower, testable capability invites less scrutiny. Claiming a path to general intelligence invites more.
Fei-Fei Li Builds Marble
Li's World Labs took a parallel path with a longer runway2. Justin Johnson, Ben Mildenhall, and Christoph Lassner joined her at launch. They brought computer-vision and graphics credentials from academic and industry work. The company emerged from stealth with $230 million in September 2024. Ashton Kutcher's Sound Ventures came in that October. Marble, its first commercial product, shipped in November 2025, roughly fourteen months after the company left stealth.
Marble generates spatially coherent, persistent 3D worlds from images, video, text, or layout inputs2. The technology aims at one gap. A flat generated image sits on one side, and an environment an agent or a person can navigate sits on the other. Momentum culminated in a $1 billion valuation round in February 2026. Autodesk alone contributed $200 million, joining Nvidia, Adobe, Intel, Samsung, Salesforce, and Databricks. Corporate investors rarely write checks that large absent a product roadmap. That stake signals what design and engineering software vendors expect. World models will reshape how professionals build digital environments.
Robots Learn to Generalize
Physical Intelligence pushed the thesis toward embodiment3. Its models control robot bodies, a step past generating images of worlds those bodies inhabit only on screen. Co-founder Lachy Grooms came from Stripe. TechCrunch has called the result Silicon Valley's buzziest robot-brain startup. The company shipped π0.5 in April 2025, described as achieving open-world generalization. Its successor, π0.7, followed exactly one year later. The company says that model shows a genuine step-change, rather than incremental improvement. Reports in March 2026 placed the company back in talks for another billion-dollar round.
Figure AI pursued the humanoid-hardware side of the same bet4. It closed a Series C exceeding $1 billion at a $39 billion post-money valuation in September 2025. A world model trained on video teaches an AI what physical interaction looks like. Putting that understanding in a humanoid body tests it. Gravity, friction, and small failures do the testing.
The Incumbents Build Their Own Worlds
Google DeepMind and Nvidia proved the thesis reached past startups56. DeepMind announced Genie 3 on Aug. 5, 2025, as a limited research preview. It generates interactive environments from text prompts at 24 frames per second and 720p. Visual memory runs up to one minute, and interaction holds for a few minutes. The model handles water, lighting, and animal behavior directly.
Jack Parker-Holder and Shlomi Fruchter wrote the announcement5. "Given a text prompt, Genie 3 can generate dynamic worlds that you can navigate in real time," they wrote. A second line marks the advance: "Genie 3 is our first world model to allow interaction in real-time, while also improving consistency and realism compared to Genie 2". They built the system to train embodied agents inside generated worlds, which treats the environment itself as training infrastructure.
Nvidia answered with Cosmos6. The platform combines world foundation models, tokenizers, and data pipelines. Robotics, self-driving cars, and industrial simulation are the targets. Cosmos 3 uses a Mixture-of-Transformers architecture that separates reasoning from generation. Nvidia's ecosystem page lists Skild AI among its physical-AI partners. Nvidia intends Cosmos as infrastructure other world-model startups build on, rather than a rival to them.
What the Departures Imply
Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le launched Discovery Loop on Aug. 5, 20267. Their names sit on MapReduce, Bigtable, Spanner, and TensorFlow, which is to say on a large share of the infrastructure this industry runs on. Discovery Loop is a Delaware public benefit corporation. It exists to automate the propose, run, evaluate loop of research. Machine learning research comes first, then science generally. Radical Ventures and Khosla Ventures co-led the seed. Google itself is a founding investor, supplying compute for at least the first year.
Their exit from an employer with vast compute and data carries a signal. Researchers with maximal internal resources judged an independent venture worth the risk. That pattern recurs across nearly every company named here.
Rodney Brooks reads the robotics leg of this wager differently. The founder of iRobot and Rethink Robotics, now cofounder and chief technology officer of Robust.AI, posted his case on his own site on Sept. 26, 2025, aimed at the training method behind the humanoid bets Figure and Physical Intelligence are chasing8. "In this post I explain why today's humanoid robots will not learn how to be dexterous despite the hundreds of millions, or perhaps many billions of dollars, being donated by VCs and major tech companies to pay for their training," Brooks wrote. Brooks spent decades building robots himself, at iRobot and then Rethink, and he aims his skepticism at the method: the premise that video of human hands alone teaches a machine touch.
Read together, these bets describe a field hedging against the possibility that transformer scaling alone stops producing proportional gains. World models offer a different scaling axis5. It is measured in physical coherence and accuracy about real environments, rather than token-prediction loss. Capital is chasing that axis, from a startup founded by a Turing Award winner to a chipmaker's own product roadmap16. The next paradigm shift arrives from researchers betting years of runway on physical understanding, well before the market prices the wager.
By the numbers
- $1.03 billion: LeCun's raise for AMI Labs by March 2026, at a $3.5 billion pre-money valuation1.
- World Labs reached a $1 billion valuation in February 2026, up from $230 million at its 2024 launch2.
- $200 million: Autodesk's contribution alone to that February 2026 World Labs round2.
- $39 billion: Figure AI's post-money valuation on a Series C exceeding $1 billion, closed September 20254.
- 24 frames per second at 720p: Genie 3's output, with visual memory up to one minute5.
- Four: Google researchers who left together on Aug. 5, 2026 to found Discovery Loop7.
- Nov. 20, 2025: the day LeCun left Meta, roughly three months before AMI Labs closed its round1.
What to watch
LeBrun's own prediction supplies the first test1. Count how many companies call themselves world-model builders by spring, then count how many of them ship a product that works. AMI Labs promised to open-source a substantial amount of code, a pledge worth checking against what actually appears in public repositories. Watch Discovery Loop's valuation talks too, since a jump from roughly $10 billion to a reported $50 billion inside a few weeks says more about the market right now than about any work the company has done7.
Sources
- Anna Heim, "Yann LeCun's AMI Labs raises $1.03B to build world models," TechCrunch, March 9, 2026, https://techcrunch.com/2026/03/09/yann-lecuns-ami-labs-raises-1-03-billion-to-build-world-models/
- Rebecca Bellan and Marina Temkin, "World Labs Coverage," TechCrunch, various dates through Feb. 18, 2026, https://techcrunch.com/tag/world-labs/
- Connie Loizos, "Physical Intelligence Coverage," TechCrunch, various dates through April 16, 2026, https://techcrunch.com/tag/physical-intelligence/
- Figure, "Figure AI News," Figure, Sept. 16, 2025, https://www.figure.ai/news
- Jack Parker-Holder and Shlomi Fruchter, "Genie 3: A New Frontier for World Models," Google DeepMind, Aug. 5, 2025, https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/
- Nvidia, "NVIDIA Cosmos," Nvidia, accessed Sept. 3, 2026, https://www.nvidia.com/en-us/ai/cosmos/
- TechCrunch Staff, "Jeff Dean and other top AI researchers are leaving Google to launch their own startup," TechCrunch, Aug. 5, 2026, https://techcrunch.com/2026/08/05/jeff-dean-and-other-top-ai-researchers-are-leaving-google-to-launch-their-own-startup/
- Rodney Brooks, "Why Today's Humanoids Won't Learn Dexterity," rodneybrooks.com, Sept. 26, 2025, https://rodneybrooks.com/why-todays-humanoids-wont-learn-dexterity/