On paper, Video Rebirth is too small to compete in the capital-intensive AI video battleground. Less than two years old, the startup has $80 million in funding and a team of 30 operating out of its Singapore headquarters and an office in Hong Kong. In a space where training cutting-edge video models costs tens of millions of dollars—and running it costs even more—Video Rebirth should be locked out of the race.

Yet, right before the public launch of its flagship model in May, the company carved out a spot on a benchmark leaderboard alongside the tech giants. Video Rebirth’s Bach model debuted at No. 6 on an Artificial Analysis text-to-video leaderboard, trailing behind models developed by Alibaba, ByteDance, Kuaishou Technology and xAI. It reigns as the highest-ranking startup model, with the cheapest price per minute of video generated among the top 10.

“For a team of our size, that was a strong signal that our architectural approach was working,” says Liu Wei, cofounder and CEO of Video Rebirth.

Developing AI video engines is just the opening act, according to Liu. By training AI to create visuals that are not just lifelike but bound by physical laws, he aims to build models that can generate hyper-realistic worlds. It’s a high-stakes field pursued by tech titans like Google, Meta and OpenAI, all of which are racing to develop the so-called world model that holds the promise of disrupting industries from autonomous driving and robotics to gaming. Against the behemoths, Liu says he’s building a “truly meaningful” world model, one that can understand its surroundings and simulate what will happen next, much like a human anticipating an outcome based on common sense and instincts.

“We do video generation in order to build a world model,” says Liu. “In three years, we’ll prove that the physical world can be simulated in real time.”

A beer commercial demo generated by Video Rebirth’s Bach model.

To pull it off, Video Rebirth in March closed a seed round totalling $80 million at an undisclosed valuation. Investors in the fundraising included AMD Ventures, the venture capital arm of billionaire Lisa Su’s U.S. AI chip developer Advanced Micro Devices; ZER01NE, the venture capital arm of billionaire Euisun Chung’s South Korean carmaker Hyundai Motor Group; Hiven, an investment firm affiliated with billionaire Lee Jay-hyun’s Korean food-to-entertainment conglomerate CJ Group; Korean game developer Actoz Soft; Shanghai-based Qiming Venture Partners and Gaw Capital, the Hong Kong-based private equity firm chaired by billionaire Goodwin Gaw.

Video Rebirth says it’s raising a new round in July, but declines to provide further details.

“Our rationale rests on the belief that video generation is far more than a tool for content creation; it represents one of the clearest and most viable pathways toward world models,” says Fang Wei, senior investment manager of Hyundai Cradle, a program under ZER01NE. “Video Rebirth shares this exact vision from day one, positioning its technology to unlock critical future applications in physical AI.”

Video Rebirth’s Bach targets enterprise clients in advertising, entertainment, filmmaking and gaming. The model’s signature feature is the ability to generate multi-shot videos of up to 45 seconds based on reference images and text prompts. By comparison, ByteDance’s Seedance 2.0, a popular model for multi-shot AI video generation released in February, is capped at 15 seconds, though it also allows inputs of video and audio. Other functions of Bach include producing clips of up to 10 seconds from text or images, as well as binding a static character to a reference video.

Video Rebirth is competing in a space that’s not just crowded but also expensive to operate, as generating videos requires significantly more computational power than text. The financial toll of the AI video race became clear when OpenAI in March abruptly decided to shut down its Sora platform, despite the fact that the mobile app had amassed nearly 10 million downloads since its launch last September and had secured a $1 billion equity and licensing agreement with Walt Disney (now cancelled). Forbes in November estimated that OpenAI was burning about $15 million a day to churn out millions of 10-second videos on user requests, with each costing the company about $1.30.

“OpenAI was held back by inference (the phase where trained AI is being used) cost,” says Liu. He adds that the cost for Bach to generate a 10-second clip is “significantly lower than other frontier models,” though he declines to share the exact figure, citing competitive sensitivity. The startup is able to bring down the inference cost thanks to its proprietary technology that Liu claims can speed up the video generation process by up to 10 times. Called multi-step sampling loss, it’s a mathematical technique that trains the model to anticipate and correct errors during the generation process, and therefore requires fewer steps to create the final video. Most traditional models, in contrast, can’t predict glitches and thereby take longer to run, according to Liu.

The financial efficiencies extend to training costs. Liu claims Bach required “a fraction of” the budget of comparable frontier models, though he didn’t elaborate further. The Video Rebirth head says he was able to do so by training on fewer, higher-quality videos, including licensed movies and music videos as well as clips filmed in-house, most of which have a resolution of 720p. Meanwhile, Bach was engineered to split the tasks of prompt adherence and visual generation, unlike other models that rely on a single “brain” to handle both tasks. This division of labor leads to compute efficiency, explains Liu.

In response to Liu’s claims, an OpenAI spokesperson says in an email “as compute demand grows, the Sora research team is refocusing on world simulation research to advance robotics and real-world physical tasks.”

A cinematic clip of a man running from a monster generated by Video Rebirth’s Bach model.

Beyond cost reduction, Liu says Video Rebirth also stands out with its ability to generate videos that follow the laws of physics, such as gravity, object collisions and lighting—a critical industry bottleneck where objects in AI-created clips are often morphing or uncanny. He adds that his AI is especially good at maintaining product consistency, a top priority for e-commerce advertisers, and excels at generating facial expressions and scenic shots for filmmakers. Hiven said in Video Rebirth’s fundraising announcement in March that it expects collaborations with the startup across CJ’s businesses, including entertainment unit CJ ENM, which produces K-dramas and films.

Video Rebirth’s edge lies in "its attention to enterprise-grade controllability and consistency,” says Hyundai’s Fang. He adds that the startup is addressing some of the chokepoints in video generation, including the AI’s ability to understand cause and effect, as well as how things move across space and time.

Alex Zhou, managing partner of Qiming Venture Partners, says Video Rebirth “could become a standard tool for professional content creation across industries such as film, advertising, gaming and e-commerce” in the next five years, just like “what Adobe has played in the traditional creative software industry.”

With the technology to generate objects and surroundings that are not just aesthetic but realistic and physically accurate, Video Rebirth is working on a world model that can create interactive 3D environments on the fly based on text prompts. Unlike traditional 3D simulations which require lines of code to make and can only react to what is pre-programmed, a world model is an AI that understands physical properties of the real world and simulates what will happen next, even in situations it has never “seen” before.

World models are still a nascent space, but a growing number of companies are betting this technology can be used to train self-driving cars to handle unexpected situations, teach robots to work smarter and speed up video game development. In January, Google began to roll out Genie 3, allowing users to generate any environment where they can navigate using arrow keys and prompt new events (such as adding a new object). Although Genie 3 supports interaction for only a few minutes, its release triggered a sell-off across gaming stocks, including Unity Software, over fears that the technology would make traditional game engines obsolete. The world model is currently adopted by Alphabet’s self-driving unit Waymo to test autonomous vehicles in scenarios from natural disasters to rare events like a malfunctioning truck blocking the road.

Other companies that are developing world models range from tech giants like Alibaba, Nvidia and OpenAI to well-funded startups such as Google-backed Runway and World Labs, cofounded by AI pioneer Fei-Fei Li.

World models are “somewhere between” hype and a game-changer, says Alec Wrubel, a Los Angeles-based associate partner at McKinsey. “World models today are largely in early development. They represent an important frontier in AI but are not yet at the level of fidelity or cost profile needed for broad deployment across industries.”

Liu plans to prove Video Rebirth’s world model is a game-changer, with the startup aiming to launch one by the end of 2026. Called Olympus, the model will work similarly to Genie 3, except it can also generate environmental sounds, such as the thump of collision or the clack of footsteps, according to Liu. ZER01NE said in the March announcement that it sees Video Rebirth as a “key partner for the future of mobility” with potential to use its technology “to train physical AI within hyper-realistic digital worlds.” Hyundai Motor is a major player in autonomous driving and it owns U.S. robot maker Boston Dynamics.

“As we scale up our world model, it will be able to simulate increasingly complex physical scenarios in real-time,” says Liu. “When that happens, the world model won’t be limited to gaming and embodied AI. We will be able to take on a wide range of industrial applications.”

Liu’s ambition to develop a world model was sparked in early 2024, when OpenAI unveiled the Sora video model, which the AI poster child dubbed a “world simulator.” Liu, then a Tencent distinguished scientist (a senior title the Chinese tech giant gives to elite researchers) leading the company’s development of its Hunyuan AI model from scratch, saw where the industry was heading.

“Even though it was only 2024, I felt that the large language model space had become very crowded, with tech giants already locking down their positions,” says Liu. “Physical AI, meanwhile, was a completely blank canvas. Sora convinced everyone that the physical world could be simulated, even if it seemed incredibly difficult at the time.”

Liu was certain that he could make the simulation happen, and he had the credentials to back up his conviction. Armed with a Ph.D. in computer science and electrical engineering from Columbia University, he has been researching machine learning since 2007, drawn in by his interest in mathematics. Over the years, he held research roles at IBM and Chinese ride-hailing giant Didi, alongside teaching stints at Rensselaer Polytechnic Institute and Stevens Institute of Technology in the U.S., before joining Tencent in 2016.

“Wei is a rare founder who combines world-class research capabilities with deep industry experience,” says Qiming’s Zhou. “He has consistently been one of the technical experts I trust most in AI. Whenever there was a major breakthrough in AI models, I often sought his perspective early on, and I know many executives across the technology industry did the same.”

Recognizing a window of opportunity in physical AI, Liu walked away from his well-paid job at Tencent in September 2024 to start Video Rebirth. To build the company, he assembled a team of cofounders, including ex-Tencent AI Lab director Lu Difu, former JPMorgan Chase quantitative developer Liu Peng and Dan Kong, who previously was a director of 42X Fund, an investment fund of Abu Dhabi-backed AI company G42.

While it took large language models more than two decades to reach the mainstream following an early breakthrough in 2003, when an academic paper outlining its blueprint came out, Liu predicts that the path to mass adoption for world models will be lengthier. He anticipates the next 12 months will focus primarily on technical breakthroughs within the laboratory.

Yet, Liu remains undeterred by the timeline. “I’ll pour absolute, undivided energy entirely into R&D until I successfully build a world model that’s commercially viable,” he declares. “That day is coming, without a doubt.”