Sand.ai closed a USD 200 million A+ round on 5 November and qualified as a unicorn on the back of it. Across ten months the Beijing company has raised USD 300 million, roughly RMB 2.001 billion.
The valuation is not a press-release number. It can be derived from a filing by Wangsu Technology on 29 September, which is more detailed than the announcement. Wangsu plans to commit RMB 300 million of its own funds, about USD 44.45 million, to subscribe for roughly 2.1054 million A+ preferred shares issued by Sand.ai’s holding entity. If every investor in the round closes, Wangsu expects to hold 4.3963 per cent.
Work that back and the post-money valuation is about USD 1.01 billion, just over the unicorn line.
The other numbers in the same filing
The same document contains a second set of figures that matter more than the valuation. In the first half of 2026, Sand.ai’s holding group recorded revenue of RMB 17.398 million and a net loss of about RMB 152 million. For the whole of 2025, revenue was RMB 366,000 and the net loss was about RMB 187 million.
Sand.ai started in January 2024 and is headquartered in Haidian, Beijing. It builds foundation models and applications for video generation. Founder and chief executive Cao Yue took a Tsinghua doctorate in 2019, was a principal researcher at Microsoft Research Asia, and is one of the authors of Swin Transformer, which won the Marr Prize at ICCV 2021. He also co-founded Guangnian Zhiwai, a company later acquired by Meituan.
The Wangsu filing names him as the actual controller of Sand.ai’s holding group. His previous venture ended in an acquisition. This time he is building an independent model supplier.
Three architectures in two years, all open
In April 2025 Sand.ai open-sourced MAGI-1, an autoregressive video generation model, when diffusion was the mainstream and few teams were working autoregressively. The company had thirty to forty people across data, algorithms and evaluation. Cao later said he had underestimated how hard the route was. The model did take first place on the video continuation task in Physics-IQ.
In March 2026 the team open-sourced Magi Human, putting text, image and audio into a single backbone network. In August it released MAGI-2 Preview at roughly 114 billion total parameters with about 6 billion active per forward pass. The distilled model generates ten seconds of 1080p video at an inference cost of about RMB 0.5.
Matrix Partners added to the round, its fourth consecutive investment. Investors are paying a billion dollars for a company at that revenue level, and what they appear to be buying is a two-year record of shipping on schedule. Autoregressive, then joint audio and video, then a hundred-billion-parameter mixture of experts. Three architecture changes in two years, each with the weights published.
The comparison inside the same category sets the scale. In July Kuaishou announced it would raise outside money for Kling at an expected valuation of about USD 18 billion. The capital increase closed on 30 October and investors paid roughly RMB 19.802 billion. Kling booked more than RMB 850 million of revenue in the second quarter of 2026, up more than 200 per cent year on year.
USD 1.01 billion against USD 18 billion is a gap of eighteen times. The revenue gap is wider still. Kling is closed and sits behind Kuaishou’s distribution and customer base. Sand.ai is an independent startup that publishes its weights.
What capital is underwriting is whether an independent third-party model supplier can still hold a position. One concern reported by 36Kr is that model suppliers are themselves building content platforms, so customers have to worry about competing with a vendor and about being tied to a single one.
Why a CDN company is leading
The lead investor has its own arithmetic. Wangsu runs content delivery and edge computing with more than 3,000 nodes of its own, over 1,000 of them overseas, and has launched edge GPU compute products. Video inference has to run close to users. The filing lists audio and video inference, low-latency edge inference, and intelligence solutions for government, enterprise and internet customers as the areas of cooperation.
It also states that specific purchase orders and revenue contribution have not been disclosed. The RMB 300 million Wangsu is investing is still equity and convertible debt. It has not yet become business.
On 16 October Wangsu signed a convertible bond agreement and an A+ preferred share and warrant purchase agreement with Sand.ai Cayman and Shanghai Sandai Technology, disclosing progress on 19 October. It first extends RMB 300 million of convertible debt to the Shanghai entity in two tranches and receives A+ warrants. Once outbound direct investment approval clears and conditions are met, the convertible amount is called back and used to subscribe for preferred shares in the Cayman entity. Completing a round and the money arriving are separated by an approval, and the 4.3963 per cent stake is explicitly conditional on every investor in the round closing.
Sand.ai’s case is therefore narrower than the headline. It is a bet that a small team that shipped three architectures in two years can keep shipping, that open weights are a viable position against platform-owned models, and that the video generation market rewards engineering cadence more than it rewards revenue today.
The company has raised enough to test that. It has not yet raised enough to be wrong twice.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.ofweek.com/ai/2026-11/ART-201715-8120-30710502.html.