A $200 million round makes AI video start-up Sand.ai a unicorn

Wangsu, a content delivery network (CDN) firm, committed 300 million yuan of its own money, 36Kr reports. Wangsu wants to work with Sand.ai on low-delay video generation at the edge of its network, close to users, Ventureburn reports.

Founder Cao Yue, formerly a researcher at Microsoft Research Asia, chose an autoregressive design, which builds a video step by step, when most of the field favoured diffusion. Sand.ai has released MAGI-1, the joint audio and video model GAGA-1, and an open-source MAGI-2 Preview with 114 billion parameters (the model’s learned settings), of which about 6 billion work on each pass. Within two months of going on sale in January, VidMuse, its paid app for anime dramas, music videos and product ads, passed $10 million in annual recurring revenue.

A distilled MAGI-2 model, a smaller copy trained to imitate the full one, makes a 10-second 1080p clip for about 0.5 yuan, 36Kr and Ventureburn report.

On our price sheet, Kling 3.0 with audio costs $0.168 a second at 1080p on its API and Veo 3.1 Fast $0.12 on Google’s API, our 8 November check.

The 0.5 yuan figure is an inference cost reported by 36Kr and Ventureburn, and Sand.ai publishes no MAGI-2 price, so twenty 10-second 1080p clips stay at $24 on Veo 3.1 Fast and $33.60 on Kling’s API as the prices an ad team can plan with.