Tencent's Former AI Chief Bets $80M on World Models Through Video Startup

Liu Wei left Tencent in September 2024 with a conviction that felt almost inevitable: the next frontier in artificial intelligence wasn't language, but the physical world itself. He had watched OpenAI unveil Sora earlier that year, a video model the company called a "world simulator," and something clicked. The large language model space was already locked down by giants. But physical AI—the ability to teach machines to understand and predict how the real world actually works—was still blank canvas. So he walked away from a senior research position to start Video Rebirth, a Singapore-based startup that would attempt something most observers thought required the resources of a major tech company.

Nine months later, Video Rebirth had raised $80 million from investors including AMD Ventures, Hyundai Motor Group, and CJ Group. The company employed thirty people across Singapore and Hong Kong. And its flagship video model, called Bach, had landed at number six on an Artificial Analysis text-to-video leaderboard—the highest-ranked model from any startup, and the cheapest to operate among the top ten. This was not supposed to happen. Video generation is brutally expensive. OpenAI had burned roughly $15 million daily running Sora before shutting it down in March, each ten-second clip costing the company about $1.30 to produce. Yet Liu claimed Bach could generate video at a fraction of that cost, thanks to a proprietary technique called multi-step sampling loss that allowed the model to anticipate and correct errors during generation, requiring fewer computational steps overall.

The Bach model itself was engineered to handle a specific market: enterprise clients in advertising, entertainment, film, and gaming. It could generate multi-shot videos up to forty-five seconds long from reference images and text prompts—nearly three times longer than ByteDance's competing Seedance 2.0. What set it apart, according to investors and Liu himself, was not just speed or cost but physical plausibility. Objects didn't morph or behave uncannily. Gravity worked. Collisions happened. Lighting obeyed the rules. For e-commerce advertisers, products stayed consistent. For filmmakers, facial expressions and scenic shots held together. These were not trivial advantages in a field where AI-generated video still often looked subtly wrong in ways viewers couldn't quite name.

But Bach was only the opening move. Liu's actual target was something far larger: a world model called Olympus, planned for launch by the end of 2026. Unlike traditional 3D simulations that required code and could only respond to pre-programmed scenarios, a world model would be an AI system that understood the physical properties of reality and could simulate what would happen next, even in situations it had never encountered before. Olympus would generate interactive 3D environments on the fly from text prompts, complete with environmental sounds—the thump of collision, the clack of footsteps. It would work like Google's Genie 3, which had already spooked gaming stocks when it launched in January, but with audio and, Liu believed, greater sophistication.

The applications were staggering if the technology worked. Autonomous vehicles could be trained in scenarios from natural disasters to rare events like a malfunctioning truck blocking the road. Robots could learn to navigate and manipulate objects in hyper-realistic digital worlds. Game developers could accelerate production. Hyundai, which owned Boston Dynamics, saw Liu's technology as a potential key to the future of mobility and physical AI. Yet the field remained nascent and contested. Google, Meta, OpenAI, Alibaba, Nvidia, and well-funded startups like Runway and World Labs—cofounded by AI pioneer Fei-Fei Li—were all racing toward the same finish line. McKinsey's assessment was measured: world models were somewhere between hype and genuine game-changer, still in early development, not yet at the fidelity or cost profile needed for broad deployment.

Liu's credentials gave him standing to make the attempt. He held a Ph.D. in computer science and electrical engineering from Columbia, had been researching machine learning since 2007, and had held positions at IBM, Didi, and teaching roles at Rensselaer and Stevens before joining Tencent in 2016. His cofounders included Lu Difu, former director of Tencent's AI Lab; Liu Peng, a former JPMorgan Chase quantitative developer; and Dan Kong, who had previously directed an investment fund for the Abu Dhabi-backed AI company G42. These were not first-time entrepreneurs. They were researchers and technologists who had decided the moment was now.

Liu acknowledged that the path to mainstream adoption for world models would likely be longer than it had been for large language models, which took more than two decades to reach the mainstream after an early breakthrough in 2003. He expected the next twelve months to focus primarily on technical breakthroughs in the laboratory. But he was undeterred. "I'll pour absolute, undivided energy entirely into R&D until I successfully build a world model that's commercially viable," he said. "That day is coming, without a doubt." Video Rebirth was already raising a new funding round in July. The bet was that a small team with the right architecture and the right founder could move faster than the giants, and that the physical world, unlike language, was still a frontier where speed and focus might matter more than scale.