Sand.ai Hits $1B Valuation: Wangsu Backs MAGI-2 Maker

👁 114 reads

Wangsu Science & Technology is paying 300 million yuan for a 4.4% stake in Sand.ai, the Beijing startup that released the open-weights MAGI-2 Preview video model in August. The September 29 transaction values the company at approximately $1.01 billion. However, corporate filings also reveal modest 2025 revenues of just 366,000 yuan alongside more than 150 million yuan in net losses during the first half of 2026 alone. I detailed the model, leaderboards, and hardware demands in the MAGI-2 Preview guide.

Who Is Investing and How Much

The strategic backer is Chinese technology firm Wangsu Science & Technology. According to China AI Dispatch, Wangsu picked up 4.4% of Sand.ai for 300 million yuan, pushing the startup into unicorn territory at a ~$1.01 billion valuation. Founder Cao Yue’s equity drops from 18.6184% to 14.7309% post-transaction. Cao Yue is an alumnus of Microsoft Research Asia and a co-author of the influential Swin Transformer architecture.

This is not Sand.ai’s first major capital injection. Three months earlier, on June 29, 36kr reported on two rounds totaling over $100 million from heavyweight investors including Matrix Partners China, Sinovation Ventures, IDG, and Baidu Ventures.

Revenue and Net Loss Breakdown

These figures reflect unaudited financial data reported by China AI Dispatch (Sep 29, 2026). While half-year revenue surged about 47-fold over the entirety of 2025, operating losses continue to outstrip revenue roughly nine times over.

Full bar width represents 186.8 million yuan. Notably, there is a discrepancy in reported metrics: Sand.ai claims on its website that its VidMuse app surpassed $10 million in Annual Recurring Revenue (ARR) within two months. That figure cannot be reconciled with official H1 revenue filings without further unreleased corporate context.

Development Timeline

• May 2025: MAGI-1 debuts as an autoregressive video model accompanied by an arXiv research paper.
• November 2025: Sand.ai decides to pivot to a Mixture of Experts (MoE) architecture, per 36kr.
• June 10, 2026: Sand.ai publishes an analysis detailing video model scaling via MoE.
• June 29, 2026: 36kr reports more than $100 million raised across two private funding rounds.
• August 3-5, 2026: MAGI-2 Preview weights hit Hugging Face under Apache-2.0 alongside a technical report. The last code commit was logged on August 6.
• August 14, 2026: Sand.ai shares details on X about a “100B+” MoE architecture featuring 114 billion total parameters and 6 billion active per token.
• September 12, 2026: A newer checkpoint, “MAGI-2 Preview (0912)”, surfaces on the Artificial Analysis text-to-video leaderboard without public weights.
• September 29, 2026: Wangsu investment closes, valuing the firm at ~$1.01 billion.
• October 11, 2026: MAGI-2 ranks 10th in image-to-video; the distilled release remains listed as “Coming soon”, and public API pricing is absent.

MAGI-2: Open by Name, Expensive in Practice

From a legal standpoint, MAGI-2’s licensing terms are pristine: true Apache-2.0 without revenue ceilings or geographic exclusions. By contrast, models like MiniMax H3 cannot legally be used in the European Union without explicit, custom authorization. But practical deployment barriers remain severe:

• The model weights weigh in at roughly 307 GB, with repository README guidelines demanding eight 80 GB NVIDIA Hopper GPUs;
• A distilled, low-step variant has been promised since August 5 and remains unavailable as of mid-October;
• There are no official ComfyUI native nodes or quantized distributions;
• You cannot run MAGI-2 through platforms like fal.ai, Replicate, WaveSpeedAI, or OpenRouter, and Artificial Analysis lists hosted API pricing simply as “Coming soon”;
• A cost breakdown published by mushroom.cv estimates that rendering a single 10-second clip on rented AWS compute costs between $25 and $42.

The model’s leaderboard placement has also cooled since launch. In the image-to-video arena, MAGI-2 holds 10th place with an Elo of 1093. While August headlines cited a 6th-place finish, subsequent entries (Wan 3.0, MiniMax H3 Max, HiDream-O1-Video, and Vidu Q4 Preview) outpaced it. In text-to-video, the unreleased 0912 build ranks 21st at 960 Elo, effectively tying LTX-2.5.

What This Means for You

• Solo creators and local workstations: Nothing changes immediately. You cannot run MAGI-2 on a consumer card like the GeForce RTX 4090, and there is no pay-per-second cloud API. For local rigs, look to models like Wan 2.2; for production APIs, top-tier options such as MiniMax H3 Max run around $4.80 per minute of generated footage.
• Enterprise teams with dedicated compute: MAGI-2 is the only open model in the image-to-video top ten with an unencumbered Apache-2.0 license. For a broader licensing and hardware breakdown, see the comparison of open video models in 2026.
• Market analysts: Wangsu’s capital runway gives Sand.ai breathing room to deliver the promised distilled variant and developer APIs. If cheaper inferences emerge, I will test and benchmark them promptly.

Explore official prompt breakdowns and timeline demonstrations in What MAGI-2 Can Do. Other recently tracked leaderboard contenders include HappyHorse 1.1 and Agnes Video 2.5.

// read moreMAGI-2 Preview: Sand.ai Open Video Model Benchmarks and Hardware SpecsAI ToolsMAGI-2 Examples: 8 Prompts, Demos, and TimelinesAI ToolsOpenAI buys back $7B in employee shares at $852B valuationMagazine

Kinetyra