Open an advertising agency’s rate card this year and a new line has appeared: AI video production. Priced by the second, three rounds of revisions included, extra rounds billed separately, with the supplier responsible for clearing the footage rights. Two years ago that line did not exist, and procurement had no idea whom to call.
A quote sheet means the work has moved from “demo” to “delivery”. Someone is willing to sign, someone is willing to pay, and someone has started asking about the delivery date. The excitement belongs to others. The contract is yours, and so is the date.
The Era of Comparing Parameters Is Over
For two years, AI video launches looked like an arms race: who had the longer clip, the higher resolution, the steadier motion, the more polished demo reel. Whoever had those was remembered, and reposted.
This year the wind shifted. The conversation moved from “what the model can do” to “how much money this step can save”. Parameters still matter, but they are no longer scarce. More and more people can coax a good image out of a model.
The Paper offered a judgment in its reporting: the sector is leaving its show-off phase, and competition is moving from model parameters to scene penetration and a closed commercial loop. In plain language, it is a contest to push the tool into someone else’s workflow.
Factories and Merchants Paid First
Those who paid early are not necessarily film companies. On 31 July, ByteDance released Seedance 2.5, lifting a single generation from fifteen seconds to thirty, with up to fifty multimodal references, enough to describe a more complex shot in one go.
Where it went matters more: XCMG uses it to generate industrial training videos, XPeng wired it into its product design platform, and Weifen Zhifei uses it to generate obstacle-avoidance and pathfinding training data for robots.
What these customers share is that the buyer does not care whether the picture is stunning. The buyer cares whether it can replace part of a shoot, part of an outsourcing contract, part of repetitive labour. What is saved is hours, not applause.
The maths on the e-commerce side is even more direct. A garment needs a model, a set change and a dozen versions. That used to mean a studio; now a batch of assets can be generated first, one picked for retouching, and the cost falls from a day of studio rent to a few minutes of electricity.
Work like this is high-volume, short-cycle and clearly specified. It does not chase cinematic feel. It needs the product colour to be accurate, the model’s fingers to be normal and the background not to give itself away — exactly the range where models currently do well.
For the merchant there is one test: will the sales this video brings in cover the production cost? If yes, order more. If not, change tools. There is no sentiment in the middle, no brand baggage either, only a return-on-spend spreadsheet.
For the client, what is being bought is not generation capability but certainty of production: order today, deliver tomorrow, and know whom to call when something breaks the day after. That is harder to achieve than any demo video.
A Seven-Person Crew Took On a Feature
Another direction is the small crew. The Korean director Yang Ik-june is making a theatrical science-fiction feature, “RAPHAEL”. The traditional route needs a crew of 150 to 300 and a budget of about two million dollars, counted in years from preparation to wrap.
With AI, the team currently numbers seven. What is saved is not the creative fee but the sets, the extras and the previsualisation — the parts that can be redone again and again, and the parts that most often blow the budget. Many roles have been folded onto one desk.
The point of such cases is not that “AI can shoot a film”. It is that the process is being written into budgets, schedules and delivery lists. Once it is in the spreadsheet, it stops being an experiment and becomes a cost line that can be discussed, and bargained down.
It Showed Up in TV Dramas First
No need to wait for science fiction. In this year’s New Year drama “The Peaceful Years” and Amazon’s “House of David”, some virtual sets and effects shots were handled by Kling, and the production teams also used image models to shortlist concept art and character designs.
That step comes earlier than the finished film. After the script is locked and before the shoot begins, a director needs hundreds of concept images to convince investors and platforms. That used to mean an artist drawing them one by one. Now a batch can be run first and reviewed.
The tool’s role here is modest: it lowers the cost of communication. It does not replace the art department. It just means fewer dozens of discarded drafts, fewer cross-department meetings, and a decision made two days earlier.
Professional Users Pay by the Month
A consumer’s first-time payment usually happens once. The steady money comes from people who treat the tool as production equipment: self-media creators, advertising hands, teams racing a short-drama deadline. They worry about delivery every single day.
Their logic is simple. If the tool means hiring one less person or pulling one less all-nighter, the monthly fee is worth it. They are not chasing a single stunning result. They need to deliver today, deliver tomorrow, and get paid on time the day after.
Kling has put its weight behind this group. Members renew monthly, not for novelty but for a production line that never calls in sick. Once a tool is in the schedule, replacing it means rescheduling the whole job.
The Money Starts Pooling in One Place
On the evening of 30 October, Kuaishou published a notice to the Hong Kong stock exchange: Kling’s capital increase had closed, with investors paying a total of about 19.8 billion yuan, and shares from investors who missed the deadline redistributed.
Set against the whole sector, the number shows capital has made a choice. It is no longer betting evenly. It is putting money behind teams that already have paying users, a workflow and a delivery record.
On the same day, Kuaishou’s own Hong Kong valuation was not high. A subsidiary that is not yet listed was valued above its parent, and that inversion is itself a statement: the market is not buying the present, but the future it believes in.
Being Able to Generate Is Not Being Able to Deliver
After payment, the real problems begin. The client wants a fixed delivery date, not the luck of a good roll. Generate a hundred, pick three, spend two days fixing them, and expect more than three rounds of revisions. That is the daily routine.
For a thirty-second film, compute, rework, rights clearance and music licensing each go into the quote. The longer the quote, the clearer it is that this business no longer makes money from surprise. It makes money by pushing uncertainty down.
So the barrier is shifting from “can you use a model” to “can you run a production line”. Many people can write a prompt. Fewer can tie prompts, assets, schedules and licences into one line. Fewer still can deliver on budget and on time.
Between being able to generate and someone being willing to pay, what sits in the middle is not technology. It is process. Quote sheets, delivery dates, revision counts, invoice headers — these unglamorous things are where an industry actually begins, and the part hardest to copy.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.