Unit economics: where the money actually goes

To make the unit economicsUnit economicsThe per-unit costs and revenue of a product, for example the cost to generate one image against the price charged for it. tangible, here is the breakdown of where the money goes for one FLUX dev image generation served through a typical SaaS product.

Cost component

Approximate cost

Captured by

GPUGPU (Graphics Processing Unit)The parallel processor that runs image and video models. Its speed and memory largely determine what you can generate and how fast. compute (raw H100 time)

$0.0015

GPU cloudGPU cloudA provider that buys GPUs and rents them by the hour or second, sitting between the hardware maker and the inference platforms. Examples include CoreWeave, Lambda, and RunPod. (CoreWeave, Lambda, etc.)

Inference platformInference platformA company that turns raw rented GPUs into pay-per-generation APIs, hosting model weights and handling scaling so developers do not manage the hardware. Examples include fal, Replicate, and Modal. markup

$0.005

fal, Replicate, etc.

Model labModel labA company that trains and ships the frontier generation models themselves, such as Black Forest Labs, Runway, or Google. license fee (FLUX dev commercial)

Variable

Black Forest Labs

Application platformApplication platformA product that wraps underlying models in user-facing creative tools and workflow, capturing the value of usability and brand. Examples include Krea, Leonardo, and OpenArt. infra

$0.001

Your servers and CDN

Application platform team & overhead

$0.005

Your salaries amortized

Customer-facing price

$0.01 - $0.10

What you charge

Notice that the actual GPU compute is the smallest line item. Most of the money goes to operational overhead at the application layer and to the inference platform that abstracts away the GPU management. This is why inference platform pricing matters so much, the platform markup is comparable in size to your entire labor cost per generation, and reducing it directly improves margin. It is also why running your own inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. becomes attractive at scale.

Audio forces a distinction image and video never had to make, the line between speech and song, and it shows up in how each is priced. Text-to-speechText-to-speech (TTS)Generating spoken audio from written text. Modern neural TTS learns the mapping from text to sound directly from data rather than gluing together recorded snippets. is metered against input, per character or via credits that map to characters, so speech cost scales with the volume of text you push through it. Music is priced per song or, more often, as a monthly subscription with a generation cap, so it sells a fixed ceiling of outputs rather than metering usage. As of July 2026, ElevenLabs runs on a credit-per-character model across paid tiers from a few dollars a month up to enterprise, Cartesia uses the same credit logic tuned for real-time voice, and Suno sells monthly plans around eight and twenty-four dollars with a daily free allowance, with Udio offering comparable tiers. One trap for operators: APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. access and the consumer interface are usually priced and licensed separately, and music platforms increasingly gate commercial download behind paid accounts, so the marketing sticker rarely equals your marginal cost inside a product.

The unit economicsUnit economicsThe per-unit costs and revenue of a product, for example the cost to generate one image against the price charged for it. then split by archetype. A live voice agent is dominated by minutes of speech generated in real time, so latencyLatencyThe delay between asking for a generation and getting the first result. Distinct from throughput.-optimized synthesis and concurrency set both cost and quality, and a caller who talks over the bot still bills you for the audio already synthesized. A dubbing or localization service pays per character per language, so margin is eaten by revisions and retakes, and the winning move is to charge per finished minute while your cost is per character. An audiobook or podcast pipeline is batch rather than real time, so it can use cheaper models and off-peak compute, and its economics are a one-time render amortized over unlimited plays. A music-bed feature lives or dies on the subscription cap, because if your users each generate more than the plan allows you either eat the overage or ration generations.

Video is where generative cost stops being a rounding error. The unit that matters is not the generation, it is the finished shot, and a finished shot is many generations. A working rule from productions actually doing this is about ten generations to land one usable two to three second shot. So a two minute film, cut from roughly twenty minutes of raw material, runs to about 1,200 seconds of raw generation before a single frame is kept. That number, seconds of raw generation, is the one that sets your budget.

Multiply it by a model's price per second and the spread across models is enormous. Pushing 1,200 seconds of 720p generation through an aggregatorAggregatorA tool or platform that wraps many image and video models behind one interface, so you switch models without switching apps. like Fal.ai, a late 2026 breakdown from the AI film school Curious Refuge puts the ceiling prices at roughly: Seedance 2.0 at $364, LTX 2.5 Pro at $204, Kling 3.0 at $202, Minimax H3 at $96, and Luma Ray 3.2 at $72. Same film, same length, a fivefold swing on model choice alone. And the floor is not $72, it is zero: LTX 2.5 run locally overnight in ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. costs only electricity.

The second lever is resolutionResolutionThe pixel dimensions of an image or video, for example 1024x1024. Higher resolution shows more detail but costs more memory and time., and it is the one most people get wrong. At Seedance 2.0's 4K rate of about $1.56 per second, twenty minutes of native 4K footage for that same two minute cut runs about $1,872. Generate the identical film at 720p and upscale it afterward, where a personal Topaz license runs about $39 a month, and the all in total lands around $139 or less, for output most viewers cannot distinguish from native 4K. That is better than a tenfold saving, and the upscale, not the native render, is the sane default.

Once you see that raw generation dominates the cost, the tactics for controlling it become obvious. A handful of habits separate a film that costs $139 from the same film that costs $1,800.

Generate at 720p, upscale last. Native 4K is the single most expensive habit in the pipeline, and it buys almost nothing a good upscalerUpscalerA model or tool that raises the resolution of an image or video, either faithfully recovering detail or inventing plausible new detail. cannot recover. Work at 720p through the whole exploration and finishing process, and upscale only the final cut. This one change is most of the tenfold saving.

Explore free, finalize paid. The hundreds of throwaway generations you burn finding a shot do not need a premium cloud model. Run a local open weightsOpen weightsModels whose trained parameters are publicly downloadable, allowing anyone to run them locally or fine-tune them. Contrasts with closed/API-only models. model like LTX overnight in ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. for the search, and spend per second cloud money only on the takes you have already decided to keep.

Route by stakes, not by habit. A hero shot that carries the scene earns a premium model and twenty rolls. An establishing shot or a two second cutaway does not. Matching model cost to a shot's importance, rather than running everything through your best model, routinely halves a budget.

Lock identity before you scale. The re-roll tax, generating a shot again and again because the character drifts, is where budgets quietly disappear. A character LoRACharacter LoRAA small model trained on images of one specific subject so that a fixed identity can be summoned consistently across an entire production. or a locked reference frame established up front turns ten rolls per shot into three.

Keyframe cheap, animate dear. A still from an image model costs cents; a second of video costs dollars. Generate and select your keyframes first, cheaply, and commit to image to video only on the frames that survive. The craft habits of bookending a shot and stacking references are, among other things, a form of cost control.

The right stack depends entirely on what you are making. A few common cases, from cheapest to most demanding:

Social clip or vertical short, fifteen to thirty seconds. The cheapest capable model, Luma Ray or Minimax H3, at 720p, with light upscalingUpscalingIncreasing an image or video's resolution after generation, often with a model that invents plausible fine detail rather than simply stretching pixels. or none. Tens of dollars, same day. Volume and turnaround matter more than fidelity, and the algorithm forgives a soft frame.

Two minute narrative short. 720p plus a final upscale, a mid tier model, roughly $100 to $150 in the cloud, or effectively free run locally overnight. Budget the iteration, not the render. This is the range where taste, not tooling, decides the outcome.

Pitch, previz, or mood film. Quantity over polish. Local or the cheapest cloud model, because you want fifty rough looks to find the idea, not one clean shot nobody has approved yet. Spend nothing here so you can spend on the version that gets greenlit.

Commercial or client hero spot. A premium model, Seedance or Kling, at high resolutionResolutionThe pixel dimensions of an image or video, for example 1024x1024. Higher resolution shows more detail but costs more memory and time. for the two or three hero shots that sell the piece, and a budget model for coverage. Keep the iteration budget high: a client is paying, in part, for the take you would otherwise have thrown away.

Music video or stylized piece. High shot count, forgiving of imperfection. A style LoRAStyle LoRAA LoRA trained on images that share an aesthetic but vary in subject, so it can render any prompt in that art style, photographic look, or brand identity. plus a budget model plus heavy selection beats a premium model used sparingly. The look carries the piece more than any single frame's fidelity, so buy shots, not resolution.