A Chinese Studio Runs a World Model on One RTX 5090 and Sells It by the Shot

The demo that sells this company is not a render. It is a live feed.

A presenter stands in an ordinary office. The floor collapses. Something with tusks comes up through it. A chair beside him warps into a wolf-headed throne and he sits down. A tree grows out of a plant pot, a curtain catches fire, and then the whole room turns into an animated grassland.

Nobody in that frame is composited. The person is real, the camera is real, and the creatures, props and set are generated in real time on a single consumer RTX 5090. That is the claim Dawa Technology — a Chongqing company that describes itself as a generative world foundation model shop — made when it launched Scenes AI in September, and it is the reason its name is suddenly showing up in investor conversations.

What they actually built

The architecture is described as heavy control layer, light generation layer, and it is worth unpacking because it is the opposite of the current fashion.

Instead of asking one large model to redraw every frame, Dawa put a control system between creative intent and pixels. That layer, called Latent Stage, is 34B parameters and models people, motion, camera, lighting, wardrobe, props and physical interaction as separate but jointly constrained variables. Change one thing and the rest of the world state is frozen.

The company reports 98.2% success on targeted continuous edits and 93.3% identity preservation. Those are vendor numbers, measured on their own tests, and should be treated as directional. Directionally, though, they describe the exact problem everyone in AI video complains about: you ask for one small change and the model rebuilds the world.

Underneath, the generation layer only computes the part that moved. Their sparse local micro-diffusion touches 10% to 20% of the frame, which they say cuts GPU power draw by 40% to 60% and makes inference three to five times faster. Stable walls, floors and faces get reused instead of re-diffused.

Then there is LiveStage, the real-time path: roughly 4B parameters active per step, 720p at 25 frames per second, running for over 120 minutes without drifting. On one graphics card. That last number is the one that matters, because drift is what kills every long generative session — the scene slowly mutates until it is a different room.

Why the offline business is the interesting part

Real-time demos are fun and hard to monetise. Dawa's revenue, according to the company, comes from something much less glamorous: short dramas.

China's vertical short-drama market is enormous and brutally cost-sensitive. Hongguo, ByteDance's free short-drama app, had roughly 403 million monthly users in July and August; Tencent's mini-program entry reached 143 million monthly users in August and 89.68 million weekly users in the first week of October, up 44.1% on the first week of September. When your audience is that large and your margins are that thin, the only number anyone cares about is cost per finished minute.

Dawa says several AI short dramas are already live on Hongguo, including titles that translate roughly as Evolving From the Tiger King, Possessed by the God of Cooking and My Life on a Deserted Island. The claim attached to them is specific and checkable in principle: per-shot retry counts fell from fifty or sixty attempts to under five, and unit production cost dropped by an order of magnitude compared with mainstream models.

If even half of that holds, it explains the funding. Five hundred million renminbi, roughly $70 million, across a Pre-A and an A round this year, with Qifu Capital, Kongkong Donghai, Yicun Capital and others participating. The company also says it confirmed close to RMB 500 million in revenue this year and is profitable at scale, which in this sector is a genuinely unusual sentence.

The 180 terabytes nobody else has

The moat they describe is not the model. It is the dataset underneath it.

In Yongchuan, Chongqing, Dawa runs an 84,000 square metre digital production facility with 11 stages that it says run above 92% occupancy. Over five years, more than 300 commercial projects have gone through it end to end, producing roughly 180 terabytes of structured process data: how the camera moved, where the actors stood, how the lights were placed, how the props were handled. Over 300 annotation dimensions per scene.

This is a different kind of data from the internet's video corpus. A YouTube clip teaches a model what a scene looks like. A shot log teaches it why the scene looks like that. For a control model, the second one is the entire ballgame, because control is the mapping from intent to result, and that mapping only exists in the process record.

It is also the least replicable asset in the company. You cannot scrape it, and it only accumulates if you are already running productions. That flywheel — shoots produce data, data improves the control model, the model wins more shoots — is the actual pitch, and it is a better one than a benchmark table.

The comparison everyone will make

Higgsfield, the San Francisco company with a similar name in the discourse, took a different route. According to the reporting around Dawa's round, Higgsfield burned $10 million across three attempts, then heard the same answer from eight senior creative directors: the missing piece is camera and shot control, not generation quality. It built that as an application-layer feature, went from $1 million to $100 million in annualised revenue in eighteen months, and is reportedly valued at $5.4 billion after its B round. Its founder's public position is that vertical companies should not chase general foundation models; the moat is model routing, with about 40% of calls assigned automatically and gross margins above 80%.

Dawa's counter-argument is that routing decides which model generates, while a control layer decides whether the output can be changed and kept consistent. One optimises cost. The other decides whether you can deliver. That is a fair distinction, and it also explains why they built a film set instead of a routing table.

Whether it is the better business is unresolved. Routing scales with demand and needs no studio. A control model needs 84,000 square metres and a decade of shoots before it is defensible.

What I would want to see before believing it

Three things.

First, independent measurement of the retry counts. "Fifty to five" is the kind of number that gets rounded in a pitch, and the difference between five and fifteen is the difference between a profitable production and a hobby.

Second, a live, unedited session of the 120-minute no-drift claim. Drift is easy to hide in a highlight reel and impossible to hide in an hour of continuous streaming.

Third, a shipped film. Dawa's credits include work on the Cannes-winning Wild Times and the theatrical release The Vanished, and it says it is planning the industry's first fully AI theatrical feature. A feature is a very public way to be wrong, which is precisely why it would settle the argument.

The thing I find most convincing is not the demo at all. It is that the company went and bought a film studio, hired crews, and spent five years recording how shots get made. That is a slow, unglamorous bet on the idea that the bottleneck in AI video was never pixels. It was knowing what you wanted, and being able to keep it.