Alibaba backs 'world models'

Alibaba just led a 2 billion yuan funding round, about $290 million, into Beijing startup ShengShu, and the money is aimed at a “general world model” instead of another text chatbot. CNBC reported the round on April 10, 2026, with TAL Education and Baidu Ventures joining Alibaba Cloud. (cnbc.com) A chatbot predicts the next word in a sentence. A world model tries to predict what happens next in a room, on a street, or around a robot arm after something moves, falls, opens, or blocks the light. (deepmind.google) That difference is why video matters here. ShengShu already makes the Vidu video generator, and Alibaba is betting that systems trained on motion, timing, and spatial change can become useful for robotics and autonomous machines. (cnbc.com) Alibaba has been moving in this direction for months. On February 24, 2026, Alibaba Cloud said its DAMO Academy released RynnBrain, an open-source embodied foundation model built for environmental cognition, spatiotemporal understanding, and task planning in robots. (alibabacloud.com) RynnBrain’s own project page describes the model as grounded in physical reality, with tools for navigation, planning, reasoning, and object localization. That is a very different product from a customer-service bot that only reads and writes text. (github.com) Alibaba is not chasing this alone. Google DeepMind said in August 2025 that Genie 3 could generate interactive environments at 24 frames per second and keep them consistent for a few minutes, which shows how fast big labs are pushing simulated worlds as training grounds for agents. (deepmind.google) Nvidia has been building the same lane from the hardware side. Its robotics blog said this week that foundation models and simulation are speeding up the jump from virtual training to real-world deployment, and Nvidia’s Cosmos line is built specifically for “physical artificial intelligence.” (blogs.nvidia.com) The reason companies care is simple: language tells a machine what a chair is, but video and sensors teach it that a chair can be behind a table, partly hidden by a person, or moved three feet to the left since the last frame. Buildings, warehouses, and robots all run into that kind of problem every second. (alibabacloud.com) If this approach works, the next layer of building intelligence will not just wait for a badge swipe or a thermostat setting. It will infer occupancy, motion, lighting conditions, and scene changes from cameras and other sensors the way a person walks into a room and instantly reads what is going on. (cnbc.com) That is why Alibaba’s $290 million check stands out. It is not just a bet that better artificial intelligence will talk better, but that the next valuable models will see space, track cause and effect, and make decisions inside the physical world. (cnbc.com)