Sand.ai's $200M bet: the video startup that refused to follow the diffusion crowd
On November 5, Beijing-based Sand.ai announced a $200 million Series A+ round co-led by Cornerstone Capital and Wangsu Science & Technology. The company says the deal makes it a unicorn. It also takes its 2026 fundraising to roughly $300 million across three rounds, which is a lot of capital for a company founded in January 2024.
The headline number is big, but the more interesting story is how Sand.ai got here. For most of the past two years, the video-generation market has run on one technical recipe. Sand.ai picked a different one, stuck with it, and has now been paid handsomely for that stubbornness.
A contrarian architecture
Most well-known video models, from OpenAI's Sora to many of the Chinese contenders, are built on diffusion. Very roughly, a diffusion model starts with noise and refines a whole clip at once until it matches the prompt. It produces beautiful short clips, but it is awkward for long, coherent sequences and for anything that needs to respond as it plays out.
Founder Cao Yue took the other road. Sand.ai's first model, MAGI-1, released in April 2025, is autoregressive: it generates video chunk by chunk, each piece conditioned on what came before, much like a language model predicts the next word.
In a June interview with 36Kr, Cao described this as a deliberate choice made when the market mainly backed diffusion. The pitch is that predicting the next moment of a scene is closer to how physics actually works. According to 36Kr, MAGI-1 held the top spot on Google DeepMind's Physics-IQ benchmark for a long stretch.
That matters beyond benchmarks. Autoregressive generation is the natural fit for longer clips, live or interactive video, and eventually "world models" that simulate environments for robots and games. Cao has said he sees video as the most important data path toward those world models.
MAGI-2 and the cost problem
The new money arrives just ahead of the full release of MAGI-2. A preview shipped on August 5: a 114-billion-parameter mixture-of-experts (MoE) model that activates only around 6 billion parameters per token. Text, video and audio tokens share a single sequence, so picture and sound are generated together in one pass rather than stitched together afterwards.
The preview outputs 10-second clips with synchronized audio, upscaled to 1088×1920, and the weights, inference code and technical report are public.
The MoE design is a business decision as much as a technical one. Cao told 36Kr that a dense model would cost three to five times more to run for the same result. In video, where every second of output burns serious GPU time, that gap decides whether a product can make money.
Even so, the preview needs eight Nvidia Hopper GPUs for inference and a checkpoint set of about 307 GB. This is not something hobbyists will run on a laptop.
Why investors are paying up
Three things stand out in Sand.ai's story for anyone tracking the AI video market.
It pairs models with products. Sand.ai launched VidMuse, a music-video agent, in January 2026. 36Kr reported it reached annual recurring revenue in the tens of millions of dollars within three months.
That is unusual for a lab this young, and it supports Cao's argument that model makers and application builders should be the same company. In his view, startups without a state-of-the-art model risk being absorbed by those who have one.
It open-sources aggressively. Releasing MAGI-1 and the MAGI-2 preview openly has given Sand.ai developer mindshare that a closed lab its size could not buy. It follows the playbook that made Chinese labs such as DeepSeek and Qwen influential: publish strong open weights and let adoption build the brand.
The investor list is deep. Before this round, the company had already closed two rounds totaling over $100 million, with backers including Matrix Partners China, Sinovation Ventures, IDG, Baidu Ventures, Source Code Capital and Lollapalooza Capital, the family office of Meituan co-founder Wang Huiwen. Bringing in Wangsu, a content-delivery and edge-computing company, as co-lead looks strategic. Video generation is bandwidth- and compute-hungry, and distribution infrastructure is a natural partner.
The hard part starts now
None of this guarantees a win. The competition is brutal. Google's Veo, OpenAI's Sora, ByteDance's Seedance and Kuaishou's Kling all have far bigger budgets and built-in distribution through platforms with hundreds of millions of users.
Sand.ai has raised around $300 million this year; its rivals measure compute budgets in billions.
There is also the question of whether autoregressive video can match diffusion on pure visual polish at scale. MAGI-1 won on physical plausibility, but consumers and advertisers usually pick tools on how good the output looks and how fast it renders. MAGI-2's full release will be the real test.
And the cost curve is still steep. Eight top-end GPUs for a 10-second clip is a long way from the economics needed for mass consumer use. The MoE approach helps, but Sand.ai will need to keep cutting inference costs to turn a research lead into durable margins.
What to watch
For founders and operators, Sand.ai is a reminder that a well-argued technical bet against consensus can still attract top-tier capital, as long as it comes with real revenue.
Three signals over the next six months will show whether the bet is paying off:
How MAGI-2's full release compares with Veo, Sora and Kling on independent quality and speed benchmarks.
Whether VidMuse and future agents keep growing revenue without heavy subsidies.
Whether Sand.ai ships real long-form or interactive video, the use case where autoregressive models should have a clear edge.
If those land, this $200 million round will look like the moment the video race stopped being a one-architecture contest.