World Models: Why AI's Biggest Names Bet Billions in 2026
Lin Junyang, Fei-Fei Li, and Yann LeCun are raising billions for world models in 2026. What a world model is, who is funding it, and why the money is moving.
By Capital & Compute
The clearest signal of where AI capital is moving in 2026 is not another chatbot. It is the resume of the people raising money to leave chatbots behind. In June 2026, Lin Junyang, the engineer who built Alibaba’s Qwen into the world’s most-downloaded open-weights model family, closed a first round for a new lab at a valuation of about $2 billion, with no product shipped and a stated focus on world models and embodied intelligence. He is the fourth marquee name in five months to raise at a multibillion-dollar valuation for the same idea. The pattern is the story: the smartest money in AI is betting that the next paradigm is a model of the physical world, not a better predictor of the next word.
What is a world model, and how is it different from an LLM?
A world model is an AI system that predicts the next state of an environment rather than the next token of text. A large language model learns the statistics of language and answers in words. A world model learns the dynamics of a place: how objects move, what a camera would see after a step forward, what happens when a hand pushes a cup. It is trained on video, 3D scenes, sensor streams, and recorded actions instead of, or alongside, text.
The cleanest demonstration is Google DeepMind’s Genie 3, described by DeepMind in 2026 as a general-purpose world model that turns a text prompt into an interactive environment a user can navigate in real time at 720p and 24 frames per second, holding the scene consistent for a few minutes. DeepMind shipped it to AI Ultra subscribers in the United States as Project Genie on January 29, 2026. The difference from a video generator is the key point: the world responds to your input as you move, rather than playing back a fixed clip.
The contrarian case for why this matters comes from Yann LeCun, the Turing Award winner who spent more than a decade running Meta’s FAIR lab. His thesis, which he has argued since 2022, is that language models will never reach general intelligence because predicting text is not the same as understanding physical reality. His preferred architecture, the Joint Embedding Predictive Architecture (JEPA), learns abstract representations of how the world behaves instead of generating the world token by token. Whether or not he is right, investors are now pricing his conviction in the billions.
A world model predicts the next state of the world. A language model predicts the next word. The bet is that the first is where intelligence goes next.
The 2026 world-model funding wave, by the numbers
Four rounds in five months tell the story, and the founders are among the most credentialed in the field.
| Item | Value |
|---|---|
| Runway | $5.3B |
| World Labs · Fei-Fei Li | $5B |
| AMI Labs · Yann LeCun | $4.5B |
| Lin Junyang's lab | $2B |
Yann LeCun’s lab, Advanced Machine Intelligence (AMI), raised what TechCrunch reported in March 2026 as a $1.03 billion seed round at a $3.5 billion pre-money valuation, the largest seed round on record in Europe. The backer list reads like a who’s who of people betting against the LLM monoculture: Bezos Expeditions, NVIDIA, Temasek, Samsung, Toyota Ventures, and individuals including Jeff Bezos, Mark Cuban, and Eric Schmidt. The lab had no commercial product at the time of the raise.
Fei-Fei Li, the Stanford researcher often called the godmother of modern computer vision, raised $1 billion for her startup World Labs in February 2026, as reported by Bloomberg and TechCrunch, with Autodesk leading at $200 million alongside Andreessen Horowitz, NVIDIA, and AMD. Unlike the other three, World Labs has shipped: its product Marble generates persistent, high-fidelity 3D environments from an image, video, or text prompt. Reporting in 2026 put the company in talks at a valuation near $5 billion, roughly five times its 2024 mark.
| Item | Prior round | 2026 |
|---|---|---|
| World Labs | $1B | $5B |
| Runway | $3.3B | $5.3B |
Runway, the video-generation company, raised $315 million at a $5.3 billion valuation and told AI Business it would use the money to pretrain its next generation of world models, a pivot from generating clips to modeling environments. That is the same migration in miniature: a company that sells text-to-video output deciding the durable asset is the model of the world underneath.
Lin Junyang’s lab is the newest and the least public. Lin resigned from Alibaba on March 3, 2026, the day after the Qwen team shipped its Qwen3.5 small models, as reported by TechCrunch and TechNode. A Caixin profile noted he was 32 and Alibaba’s youngest P10, the company’s senior leadership tier. By June, according to reporting from The Information summarized by China outlet 36Kr’s Intelligence Emergence, Tencent had made a strategic investment in his new lab at a post-money valuation around 13.5 billion yuan (roughly $2 billion), with HongShan (formerly Sequoia China) and Gaorong Ventures reported in talks. Chinese corporate records cited in the earlier 36Kr report show Lin registered or took control of three Shanghai entities between May and June 2026, including Yuyong (Shanghai) Technology and Shanghai Bulage Technology, the legal scaffolding of a lab being built in public view.
None of these figures sit in isolation. Global startup funding hit a record $297 billion in the first quarter of 2026, according to TechCrunch, with AI absorbing the majority of it. World models are the part of that flood with the most pedigree per dollar and the least revenue per dollar.
Feb 2026
World Labs raises $1B
Fei-Fei Li's startup, with Autodesk leading. Later reported in talks near $5 billion.
Mar 2026
AMI Labs raises $1.03B
Yann LeCun's lab, at a $3.5 billion pre-money valuation: Europe's largest seed round.
Jun 2026
Lin Junyang's lab, about $2B
Tencent invests; the former Qwen lead enters the race with no product yet.
Why investors write billion-dollar checks before there is a product
Three of the four labs above raised at multibillion-dollar valuations with nothing to sell. That is not irrational, and understanding why explains the whole wave.
The first reason is the bet on a paradigm shift. LLM capabilities are converging, and the price of using them is falling toward commodity levels, a trend visible in the tight $10 to $20 band of AI coding plan pricing where a dozen competent models cluster within a few dollars of each other. When a capability commoditizes, the returns move to whatever comes next. World models are the most credible candidate for “next,” so the capital chasing outsized returns flows to the labs furthest ahead on it, regardless of revenue.
The second reason is that founder pedigree is the only available signal this early. There is no revenue to underwrite and no benchmark that settles the question, so investors price the people: the architect of the most influential open model family, the researcher who built the field of large-scale computer vision, the scientist whose architecture the field is named after. In a paradigm bet, the team is the moat, because the work that matters has not been done yet.
The third reason is strategic, and it is loudest in China. Tencent backing a former Alibaba star is a move in a talent war, not just a financial position. Owning a stake in the lab of the person who built a rival’s flagship model is cheap insurance against being left behind, which is why a strategic investor will pay a price a pure financial investor might not.
The compute bill is the real story
For a site that tracks the economics of AI infrastructure, the interesting number is not the valuation. It is the cost of the thing being built.
World models are far more expensive to train than chatbots because their training data is heavier per unit of learning. A text token is a few bytes. A second of video is millions, and a world model has to learn the physics implied across frames, not just the next symbol. The scale shows in NVIDIA’s own platform: Cosmos 3, the open world foundation model NVIDIA launched on June 1, 2026 for physical AI, was trained on 20 trillion tokens of multimodal data including nearly a billion images and 400 million real and synthetic videos. Curating, storing, and training on that corpus is a different order of capital expenditure than fine-tuning a language model on web text.
Inference is heavier too. A chatbot answers once and stops. A world model like Genie 3 renders a navigable environment continuously, at 24 frames per second, reacting to input the whole time. The token economics that already make agentic AI expensive, worked through in our breakdown of what Claude Code costs per task, get worse when the output is a live world instead of a block of text. The unit cost of a useful world-model session is an open question, and it is the question that decides whether any of these valuations make sense.
That cost is also why NVIDIA sits at the center of the wave as both supplier and investor. It backed AMI Labs and World Labs, ships Cosmos as an open model, and convened a Cosmos Coalition of world-model builders that includes Runway. When the picks-and-shovels vendor is also funding the prospectors and giving away a reference model, it is hedging that the gold rush happens on its hardware regardless of which lab wins.
The open-versus-closed question Lin carries with him
Lin Junyang’s raise carries a subplot the others do not. His reputation was built on open weights: he turned Qwen from an internal project into the most widely adopted open model family in the world, the same lineage whose coding plans sit in the China tier of our pricing comparison and whose open releases pressure the closed frontier on cost. Whether his new lab keeps that open posture or closes up like a typical frontier startup is one of the more consequential unknowns in the space.
It matters because the open-versus-closed split has been the main force holding down AI prices, a dynamic that runs through the entire 2026 AI coding agent landscape. NVIDIA chose openness with Cosmos. World Labs and AMI Labs are building commercial products. If world models follow LLMs, an open-weights contender will eventually drag the price of the capability down, and the lab most likely to produce one is run by the person who already did it once. A closed Lin lab would be a meaningful signal that the economics of world models do not support giving the model away.
Bottom line
World models are the biggest pre-revenue bet in AI in 2026, and the bet is being placed by the most credentialed people in the field: Lin Junyang at about $2 billion, Yann LeCun at $3.5 billion pre-money, Fei-Fei Li near $5 billion, Runway at $5.3 billion. The thesis is coherent: language is commoditizing, physical and spatial intelligence is the next frontier, and the team is the only thing worth pricing before the work is done. The risk is equally clear and rarely stated next to the valuations. World models cost far more to train and run than chatbots, none of these labs has shown the capability pays for that compute, and the open-versus-closed question that governs pricing is unresolved. Watch two numbers as this plays out: the cost per useful world-model session, and whether an open-weights contender appears. Those, not the funding headlines, will decide whether 2026’s biggest bet pays off. Every figure here is sourced and dated, in line with our editorial standards.
Sources
- TechCrunch (2026). Yann LeCun’s AMI Labs raises $1.03 billion to build world models. March 9, 2026. techcrunch.com
- Bloomberg (2026). AI Pioneer Fei-Fei Li’s Startup World Labs Raises $1 Billion. February 18, 2026 (paywalled). bloomberg.com
- TechCrunch (2026). World Labs lands $200M from Autodesk to bring world models into 3D workflows. February 18, 2026. techcrunch.com
- AI Business (2026). AI Startup Runway Raises $315M, Pivots to World Models. aibusiness.com
- TechCrunch (2026). Alibaba’s Qwen tech lead steps down after major AI push. March 3, 2026. techcrunch.com
- TechNode (2026). Qwen technical lead Lin Junyang leaves Alibaba, sources say. March 4, 2026. technode.com
- Caixin Global (2026). In Profile: The Alibaba Engineer Behind Qwen Who Walked Away. March 10, 2026. caixinglobal.com
- 36Kr / Intelligence Emergence (2026). Report: Lin Junyang Secures Tencent Investment with First-round Valuation of 13.5 Billion Yuan (summarizing reporting by The Information). eu.36kr.com
- 36Kr / Intelligence Emergence (2026). Lin Junyang Launches Business: New Company Valued at Around $2 Billion. eu.36kr.com
- Google DeepMind (2026). Genie 3: A new frontier for world models. deepmind.google
- Google (2026). Project Genie: AI world model now available for Ultra users in the U.S. January 29, 2026. blog.google
- NVIDIA (2026). NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI. June 1, 2026. nvidianews.nvidia.com
- TechCrunch (2026). Startup funding shatters all records in Q1. April 1, 2026. techcrunch.com