NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Inside Vidu's 1.9-second video: a US$83 million wager on China's AI video (AI视频) lead

Shengshu Technology's Vidu closed a >600 million yuan round and open-sourced a framework rendering five seconds of video in under two.

2026-11-15 · 915 words · NeuroAI
Inside Vidu's 1.9-second video: a US$83 million wager on China's AI video (AI视频) lead

A filmmaker in Hangzhou opens her laptop at midnight. She types a sentence — "a paper lantern drifts down a rainy alley, neon reflecting on wet stone" — and nineteen seconds later she is scrubbing through a clip good enough to storyboard a commercial. Nineteen seconds is not a rendering time anyone expected from a Chinese large model (大模型) two years ago. Those nineteen seconds are the hosted SaaS end-to-end latency; the 1.9-second figure in the title is a raw framework speed measured on a single RTX 5090 consumer card — a different metric, not a contradiction.

Vidu is the video model built by Beijing Shengshu Technology (生数科技), developed with researchers from Tsinghua University. It first went public on April 27, 2024, billed as China's first long-duration, high-consistency, high-dynamics video large model, initially generating 16-second 1080p clips from text or images using a U-ViT architecture that fuses diffusion and transformer methods. Earlier this year the company turned that research into a business story the market could price.

A funding round that signals confidence

In early 2026, Shengshu closed a Series A+ round of more than 600 million yuan (≈US$83 million), led by Zhongguancun Science City and Xinglian Capital, with strategic investment from Wanxing Technology (万兴科技), Visual China (视觉中国) and TRS (拓尔思), and follow-on capital from existing backers including Qiming Venture Partners (启明创投) and the Beijing AI Industry Investment Fund.

That size of check matters because video generation is the most compute-hungry corner of generative media. Training and serving these models burns through GPUs at a rate that quickly separates funded labs from hobby projects. A round of this magnitude tells the market that institutional investors expect video to become a paid, production-grade tool rather than a free demo.

The 1.9-second claim

The number that travels furthest is speed. In December 2025, Shengshu open-sourced TurboDiffusion, a framework that it says generates five seconds of video on a single consumer RTX 5090 card in 1.9 seconds — roughly a 100–200x efficiency gain over its earlier pipeline. The company frames this as moving video generation from "minutes per clip" to "seconds per clip" on hardware a studio can actually buy.

Speed is not vanity here. Video teams judge tools by throughput: how many usable seconds a studio can produce per GPU-hour. On the benchmarking site Artificial Analysis, Vidu's generation speed has been measured at about 10x that of OpenAI's Sora 2 and roughly 2x that of Google's Veo 3 Fast and Grok's image-to-video model. Those are third-party latency comparisons, and they describe speed, not artistic quality — but for anyone running a content pipeline, latency is the bill.

What Vidu actually sells

Vidu is not one product. Shengshu has assembled a stack: Vidu MaaS for developers, Vidu SaaS for creators, and Vidu Agent for task-based generation. The company reports that in 2025 its users and revenue each grew more than tenfold, and that Vidu has become a default choice for commercial content creation across advertising, animation, film, education, gaming, hardware, cultural tourism and broadcasting.

The technical story runs through reference-to-video consistency. Version 1.5, released in November 2024, added multi-entity consistency — letting a creator hold a character, an object and a setting visually stable across a single clip. Later Q-series updates pushed subtle facial micro-expressions and synchronized audio. These are the features that separate a "fun clip" from a "usable shot," and they are why advertising agencies, not just individual creators, have started to pilot the tool.

Why China is investing here

The strategic logic is straightforward. A sovereign, fast, cheap AI video (AI视频) large model (大模型) lowers the cost of producing localized advertising, training material, short drama (短剧) and game cinematics in dozens of languages. Shengshu is one of several Chinese teams — alongside Kuaishou's Kling, ByteDance's Seedance and Alibaba's Wan — racing to own that layer. The A+ round suggests investors believe the winner will capture recurring enterprise revenue, not one-off consumer curiosity.

The open-weight question

Shengshu's strategy mirrors a wider Chinese pattern: give away the efficiency technology, keep the flagship model behind a paid API. TurboDiffusion is open, but Vidu's best quality tiers remain commercial. That hedging lets the company recruit developers and burnish its research reputation without surrendering the revenue engine. It also blunts a criticism often aimed at closed Western labs — that they hoard capability — while still protecting the margin that funds the next training run. Whether the community adopts TurboDiffusion widely enough to matter is the open question.

Honest limitations

This article relies on company announcements and a Xinhua-affiliated financial report for the funding figure, the TurboDiffusion benchmark and the revenue-growth claim; independent audited financials are not public. The "10x faster than Sora 2" comparison is a latency metric from Artificial Analysis and does not measure visual quality, prompt adherence or safety. I did not test Vidu myself, and per-second generation quality varies sharply with prompt complexity. The piece also does not cover content-labeling (内容标识) obligations for published clips, enterprise pricing tiers, or how Vidu compares with rivals on long-form narrative coherence rather than single-shot consistency.

What readers can do now

  • If you produce video at volume, run a side-by-side pilot of Vidu against one Western model on your own prompts; judge by throughput and consistency, not demo reels.
  • For Chinese-language or China-market content, test Vidu's reference-to-video feature for brand-character consistency before committing to a workflow.
  • If you run your own GPUs, track the open-source TurboDiffusion release — the efficiency claim, if reproducible on your hardware, changes your cost math.

Related coverage

Cite this article

NeuroAI (脑机网), "Inside Vidu's 1.9-second video: a US$83 million wager on China's AI video (AI视频) lead," 15 November 2026, https://neuroai.site/a/na-aigc-vidu-shengshu-video

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Generative Media” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories