By Hamza Faisal
Chinese AI video startup Sand.ai has raised $200 million in a Series A extension, pushing the two-year-old company into unicorn territory as investors pour money into the race to build the next generation of video models.
The round, announced on November 5, was co-led by Cornerstone Capital and Wangsu Science & Technology, with Matrix Partners China, Panlin Capital, Times Capital and state-backed funds including Shanghai Guofang also taking part. It is Sand.ai's third funding round of 2026, bringing its total raised in the past 10 months to about $300 million.
The company did not disclose an exact valuation, but reports put it above $1 billion. Wangsu had already signalled its confidence in September, when it disclosed plans to invest RMB 300 million in Sand.ai at an implied valuation of about RMB 6.8 billion.
A research lab betting on video as the path to world models
Sand.ai was founded in January 2024 by Cao Yue, a former principal researcher at Microsoft Research Asia and co-author of the Swin Transformer, one of the most influential computer-vision papers of recent years. The Beijing company builds foundation models for video under its MAGI series, arguing that video is the most direct route to AI systems that understand the physical world.
Its earlier Magi-1 model, released in early 2025, topped Google DeepMind's Physics-IQ benchmark for physical realism, according to the company. In August, Sand.ai open-sourced MAGI-2 Preview, a roughly 114-billion-parameter mixture-of-experts model that activates only about 6 billion parameters per inference, a design meant to keep the cost of generating video down as models grow. The full MAGI-2 model is due to be released soon.
The company says the new money will go into research and development of its video foundation models, training and inference infrastructure, product iteration and global expansion.
Early revenue, and a very expensive race
Sand.ai is not only a research outfit. Its music-video creation product VidMuse, launched in January 2026, passed $10 million in annual recurring revenue within two months, giving the company an early commercial foothold while its models are still being built out.
Even so, the contest it has entered is one of the most expensive in AI. Kuaishou's Kling unit closed a roughly $2.9 billion raise in October at a reported valuation near $18 billion, and OpenAI's Sora 2 has set a high bar for realism. Video models demand enormous computing power to train and run, which is why even well-funded startups keep returning to investors.
For Sand.ai, the open-source strategy is part of the bet: by releasing large models publicly, it hopes to widen adoption and attract developers who might otherwise build on a rival's system. Whether a $300 million war chest is enough to keep pace with rivals raising billions is the question the next year will answer.
What happens next
The immediate test is the full release of MAGI-2, which the company has said is coming soon. Its reception among creators and developers will show whether Sand.ai's mixture-of-experts approach can match the quality of closed rivals at a fraction of the running cost.
If it lands, Sand.ai will have proved that a focused, research-led startup can still compete in a field where scale usually wins. If it does not, the company will find itself in a market where the biggest players are spending ten times as much.



