Sand.ai raises over $100 million in new funding for video generation model
The AMW Read
Novelty 2: Sand.ai is a known player but its MoE bet and open-source commitment update the video generation landscape. Significance 2: The $100M+ raise and architectural pivot (autoregressive + MoE) have segment-level impact on multimodal/generative media.
Sand.ai raises over $100 million in new funding for video generation model
Sand.ai, a Beijing-based video generation startup founded by Cao Yue (曹越) in January 2024, has completed two funding rounds totaling over $100 million. The investment comes from a consortium that includes Look Capital, Lollapalooza Capital (Wang Huiwen's family office), Jiukun Venture, Matrix Partners, MSA Capital, Sinovation Ventures, Shunwei Capital, Matrix Partners China, China Renaissance, Hongtai Capital, Today Capital, Huaye Tiancheng, Yunhui Capital, IDG, and Baidu Ventures. The company plans to release a new MoE-architecture video generation model in Q3 2026, which it claims will achieve open-source SOTA performance across general generation, synchronized audio-video output, multi-shot narrative, and multi-reference generation.
The raise is significant because it funds a bet on three non-consensus technical choices. Sand.ai committed early to autoregressive (next-frame prediction) architecture when the market favored diffusion models; it achieved audio-video synchronous generation months after Google Veo-3 but before any other major player; and it is now converting its dense models to Mixture-of-Experts (MoE) to break what Cao calls the 'impossible triangle' of cost, speed, and quality in video generation. The company's Magi-1 model has held the #1 spot on Google DeepMind's Physics IQ benchmark, and its music agent product VidMuse reportedly reached $10M ARR within three months of launch. Sand.ai exemplifies the pattern of a foundation-model startup making counter-consensus architectural bets under compressed capital-cycle conditions.
Cao Yue's trajectory — from founding Lightyear (光年之外), its abrupt end, and now restarting with Sand.ai — is a canonical case of founder resilience in the Chinese AI ecosystem. The company's strategy of simultaneously building SOTA video models and consumer applications (VidMuse, digital human products) mirrors the hybrid model-plus-product approach seen at Anthropic with Claude Code. Sand.ai's commitment to open-sourcing its MoE model also positions it within the open-weight competition trend against incumbents like DeepSeek and ByteDance's Seedance. The real open question remains whether video generation alone can sustain the 'world model' narrative Sand.ai is aiming for, especially as Sora has been suspended and the definition of 'world model' remains contested among figures like Yann LeCun and Fei-Fei Li.