Sand.ai raised $200 million ahead of the release of video model MAGI-2

In ten months, Sand.ai has transformed from an early-stage project into one of the most discussed artificial intelligence startups in China. The company raised $200 million in a Series A+ round, aiming to become a unicorn valued at over $1 billion. Cornerstone Capital and Wangsu Science & Technology led this round. New investors include Times Capital, Panlin Capital, Orient Securities Capital, Pudong Venture Capital, and Shanghai Guofang.

Sand.ai has raised approximately $300 million in funding and is now targeting competition with video generation giants—Kling and Seedance. The company was founded in 2024 by Chao Yue, who previously worked as a senior research scientist at Microsoft Research Asia. His prominence stems from the joint development of the Swin Transformer model, which won the best paper award at the ICCV conference.

This experience shaped the company's strategic decision: Sand.ai chose an autoregressive approach to video, while most of the industry favored diffusion models. Chao later admitted that the team underestimated the complexity of the chosen path.

In May 2025, Sand.ai released the open-source project MAGI-1, which won the Physics-IQ video continuation benchmark. Then, in August 2026, a preliminary version of MAGI-2 was presented—a mixture-of-experts model with 114 billion parameters. Because it is only active about 6 billion times per pass, its cost remains low. The distilled MAGI-2 model can generate a 10-second video in 1080p resolution for approximately 0.5 yuan.

According to BigGo Finance data, the cost per second of video is about one-tenth the cost of mainstream models. Developers also point to the MagiAttention library, used by Tencent's Hunyuan team and supported via NVIDIA's NeMo AutoModel. This is a quiet but significant endorsement. Furthermore, there is a product side: the AI video creation application VidMuse generated $10 million in annual recurring revenue within two months.

It is important to note that many AI labs demonstrate impressive examples but struggle with product sales. Sand.ai demonstrates the ability to do both, which is exactly what investors are currently looking for.

Wangsu is not just providing financial support; 300 million yuan ($45 million) of capital from this company, a giant in CDN and edge computing, is directed toward this project. Wangsu is interested in collaborating with Sand.ai on low-latency edge video inference. Given clients such as ByteDance, Tencent, and JD.com, this partnership could provide Sand.ai's models access to very large customers.

The raised funds will be used for model research, training, inference infrastructure, product improvement, and expansion into overseas markets. In a broader context, the AI video race in China is accelerating as major players already have their own platforms. Kuaishou recently raised about $2.9 billion, valuing the company at $18 billion. ByteDance's Seedance reportedly generates over 1 billion yuan in monthly revenue.

Sand.ai is choosing a different path: it is independent, open-source, and operationally economical. This is attractive to developers and brands that do not want to depend on competitors' platforms. Nevertheless, risks exist: foundational models continuously consume capital, and early successes can quickly fade. However, if MAGI-2 meets expectations, Sand.ai could become a neutral choice in the AI video sphere, which warrants close monitoring.

GenRobot AI, a Beijing-based company that provides data infrastructure for robot development, has successfully closed a Series A+ funding round totaling $100 million. This financing closure occurred in the third quarter of the year, as announced on November 2.

Investors participating in the funding round included SAIC Motor, Eastworlds from Malaysia, Baidu, Shoucheng Holdings, and Galaxy General Robotics. Additionally, financial investors included GL Ventures, CITIC Goldstone, CCB International, and Huatai Zijin.

GenRobot AI is engaged in creating software and hardware for collecting demonstrations of human actions, which are subsequently used to train robots. The company's recording systems capture visual information along with hand and body movements when performing tasks in daily life, manufacturing, and logistics. Subsequently, the company's Matrix platform synchronizes, cleans, and labels these recordings, providing robot developers with the necessary data for training and evaluation.

Jingshuo, a subsidiary of WeRide specializing in developing data infrastructure for physical artificial intelligence, has successfully raised tens of millions of dollars in a pre-Series A funding round.

Investors in this round include Vision Knight Capital, Hong Kong Science and Technology Parks Corporation, and Sirius Capital.

The company plans to use the raised funds to expand its research and development team, as well as to develop its data platform, world models, and systems for creating reusable robot skills. Additionally, Jingshuo intends to produce more high-quality datasets covering various tasks and operating conditions.

Jingshuo originated from WeRide's efforts in physical AI research. Its CEO, He De, was the first employee of WeRide and played a key role in establishing engineering teams in China, the US, and Singapore. After completing the financing, Jingshuo remains a wholly-owned subsidiary of WeRide, operating autonomously, while its financial results are consolidated in the parent company's reports.