Core argument: The world-model field has four kinds of builders, each with a different source of advantage: the giants hedging with homegrown models (Google DeepMind, NVIDIA), star founders raising historic rounds on the strength of their names (World Labs, AMI Labs), specialists compounding proprietary data (Decart, General Intuition, Odyssey, Emulate), and a China track where the same technology sits closest to paying customers through driving (Huawei, NIO) and open-source strategy (Kunlun). The pattern that cuts across all four: the models themselves reproduce fast — open weights are everywhere — so the durable bets are on data and compute. Watch who owns those.
This is the fourth chapter of an ongoing series on world models. So far: why LLMs hallucinate, what separates a world model from an LLM, and how the four technical routes actually work. This chapter maps who is building them. Figures below were checked in early October 2026; this is the fastest-moving corner of AI, so treat amounts as snapshots.
1. The short answer
Ask who is building world models and the honest answer is: nearly everyone serious about AI. But the field organizes into four groups, and the differences between them are more revealing than the list.
The giants — Google DeepMind and NVIDIA above all — build their own models and control the compute they run on. The star founders — Fei-Fei Li and Yann LeCun — raised historic sums on personal reputations and specific theses. The specialists — Decart, General Intuition, Odyssey, Emulate — are smaller, faster, and built on proprietary data no one else has. And China runs a parallel track where world models are closest to actual revenue, through assisted driving at Huawei and NIO, with Kunlun pushing an open-source line for embodied AI.
Three of this year’s numbers frame the whole map: an $8.2 billion acquisition, a $1.03 billion seed round, and a startup whose valuation went from $2.3 billion to $6 billion in two months.
2. The gold rush, with a twist
Gold rushes have a familiar cast. There are the miners, who chase the gold directly. There are the shovel sellers, who profit from everyone else’s chase. And there are the land claimants, who position themselves early on ground that will matter later.
Map the field onto that cast and it almost works. The model labs are the miners, digging for capable systems. The compute sellers — NVIDIA first among them — sell the shovels. And the data owners are the land claimants: whoever holds vast gameplay footage, robot logs, or driving miles holds ground the models must eventually cross.
The analogy breaks in an instructive way. In a real gold rush, the three roles stay separate — miners mine, sellers sell. Here, every giant plays all three at once. Google DeepMind mines with Genie 3 while Google sells the compute it trains on and sits on YouTube-scale data. NVIDIA sells shovels to everyone while mining with its own open models. Amazon, meanwhile, appears as an investor in one startup, an infrastructure partner to another, and a rumored buyer of a third. When the roles collapse into each other like this, the frame to use is not a gold rush. It is infrastructure — and infrastructure outlives any single strike.
3. The four groups, in detail
The giants: hedge, then standardize
Google DeepMind runs the most visible world-model program: Genie 3, previewed in August 2025 and since opened to the public through Project Genie, generates interactive 3D environments in real time. The bet is breadth — a general-purpose world model that serves research, gaming, and agent training at once.
NVIDIA’s bet is different and quieter: own the standard. At GTC Taipei in May 2026 it released Cosmos 3, a world foundation model for physical AI that the company calls the first fully open “omnimodel” — one system handling reasoning, simulation, and robot action generation. That followed an October 2025 release of a large open physical-AI dataset, including over 1,700 hours of multimodal driving sensor data. The playbook is familiar from CUDA: open the models, seed the ecosystem, and sell the compute that everything eventually runs on.
The star founders: thesis-driven billions
Fei-Fei Li’s World Labs spent 2025 shipping Marble, a commercial model that turns text or images into persistent, explorable 3D worlds, and September 2026 shipping Atlas, an “omni” model that folds video generation, 3D reconstruction, and camera-controlled simulation into one system. Then came the field’s defining deal: on September 28, AMD agreed to acquire World Labs for roughly $8.2 billion in stock, with Li joining AMD as chief scientist. The deal was announced in late September and remains subject to regulatory approval.
Yann LeCun, the Turing Award winner who spent years arguing that autoregressive language models are the wrong road, put his thesis to the market in March 2026: AMI Labs raised a $1.03 billion seed round — reported as the largest in European history — to build world models explicitly against the LLM-first grain.
The bet in both cases is credibility. These founders raised on the strength of decades of work, and their theses, not their products, are what the money priced.
The specialists: compound the data
The most instructive story belongs to General Intuition. It trains agents on millions of hours of gameplay footage — data no competitor can legally scrape — and its valuation ran from a $2.3 billion Series A in June 2026 to a $6 billion round two months later, as it pivoted toward robotics. Data, priced in real time.
Decart, the Israeli real-time video model company, raised $300 million at a $4 billion valuation in May 2026, struck an infrastructure deal with AWS in August, and by late August was reportedly exploring a sale in the $6-7 billion range with Amazon named among potential buyers. Odyssey raised $310 million at a $1.45 billion valuation in June for Hollywood-grade world simulation. And Emulate — founded by the three researchers behind Genie 3 — has reportedly been in talks for a seed round near $700 million at a multi-billion valuation while remaining pre-product.
The specialists’ bet is speed plus data moats: footage, gameplay, and rendering pipelines that incumbents cannot easily replicate.
China: closest to the paying customer
The Chinese track looks different because it monetizes through driving today. Huawei’s ADS 5, announced in April 2026, runs on an updated world-model architecture whose cloud training uses multi-agent adversarial scenarios — and the company reported over 10 billion kilometers of cumulative assisted-driving mileage by April. NIO has pushed its own world model to more than 460,000 cars, pairing it with closed-loop reinforcement learning. And Kunlun has open-sourced its Matrix-Game line — among the first large open interactive world models — competing for the same developer mindshare NVIDIA courts.
The bet here is distribution: world models that ship in millions of cars collect more real-world driving data than any simulator can fake, and the open-source line builds the ecosystem the way Android did.
4. My take
Three readings, in descending order of confidence.
First: judge these companies by their data and compute position, not their demos. The models reproduce quickly — within months of any release, open alternatives close most of the gap. What does not reproduce is General Intuition’s gameplay archive, NIO’s driving miles, or NVIDIA’s installed base. Every premium valuation on this map traces to one of those two moats, and every company without either is priced on team alone.
Second, medium confidence: the consolidation has started and the buyer list is the tell. AMD bought World Labs. Amazon invested in Odyssey, partnered with Decart, and is rumored to be circling it. Chipmakers and clouds buying modeling teams is the pattern I expect to define 2027, because the physical-AI story — robots, autonomous vehicles — requires simulation environments, and the fastest way to own those is to buy the people who build them.
Third, also medium: mind the gap between valuations and revenue. Most of the companies on this map have little revenue today; the clearest exception is driving, where Huawei and NIO bill through cars on the road. Corrections in this category are likely — the two-month doubling of General Intuition’s valuation is not a pace that survives contact with audited financials. But the direction of the field is not in question, and a correction would not change who owns the data.
5. Questions people actually ask
Why would a chipmaker buy a world-model company? Because physical AI needs simulation. Training robots and autonomous vehicles at scale requires generated environments, and the company selling the compute wants to own the place where that compute gets spent. AMD’s stated rationale for World Labs was exactly this.
What is China’s approach in one line? Monetize through driving, where world models already ship in millions of cars, and give away open-source models to win the developer ecosystem — two different strategies run by two different kinds of companies.
Which of these should I actually watch? The ones compounding data, and the deal flow around them. General Intuition’s two-month valuation jump and Decart’s reported sale exploration are the signals to watch: they tell you where the next acquisitions and the next corrections will land.
Sources
- TechCrunch, “AMD will acquire Fei-Fei Li’s World Labs for $8.2 billion”, September 28, 2026; see also Fei-Fei Li’s own account of the deal.
- TechCrunch, “Yann LeCun’s AMI Labs raises $1.03 billion to build world models”, March 9, 2026.
- TechCrunch, “General Intuition’s $2.3B bet that video games can train AI agents for the real world”, June 25, 2026.
- NVIDIA Newsroom, “NVIDIA launches Cosmos 3, the open frontier foundation model for physical AI”, May 31, 2026.
- Google DeepMind, “Genie 3: A new frontier for world models”, August 2025.
- World Labs, “Atlas: A World Model for Spatial Intelligence”, September 2026.
- Decart funding and AWS partnership: Columbus Jewish News (May 19, 2026) and 36Kr English (August 28, 2026) — outlet archives, article paths not retrievable at citation time.
- Odyssey Series B: The AI Insider (June 18, 2026) — outlet archive, article path not retrievable at citation time.
- General Intuition’s August 2026 round at a $6 billion valuation: InvestGame (June 29, 2026) and The AI Insider (August 2026) — outlet archives, article paths not retrievable at citation time.
- Huawei ADS 5 (April 2026), NIO world model (January 2026), and Kunlun Matrix-Game (2025-2026): reported by Tencent News, Sina, and Securities Times (Chinese outlets) — outlet archives, article paths not retrievable at citation time.
Next in this series: artificial general intelligence — the term everyone uses and few define. What it actually means, why the abbreviation itself is a trap, and where the serious definitions disagree.