DeepSeek AI app logo photo illustration, of the Chinese artificial intelligence company that develops open-source large language models LLM in a fraction of the cost of the US based competition. Close up of DeepSeek application thumbnail logo displayed on a smartphone screen next to other AI providers like ChatGPT of OpenAI and Copilot of Microsoft. Amsterdam, the Netherlands on January 28, 2025 (Photo by Nicolas Economou/NurPhoto via Getty Images)NurPhoto via Getty ImagesIn the first half of 2026, venture investors committed more than $3 billion to startups building world models, AI systems that learn the physics of an environment instead of the statistics of text. Yann LeCun walked out of Meta and raised a $1.03 billion seed at a $3.5 billion valuation, the largest seed round in European history. The wager underneath every check is that the next foundation model will simulate reality rather than describe it.As frontier language models commoditize and inference margins compress, investors are repricing where defensibility sits in the AI stack. World Labs, founded by Stanford’s Fei-Fei Li, raised $1 billion in February 2026 at a $5.4 billion post-money valuation, lifting its total to $1.23 billion. Decart closed $300 million in May at a $4 billion valuation. Odyssey, the self-driving spinout, raised a $310 million Series B in June at $1.45 billion. General Intuition, eight months after a $133.7 million seed, is in talks to raise roughly $300 million at just over $2 billion. Vinod Khosla, who led that seed, has said he expects multiple hundred-billion-dollar companies to emerge from the category. The pattern repeats; each round prices a thesis that physical AI, not another chatbot, is the next layer worth owning.The open racesThe field has not agreed on what a world model even is. Fei-Fei Li forced the question in June with a functional taxonomy that splits these systems into renderers, simulators, and planners. A renderer outputs pixels for human eyes and carries no understanding of three-dimensional structure, which is where she places most current demos, including Google DeepMind's Genie 3. A simulator outputs state that holds up under geometry and Newton's laws. A planner outputs actions. Most products marketing themselves as world models today are renderers.Three races run underneath that distinction. The first is long-horizon consistency. Genie 3 generates navigable worlds at 24 frames per second but stays coherent for only a few minutes, with memory of changes lasting about a minute. The second is representation. World Labs anchors worlds in Gaussian splats and physics engines, while AMI builds on Joint Embedding Predictive Architecture, learning abstract representations of reality in compressed latent space rather than predicting frames. The third is the data moat. OpenAI reportedly offered $500 million for Medal's gameplay archive before General Intuition spun out of it, because action-labeled first-person video teaches the mapping from decision to consequence that spectator footage cannot. Whoever controls that kind of interactive data controls a moat that compute alone cannot buy.MORE FOR YOUThe lineagesThe founders cluster into recognizable camps. The self-driving diaspora produced Odyssey, whose CEO Oliver Cameron ran product at GM's Cruise and whose CTO Jeff Hawke engineered at Wayve, the UK lab that itself raised $1.2 billion in February. The academic-lab camp produced World Labs, out of Stanford and Google, and AMI, out of Meta's former chief AI scientist. The OpenAI orbit shows up as capital, with co-founder Andrej Karpathy angel-investing in Decart alongside former Disney chief Michael Eisner. Decart's own founders came out of Israeli military intelligence Unit 8200, and General Intuition grew from a consumer gaming platform.One name appears across nearly every cap table. Nvidia has committed over $40 billion in AI equity in 2026, frequently taking a stake in exchange for long-term GPU commitments, and it backs World Labs, AMI, Decart, and Odyssey. Amazon is countering with its Trainium silicon, signing Odyssey and Decart as customers and tying both to AWS.The market map of world models: ranked by capital raised. The feature column marks where each sits on the renderer-to-simulator spectrum.Josipa Majic PredinSkepticism is warranted. AMI CEO Alexandre LeBrun predicted that within six months every company would relabel itself a world model to raise money, and valuations are running well ahead of shipped product and disclosed revenue. The honest benchmark is sim-to-real correlation, whether a policy trained inside a generated world performs in the physical one. For founders, the category will sort along Li’s line. Renderers compete on visual polish that commoditizes quickly, while simulators that hold their geometry become the training substrate for robotics and agents. For investors, the test through 2026 is simple; watch which of these worlds survive an agent, not a viewer, walking through them.