More than 600 billion yuan has been poured into the vague technology: capital is footing the bill for the definition of "World Model"
666 billion yuan poured into the world model sector amid conceptual generalization, while relevant technologies are far from meeting expectations.
In the first seven months of 2026, China's primary market invested 666 billion yuan in the "world model" concept. Among 64 "world model" enterprises, 58 have obtained financing, with a financing success rate close to 90%: Eximverse raised 35 billion yuan in just three months, with its valuation soaring to 3 billion US dollars; Shengshu Technology has spread news of a Hong Kong IPO; Runway's annual recurring revenue has exceeded 200 million US dollars, with a valuation of 5.3 billion US dollars. These figures have heated up the "world model" concept, but Li Feifei, co-founder and CEO of World Labs, poured cold water on the hype with one sentence: The world model is "one of the most important and most abused terms" in the AI field.
There is a huge definition gap between the total financing of over 600 billion yuan and the technical essence covered by the broad concept of "world model".
01
The Same Basket Holds Multiple Types of Businesses
Wang Zhongyuan, Dean of the Institute of Intelligent Innovation, divided products named "world models" into four categories at the 2026 Zhiyuan Conference: language-centered, pixel-centered, 3D structure-centered, and visual representation-centered. Li Feifei's classification is more concise: the one that outputs pixels is a renderer, the one that outputs environmental states is a simulator, and the one that outputs action instructions is a planner (sometimes also called a "decision-maker"). The two classification methods point to the same fact: Marble from World Labs, GWM Worlds 2 from Runway, GigaWorld from Eximverse, and Guanglun Intelligence are not doing the same thing at all, but are defined by the same concept: the world model.
The commercial essence of Marble is a content tool sold to game development, film and television special effects and VR teams: the standard version is $20 per month, and the flagship version is $95 per month, including 75 generation credits and commercial usage rights. Marble delivers roamable 3D scenes and optional collision meshes, with evaluation dimensions of creator control and export interoperability, rather than closed-loop policy success rate. Its core advantages are high 3D consistency and fast generation speed, at the cost of no physical simulation.
Runway has taken a different path: in September 2026, Runway released the second-generation world model GWM Worlds 2, focusing on real-time interactive world simulation, clearly repositioning itself from a "video supplier" to a "simulation base for physical AI". It does not sell video clips, but the underlying simulation layer of the clips — a base that can learn physical laws from videos and support training for robots and autonomous driving. With annual recurring revenue of 200 million US dollars and a price-to-sales ratio of about 26.5 times, this is a valuation narrative for a SaaS company transitioning to an infrastructure company, which is not the same type of business as selling generation quotas by subscription times like Marble.
The domestic market also shows drastic differentiation. Eximverse's GigaWorld-1 scored over 60 points in the WorldArena evaluation, surpassing Google, NVIDIA and Alibaba. At the same time, the company also launched the Maker H01 robot body, with the narrative logic of embedding the world model into the "brain" of embodied intelligence. Shengshu Technology raised 26 billion yuan in two months, with a post-investment valuation of over 120 billion yuan. Its path is to move closer to the "world model" from video generation, run commercialization first, and then upgrade its technical concept; Zhixiang Future raised over 21 billion yuan in three months, also relying on the prerequisite experience accumulated from video generation, first supporting itself with scenarios that are willing to pay for "generation results" such as film, television and marketing, instead of tackling hard bones like physical simulation that take three to five years to verify right from the start. There is also Guanglun Intelligence, which does not directly develop world models, but provides simulation data and evaluation infrastructure for world model teams. All of the world's top three world model teams cooperate with it, and more than 80% of the simulation assets of the world's mainstream embodied intelligence teams come from it. This company became the first unicorn in the global embodied data field in March 2026. In essence, it sells "shovels" rather than gold mining stories.
02
Rendering Accuracy Is Rising While Physical Understanding Scores Remain Stagnant
The heat on the financing side and the coldness on the technology side form an embarrassing contrast.
The S-Lab team of Nanyang Technological University released the Apple-π benchmark in July 2026, which is the first video model evaluation tool with physical laws as the core. It does not test whether the picture looks good, but whether the model can perceive physical quantities, express governing laws, and deduce subsequent motions that conform to the laws. 11 representative models participated in the evaluation, and the highest score was only 0.473. Physical motion reasoning was clearly pointed out as the main bottleneck. A paper presented at WACV 2026 is more straightforward: a number of current models such as OpenAI's Sora, Runway, Pika Labs' Pika and Google Research's Lumiere have "extremely limited" physical understanding, and this limitation has nothing to do with visual realism: The more realistic the picture is, the easier it is to mistakenly think that the model "understands", but looking similar and truly understanding are two completely different things.
More specific problems appear in action-conditioned world models. The WorldOracle benchmark jointly proposed by Princeton University, the University of Texas at Austin, NVIDIA and Zhiyuan uses real robot operation records to compare the predictions of the model, and found a pattern that is almost black humor: the movement path of the robotic arm is predicted quite accurately, but for whether the object is grasped and where it goes after being grasped, the model often "fills in the blanks successfully", which is quite in line with the "Ah Q spirit". ByteDance's IRASim scored 78.8 points for the robotic arm and only 29.1 points for operating objects; NVIDIA's Cosmos 3 scored 89.4 points vs 45.4 points. After the prediction time is extended, the score of another NVIDIA model DreamDojo drops from 50.56 to 27.08, and that of Cosmos 3 drops from 53.99 to 26.60, almost halved. There is a whole technical gap covering physical simulation, long-term stability and real machine verification between a model that can make the robot "look like it has completed the task" in the video and a model that can make the real robotic arm not grasp empty and not knock over cups on a cluttered desktop.
The actual measurement feedback of Marble also confirms the same thing: the light and shadow are coherent when moving over a short distance, and the indoor scene looks good, but distortion occurs when moving over a large range, and the performance is "quite negative" in outdoor scenes. A developer said bluntly on Hacker News: "As a 'world' model, it doesn't seem to quite understand what physical space is." This is not the shortcoming of Marble alone. Most of the products currently classified under the name of "world model" are making rapid progress in "generating things that look reasonable", while making slow progress in "predicting what is physically truly correct". And the latter is exactly the full value of simulators and planners (decision-makers).
03
Capital Places Bets At The Pace Of The Industry Track While Technology Climbs At The Pace Of Disciplines
The signal from the policy side is clear, but the wording is intriguing: the 15th Five-Year Plan lists "world model" as a cutting-edge technology to be tackled, and the State Administration for Market Regulation is stepping up the formulation of national standards, explicitly requiring to "quickly gather industry consensus". At present, there is no such consensus yet. In Wang Zhongyuan's words: "Everyone has not reached a consensus on what the definition of the world model is." However, in a field where even the boundaries and definitions have not been delineated, 666 billion yuan and 58 financed companies have poured in first, and the screening cost has been transferred to downstream parties — car manufacturers, robot companies and factories. Under the "world model" label, they have to distinguish by themselves whether the product in front of them is an easy-to-use rendering tool, an inaccurate simulation environment, or a decision module that can really command the robotic arm to complete actions.
Runway's annual recurring revenue is real, Eximverse's evaluation score is real, and the 0.473 score of Apple-π is also real. The problem is not that someone is "glorifying" themselves, but that after different types of products are put into the same valuation coordinate system, a misplaced time window appears between the exit rhythm of capital and the maturity rhythm of technology. Renderers can make a beautiful annual revenue curve within two years relying on the subscription system. The value of simulators takes three to five years of real machine data accumulation to be verified, and the commercialization of decision-makers will definitely come later. When the three share the label of "world model" and the same narrative of "next-generation general artificial intelligence infrastructure", the category that first runs through revenue will first define the market's price expectation for the concept of "world model", while the category that is really working on physical understanding will have to prove that it deserves the high valuation in a raised valuation framework.
The 666 billion yuan figure did not buy a clearly defined technology, but the right to define the concept that is being contested. The contest for the right of definition is never just an academic debate — Who is qualified to say what counts as a "world model" is qualified to determine the valuation anchor point for the next round of financing. As for when the physical understanding capability will catch up with the rendering accuracy, the 0.473 score of Apple-π will not automatically become 0.8 just because there is an extra zero after the financing amount. This is an engineering and scientific problem, and engineering and science do not participate in roadshows. This misalignment will not disappear automatically just because the next round of financing arrives.
This article is from the WeChat official account "Caineng Circle", authorized to be released by 36Kr.