Photo by André Eusébio / Unsplash
Economies of Scale in the Age of Generative AI: How Falling Marginal Costs Are Reshaping Visual Production
Everyone who has taken introductory microeconomics has sketched this U-shaped average cost curve: as output expands, cost per unit decreases until the firm fully utilises its economies of scale, then begins to increase again as capacity becomes strained.
Generative AI models provide an extraordinarily crisp use case for this concept, and also subvert it in a fascinating way. There is no visible point at which the average cost starts to rise again as a generator like Seedream 5.0 Pro cranks out additional output. For all practical purposes, after the initial fixed investment, producing one more unit is essentially free; the standard U-shaped curve no longer holds.
The Fixed-Cost Mountain
Even before any model, such as Seedream 5.0 Pro, generates an image, its developers must overcome a fixed-cost barrier that far exceeds that of conventional manufacturing.

Comparable systems give some sense of scale: Google spent an estimated $192 million to train one model, Gemini 1.0 Ultra, with costs evenly divided between compute hardware and salaries for the research team that built it. Image-generation models are generally smaller, but still not inexpensive.
Training a diffusion model of an earlier generation (e.g., Stable Diffusion) has been estimated at somewhere around $600,000. Regardless, they are fixed costs in the textbook sense: the bill doesn't change whether the finished model generates ten images or ten million.
The Near-Zero Marginal Unit
At the end of training, this cost structure flips entirely. Generating a single additional image with something like Seedream 5.0 Pro costs almost nothing compared to that training bill — just seconds of server time and a tiny trickle of electricity.
This allows a brand to place an identical product on a marble counter, at a beach, or in a studio set for the cost of a few cents of compute, instead of scheduling an actual shoot. Platforms built on these models market this exchange precisely: replace a physical photography budget with a prompt and a reference image, and the environment changes at no cost.
This is marginal cost collapsing towards zero, and displacing a cost that used to be real, wholly variable, labour-intensive, billed per shot for the studio rental, a photographer's time, a stylist's fee, and often a model's fee.
Production function and factor substitution
This is a classic example of capital-for-labour substitution in a firm's production function. A traditional product shoot is a fixed ratio of labour — photographers, stylists, retouchers — to physical capital, such as studios and lighting gear, used to create each image.
This means that most of that labour is replaced by a single investment in a different form of capital: trained model weights and the GPU infrastructure needed to run them for each AI-generated shoot. Capital, once created, can be reused in perpetuity with nearly zero re-creation cost. So as economies of scale go, these aren't the marginal efficiencies of operating a larger factory line — they're one huge sunk cost being spread across an exponentially increasing volume of production.
Why this is happening now, and what it means
A couple of forces are propelling this transition. GPU compute is getting cheaper by the hour, and competition in AI image-generation has reached a fever pitch, with firms racing to capture commercial design budgets. Training costs, however obscene — running into the tens or hundreds of millions — are one-time lumps, amortisable across a global customer base rather than charged directly to a single recurring client.
One important caveat is that inference isn't really free at scale: those servers processing millions of image requests are still a cost centre. The "nearly free" price tag for a single image doesn't account for the monthly bill required to keep that infrastructure operational. The short-run effect resembles simple job displacement among those in commercial and product photography, such as photographers, models, and set designers involved in the more routine parts of the work.
This longer-run effect is familiar from previous waves of falling marginal costs: markets end up producing orders of magnitude more output per dollar than before. Labour, to the extent that it survives the ubiquity of models, is pushed towards the parts of a job — creative direction, brand judgement, taste — that a model still cannot consistently do by itself.