AI Film Compute Is Collapsing, But Quality Is What Actually Ends Hollywood's Old Model
Higgsfield's The Cully Hill Boys premiered in New York on August 5, 2026, a 110-minute AI-generated action-comedy made in four weeks for about $2 million, with roughly $1 million of that going to compute. That $1 million is a real number for what the production cost at the time it was made. It is also a number that has already come down since, and will keep falling on a curve steep enough to change what a movie is worth making for.
The generation rates that produced this film in mid-2026 are not the rates a producer will pay next quarter. Video generation pricing is dropping every month, control tools are cutting the number of attempts each shot takes, and open-weight models are moving the work off metered cloud APIs and onto owned hardware. Run those trends forward and the same film costs a fraction of $1 million within a year. Run them a little further, and the economics that justified $100 million studio budgets stop making sense.
Why the Million Is Already Smaller
The $1 million compute figure is best understood as a snapshot of frontier pricing on an unstandardized workflow. Most of that money went to iteration, re-runs, upscaling, and experiments the team ran before the pipeline settled. Those are the parts that shrink fastest as the tools mature and the workflow becomes repeatable.
The raw generation layer is the cleanest place to watch the decline, because it is the easiest to price. A 110-minute film is 6,600 finished seconds. At a typical pace of three seconds per shot, that is roughly 2,200 shots, and real production burns three to ten attempts to land each usable shot. That puts total generated video somewhere between 40,000 and 66,000 seconds. Seedance 2.5, the model used for the film, runs near $0.23 per second at 720p as of August 2026. Multiply it out and the raw generation of the entire film sits around $9,000 to $15,000 at today's rates.
That gap between the generation layer and the full compute bill is the overhead that standardization removes. As the workflow stops being experimental, the compute number moves toward the generation layer, and the generation layer itself keeps dropping.
The Rate of Decline
Three separate cost declines are compounding, and none of them depends on a single vendor's roadmap.
Per-second generation price is falling first. Models that cost $0.20 to $0.50 per second in early 2025 now have competitors at $0.02 to $0.09 per second. As of August 2026, premium closed models like Sora 2 Pro and Veo 3.1 sit at $0.30 to $0.70 per second at 1080p, mainstream models including Seedance 2.5 run $0.12 to $0.25, and the cheapest production-grade options drop to $0.02 to $0.09. Epoch AI and Stanford's AI Index put LLM-class inference decline near 40x per year at mid-range capability. Video is younger and falling faster off a higher base. A conservative planning assumption is a 2.5x to 5x drop per year at constant quality, which works out to somewhere between 8% and 13% per month.
The second decline is iteration rate, and it moves the number more than price does. The dominant cost is not what a second of video costs but how many attempts it takes to keep one. Control tools that regenerate a single section of a shot, hold a character's face consistent across cuts, and lock reference images push the three-to-ten attempt range toward one-to-two. Cutting attempts in half cuts cost in half on top of any price drop.
The third is the shift to local hardware. LTX-2.5 is a 22-billion-parameter open-weight model with synchronized audio that runs on a Mac with enough memory, at one-fifth to one-tenth the cost of the cloud API. As open models stay within a generation of the frontier, more of a production moves from metered API calls to fixed-cost owned hardware where the marginal cost of another take is close to zero.
Where the Number Lands
Applying those curves to the same 110-minute film at the generation layer gives a clear trajectory.
Today the quality-matched price is near $0.15 per second, shots take about six attempts each, the production generates roughly 40,000 seconds, and about 10% of that runs locally. Generation compute lands at $9,000 to $15,000.
Six months out, price falls toward $0.08, attempts drop toward four, generated seconds fall as fewer takes are wasted, and the local share climbs past 25%. Generation compute moves toward $3,000 to $6,000.
At one year, price is near $0.05, attempts are down to three, and more than a third of the work runs on owned hardware. Generation compute reaches roughly $2,000 to $4,000.
By 18 months, price is around $0.03, shots take one to two attempts, and the majority of generation runs locally. Generation compute lands at $500 to $1,500.
The full production cost follows the same shape. Overhead, tooling, and team all compress as pipelines standardize, and all-in cost excluding licensing moves from the $1 million to $2 million range today toward $50,000 to $150,000 within 18 months. The one line that may rise is licensing, which is contractual rather than compute-driven and could climb as named talent recognizes demand.
Even If the Cost Froze Today
Here is the part that matters more than the exact numbers. Suppose the price stopped falling right now and stayed at 2026 rates permanently. The old model would still be finished.
Quality is climbing at the same time cost is dropping. Each model generation renders more convincing faces, cleaner physics, and longer coherent shots. The direction is toward output an ordinary viewer cannot distinguish from a conventionally shot film. Once the content matches and no one watching can tell the difference, the spending logic of the traditional business breaks.
A studio film built the old way carries the cost of physical locations, crews numbering in the hundreds, insurance, travel, catering, set construction, and months of scheduling around the availability of people and places. The AI production carries none of that. If two finished films are indistinguishable on screen and one cost $150,000 while the other cost $80 million, the $80 million number has to justify itself on something other than what ends up on the screen. For most commercial content, it cannot. A studio weighing a green light will not keep paying a hundredfold premium for an output the audience cannot separate from the cheaper one.
That does not erase the top of the market overnight. Prestige projects, live performance, and work that sells partly on the fact that real people were really there will hold value for a while. But the broad middle of film and streaming content, the volume business that fills catalogs and feeds, moves toward the cheaper method as soon as the quality lines cross. The huge-budget, multi-location model does not vanish in a single year. It fades, one project at a time, as each new film proves the gap has closed.
The New Stars Are Already Here
The cast of The Cully Hill Boys points at who benefits. The film starred the licensed likenesses of UFC champion Israel Adesanya, MMA veteran Quinton "Rampage" Jackson, streamer N3on, and streetball creator Matt Kiatipis, all contracted for likeness and voice before production. None of them had to stand on a set. Their faces and voices were licensed, generated, and then, per Higgsfield, deleted within 30 days of wrap.
That arrangement rewires who a movie star can be. A traditional actor's time is the bottleneck, since a person can only be on one location shooting one scene at a time, and that scarcity is what caps how many projects they can headline. A licensed likeness has no such limit. A streamer with a large audience can license their appearance into as many films as they can negotiate, running in parallel, with no location, no travel, and no scheduling conflict. The people who already command attention on YouTube, Twitch, and social platforms bring the one asset AI production still cannot manufacture, which is a built-in audience that trusts a specific face.
This is a different pipeline into stardom than the one Hollywood controlled. The gatekeeping that determined who got to be on screen loosens when being on screen no longer requires being on set. Creators who built their following outside the studio system can now appear in scripted features without leaving their setup, and they can do it across many projects at once. For a generation of internet-native performers with millions of followers each, the likeness becomes a licensable asset that works while they are doing something else.
What to Watch
For anyone modeling this, the practical signals are concrete. Watch iteration rate more closely than sticker price, since the tooling that gets a usable shot in one or two tries captures most of the remaining cost reduction. Watch the quality line, because the moment ordinary viewers stop being able to tell AI-generated film from conventionally shot film is the moment the budget argument ends. And watch the licensing market for likenesses, because that is where the value that used to sit in physical production and A-list scheduling is moving. The rates cited here are snapshots that move quickly, so verify against current vendor pricing before committing a budget.
David Borish is a journalist and analyst covering frontier AI, enterprise technology, and emerging science, and the author of the forthcoming book The Tony Hawk Paradox. His work is at davidborish.com.
