CDN Veteran Wangsu Puts 300 Million Yuan Into Video Model Startup Sand.ai - a $960 Million Implied Valuation for a Company That Lost $21 Million Last Half

Chinese edge and CDN operator Wangsu Technology (300017.SZ) has become the latest strategic investor in Sand.ai, the Beijing video-generation startup behind the open-source MAGI model family. In a Shenzhen exchange filing reported on October 8, Wangsu said it will invest 300 million yuan (about $44.5 million) of its own cash to subscribe 2,105,352 A+ preferred shares of Sandai Holdings Limited at $21.1137 per share, holding 4.3963 percent once the full round closes - an implied valuation of roughly 6.8 billion yuan, or about $960 million, based on the per-share price.

The structure is a two-step bridge. As a transitional arrangement, Wangsu first extends a 300 million yuan convertible bond to subsidiary Shanghai Sandai Technology and receives A+ purchase warrants; when it exercises the warrants, the redeemed bond proceeds pay for the shares. Founder Cao Yue and co-founder Zhang Zheng hold 18.62 percent and 5.44 percent respectively through BVI vehicles, the ESOP platform holds 15.53 percent, and financial investors from earlier rounds hold 60.41 percent in preferred shares. Wangsu's stated rationale: pairing its edge network with a video model lab for low-latency, audio-video inference workloads.

What Wangsu is buying is one of China's more credible independent video research teams. Sand.ai was founded in January 2024 by Cao Yue - a Tsinghua special-grade award winner, Marr Prize recipient and core author of Swin Transformer, the hierarchical vision architecture that underpins much of modern vision modeling - with a team drawn from Microsoft Research Asia and Alibaba's DAMO Academy. Its catalog: Magi-1, an autoregressive video generation model; an audio-visual synchronized generation model (branded MagiHuman, previously described as GAGA-1); VidMuse, a music-video agent launched in January 2026; and MAGI-2 Preview, released and open-sourced in August as what the company called the first 100-billion-parameter open-source MoE video model - 114 billion total parameters with roughly 6 billion activated per forward pass, ranked sixth on the Artificial Analysis image-to-video leaderboard at release.

The financials explain why the round matters. Per the filing, Sand.ai booked revenue of 366,000 yuan for all of 2025 - about $51,000 - against a net loss of 187 million yuan; in the first half of 2026 revenue reached 17.4 million yuan (about $2.4 million) against a net loss of 152 million yuan. The company had raised two rounds totaling over $100 million as of June 2026, with backers including Kuaishou co-founder Su Hua, the family office of Meituan co-founder Wang Huiwen, Matrix Partners China, Source Code Capital, IDG and Sinovation Ventures.

For Wangsu, a CDN veteran whose core business has been squeezed for years, the stake is a wedge into the video model economy - a bet that whoever generates the world's video will need edge inference close to viewers. For Sand.ai, the money buys runway while its losses outpace revenue by nearly nine to one. Both are, in effect, wagering that open-weight video models can hold a commercial niche against ByteDance, Alibaba and Tencent - and against Kuaishou's Kling, which is reportedly preparing its own Hong Kong IPO. At a $960 million implied valuation, the market is pricing in that the technology, not the incumbency, decides.