What the filing adds

Funding announcements report the money in. A co-lead investor's stock filing is the rare case where a private AI startup's actual books enter the public record. Tech in Asia, citing Cailianshe's reporting of Wangsu's stock filing, relays unaudited first-half 2026 figures for Sand AI Group: revenue of RMB 17.4 million (approximately US$2.61 million) and a net loss of approximately RMB 152 million (US$22.8 million).

Read against the round, the loss is roughly 8.7 times the revenue. Using the reported valuation of more than US$1 billion, the implied valuation is more than 380 times Sand AI Group's reported H1 2026 revenue of approximately US$2.61 million. This compares a reported valuation with a six-month revenue figure, not annualized revenue.

These figures do not establish whether the company is commercially unsuccessful or whether its spending is justified. They establish something narrower: the disclosed first-half revenue base is small relative to the reported valuation, while the company remains heavily loss-making. At this point, the $200 million round is funding future research, infrastructure and product development against a limited disclosed revenue base.

The $10 million ARR and the $2.6 million half

The company's claimed commercial traction tells a different story. Alpha Maven, Dealroom and VentureBurn have relayed a company claim that VidMuse, Sand.ai's paid video creation service, passed US$10 million in annual recurring revenue within two months of its January 2026 launch. This is a company-reported ARR figure, not independently verified booked revenue.

ARR is an annualized recurring-revenue run rate, not the same thing as revenue actually booked during a reporting period. The sources do not provide the underlying calculation or customer-level data supporting VidMuse's claim. The filing's $2.61 million is booked revenue for the whole group across six months. The two numbers can both be technically accurate while describing very different companies, and no published source reconciles them. Until one does, the ARR belongs in the pitch column and the filing figure in the evidence column, and a reader should know which column each number came from.

The model is the strongest part of the story

What Sand.ai gives away is more concrete than what it sells. MAGI-2 Preview's weights and inference code are public under Apache 2.0. The repository describes MAGI-2 Preview as a 114-billion-parameter MagiMoE model, with approximately 6 billion active parameters per token. It supports text-to-video and image-to-video generation, producing 10-second clips with sound generated alongside the video. Generation runs in two stages: an initial 512 × 896 preview followed by a refiner that produces 1088 × 1920 output. VentureBurn reports that the company points to MAGI-1's earlier win on the Physics-IQ video continuation benchmark as evidence of its research progress. That is a company-side benchmark claim, not an independent evaluation presented in this funding announcement.

The openness has a hardware condition. The downloadable preview is approximately 307 GB of weights: around 228 GB for the preview transformer, 56 GB for the Qwen3.5-27B text encoder and 14 GB for the refiner. The repository's documented setup requires eight NVIDIA Hopper GPUs. The video VAE is based on Wan 2.2. VentureBurn and AI Tool Lab report a cost estimate of approximately ¥0.5 (roughly US$0.07) for a 10-second clip from a distilled version of the full MAGI-2 running at scale. That distilled model has not been released, and the company has not provided a release date. This cost estimate does not describe the downloadable MAGI-2 Preview. The open weights are available now. The cheap inference is a plan.

Scale puts the bet in proportion. Dealroom reports that Kling's unit closed a financing round of approximately US$2.9 billion in October at a reported valuation of around US$18 billion, with quarterly revenue above US$116.9 million, up more than 200% year over year. That quarterly revenue is roughly 45 times Sand AI Group's reported H1 2026 revenue of US$2.61 million, a six-month figure. The periods are not directly comparable, but the difference illustrates the scale of the two businesses' reported commercial activity. For Sand.ai to narrow that gap, the full MAGI-2 would need to arrive, and the company would need to turn its technical work into commercial adoption. Neither the timing nor the commercial outcome is established.

Sources

  1. [1] Tech in Asia, "Chinese video startup Sand AI raises $200m" (November 5–6, 2026; read first-hand at Gate 1: the $200 million Series A+ led by Cornerstone Capital and Wangsu Science & Technology, use of funds, the 10-month total of US$300 million, Wangsu's previously disclosed plan to invest RMB 300 million (US$45 million) including convertible bonds and warrants, and the unaudited figures in Wangsu's filing sourced to Cailianshe — revenue of RMB 17.4 million (US$2.61 million) and a net loss of about RMB 152 million (US$22.8 million) in H1 2026, with the USD figures being the report's own conversions; the co-lead firm's own English name is CoStone Capital, per the firm's site)Read source
  2. [2] Dealroom News, "Sand.ai raises $200M Series A extension, joins unicorn ranks" (November 5, 2026; read first-hand at Gate 1: the co-leads and investor list, founder Cao Yue, the third raise of 2026, and the Kling comparison — a roughly $2.9 billion October raise at a reported valuation near $18 billion, with second-quarter revenue above US$116.9 million, up more than 200% year on year)Read source
  3. [3] Sand.ai, MAGI-2 Preview official repository (Hugging Face; read first-hand at Gate 1: 114B parameters with about 6B active per token, MagiMoE architecture, Apache 2.0 license, roughly 307 GB of weights — 228 GB preview transformer, 56 GB Qwen3.5-27B text encoder, 14 GB refiner — eight NVIDIA Hopper GPUs required, 10-second clips only, two-stage generation from 512×896 to 1088×1920, video VAE from Wan 2.2)Read source
  4. [4] Alpha Maven, "Chinese AI video generation startup Sand.ai enters unicorn club after $200m round" (November 5, 2026; read first-hand at Gate 1: VidMuse's company-claimed ARR of over $10 million within two months of its January launch, the reported valuation above $1 billion, the model lineup)Read source
  5. [5] VentureBurn, "Sand.ai Raises $200 Million To Rival Kling, Seedance AI Video" (November 5, 2026; MAGI-1's Physics-IQ video continuation benchmark win as a company-side claim, the distilled-model cost reporting, MagiAttention)Read source
  6. [6] AI Tool Lab, "MAGI-2 Review 2026" (reported inference cost at scale of about ¥0.5, roughly $0.07, per 10-second clip, with the full distilled version described as "coming soon" and not part of the preview release)Read source