Shengshu’s First AI Feature Film Misses the Mark as Its Hong Kong IPO Nears

NextFin News — Shengshu Technology spent years arguing that its Vidu model belongs inside professional film pipelines. China’s first AI-assisted feature to clear official release approval was meant to prove the point. Eleven days after opening, Sanxingdui: Future Past had taken in about 1.68 million yuan at the box office. On weekdays, average attendance per screening fell below one person.

Bona Film Group carries the producer’s stigma. Shengshu, listed as chief technical support, faces a different problem. A commercial showcase timed to a listing campaign became a public counterexample instead.

The company is less than four years old, tied to Tsinghua University’s AI research circle, and has raised more than 60 billion yuan this year. Bloomberg has reported a post-B-plus valuation above $2 billion. Hong Kong flotation plans are under way. Market mood has shifted since the early-year wave of AI listings. MiniMax has traded below its issue price; hardware names in the U.S. have sold off; “AI bubble” talk has returned. Buyers care less about who shipped first and more about who can keep earning after prices collapse.

Founder Zhu Jun is a vice dean at Tsinghua’s AI institute. The team has long stressed an early U-ViT architecture and positioned Vidu, launched in 2024, as one of the first Chinese models to answer OpenAI’s Sora in public.

Creators and overseas reviewers most often praise reference-image-to-video: a few stills of a character, prop, or set, held consistent across shots. That fits anime-style work, brand campaigns, and series with fixed casts. Consistency tends to slip once reference counts rise. On longer clips and live-action faces, rivals such as Kuaishou’s Kling often rank higher in independent tests.

Third-party leaderboards tell the same split story. Vidu has touched the top of some image-to-video tables and disappeared from strong positions on pure text-to-video boards. Static or tightly controlled frames in the Sanxingdui film looked serviceable; performance and extended narrative stretches did not.

Speed and price are the other pitch. Shengshu has highlighted open tooling that can produce short clips in seconds on a single consumer GPU. Early commercial traction was real: the company said Vidu crossed $20 million in annualized recurring revenue within eight months of launch, with more than half of generation volume coming from ads, e-commerce, and animation. Those customers buy seconds of video the way they buy stock footage—high volume, low loyalty, and endless price pressure. Recent list prices for lower-resolution output have fallen toward a few fen per second. Kling, MiniMax, and global models have followed. Platform owners can subsidize inference; a pure-play startup cannot.

Video APIs alone rarely support a multi-billion-dollar story. Over the past year Shengshu has pushed harder into “world models”: open-sourcing Motus with Tsinghua, releasing Motubrain, and showing Motus2 aimed at robotic manipulation. By the time of its $500 million B-plus round, it was describing itself as a general world-model company—Vidu as generation in the digital world, robotics as the physical half.

The label is fashionable. Dozens of firms have claimed it this year. Fei-Fei Li’s World Labs has called the term among the most important and most abused in the field. Chinese research leaders have noted that current leaderboards still measure generation quality, not physical understanding. Yann LeCun has long argued that predicting pixels is a poor route to models of action. Academic work has shown video generators can look flawless inside their training distribution and fail outside it—closer to nearest-neighbor mimicry than learned physics.

Shengshu’s founder counters that video already encodes time, space, motion, and cause, and cites internal multi-task gains over selected baselines. Those results remain largely company-reported and simulation-based. Factory and household deployment is still distant. OpenAI’s decision to wind down consumer Sora products while treating world simulation as longer-horizon research underscores the commercial gap: video generation is hard to monopolize; embodied research is expensive to fund without a cash engine elsewhere.

Alibaba Cloud led a large B round and listed Vidu on its model platform the day before. The same group builds competing video systems. Capital and distribution help; they also make Shengshu replaceable inside a bigger ecosystem.

IPO timing has already slipped—from talk of a first-half filing to a more cautious “next year” window. Higher private valuations raise the bar for a public debut just as secondary investors demand proof of durable revenue. The Sanxingdui film was supposed to supply narrative cover. Instead it clarified where AI video sits today: useful for controlled, short-form commercial work; not yet a substitute for cinematic craft audiences will pay to sit through.

Shengshu is not alone. When this listing wave settles, fewer independent video-model companies are likely to remain on the board. The ones that do will need either a cost structure that survives commodity pricing or a world-model story that survives contact with real robots and real physics—not only with a theatrical poster.