ChaoYong AI Editorial Team

On November 5, Sand.ai (Sanda Technology), a video generation foundation model company, announced the completion of a $200 million Series A+ financing round.

This marks the company's third financing round in 2026, bringing its cumulative funding to $300 million over just 10 months. With this round, the company has officially entered the unicorn club.

The round was co-led by Cornerstone Capital and Wangsu Science & Technology, with participation from Times Capital, Panlin Capital, and Orient Securities Capital. Tianchuang Capital made strategic investments alongside Shandong Tietou Industrial Investment and Shandong Railway Fund. Shanghai state-backed institutions, including Pudong Venture Capital and Shanghai Guofang, also joined, while existing shareholders such as Matrix Partners China, Huaye Tiancheng, and CAS Star continued to increase their stakes.

The proceeds will be allocated toward video foundation model R&D, training and inference infrastructure, product iteration, and global market expansion.

Founded in January 2024 and headquartered in Haidian, Beijing, Sand.ai is led by Founder and CEO Cao Yue, who earned his Ph.D. in Software Engineering from Tsinghua University in 2019. He previously served as a principal researcher at Microsoft Research Asia and headed the Vision Model Research Center at the Beijing Academy of Artificial Intelligence. He is also a co-founder of Light Year.

Cao Yue is one of the primary authors of Swin Transformer, which received the ICCV 2021 Marr Prize.

Founder and CEO Cao Yue

The company's technical roadmap has undergone three major shifts. In April 2025, it open-sourced the autoregressive video model MAGI-1, which supports continuous generation and long-duration extension. In March 2026, it open-sourced MagiHuman, an audio-video foundation model that uses a unified single-stream Transformer to jointly model text, visuals, and sound. In August 2026, it open-sourced MAGI-2 Preview, featuring approximately 114B total parameters and an ultra-fine-grained MoE architecture that activates only about 6B parameters per forward pass, achieving an inference cost of 0.5 yuan for a 10-second 1080P video.

This makes Sand.ai one of the earliest teams globally to scale video generation models to the hundred-billion-parameter level and open-source the weights. At the time of release, it ranked sixth globally on the Artificial Analysis image-to-video leaderboard.

On the commercialization front, the company's video generation product VidMuse surpassed $10 million in annualized recurring revenue within two months of launch.

Cornerstone Capital expressed long-term confidence in the company's technical roadmap. Wangsu Science & Technology and Sand.ai have entered into a strategic partnership to collaborate on audio-video inference, low-latency edge inference, and industrial applications.

Three financing rounds totaling $300 million in 10 months—Sand.ai's pace reflects the broader shift of video generation from "model demos" to "inference cost and product revenue."

The hundred-billion-parameter MoE architecture brings single-inference cost down to 0.5 yuan, and VidMuse reached $10 million ARR in two months. These two data points demonstrate the commercial viability of the technical roadmap more convincingly than the financing amount itself.

However, competition in the video generation track is shifting from model capability to controllability, inference efficiency, and ecosystem coverage. Kuaishou Kling and ByteDance Seedance each have their own scenario entry points. Whether Sand.ai's strategy of open-source weights plus independent products can generate sustained payments in the developer ecosystem and overseas markets is more worth watching than the parameter scale of the official MAGI-2 release.