The model is not the product anymore

Everything before this chapter was the machine: how generative models work, and how to wield them. Here the book turns. The rest of it is about the medium being built on top of the models, the products, workflows, companies, and layers that turn a model into finished work and carry that work to an audience. This part explains why, in 2026, knowing the best model matters less than knowing the whole board.

Everything in this book up to here treated the model as the thing. For a while that was right. This part explains why it stopped being right, and what took its place.

Recall two facts the book has already established. First, the image models have largely converged: the quality gap between FLUX, Nano Banana, GPT-Image, Ideogram, and the 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. is real but narrowing, and the open weights keep closing it. Second, inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. is a race to the bottom, with aggregators and hosts competing on price per image and per second, and that price only moves one way. Put those together and you have a commodity. When a capability is excellent everywhere and cheap everywhere, it stops being where the value is.

This is an old pattern. Value migrates away from the layer that commoditizes. It moves up, toward whoever owns the workflow and the customer, and down, toward whoever owns the scarce input or the audience. In generative media in 2026 both moves are happening at once, and the model sits in the thinning middle.

The evidence is now concrete, not just intuition. By one widely-cited estimate the application layer captured more than half of all enterprise generative-AI spending in 2025, ahead of the foundation models themselves. And a single AI-video app, Higgsfield, went from being founded in 2023 to a reported five-hundred-million-dollar annualized revenue run rate about a year later, most of it from enterprises paying per seat and per render, while several of the model labs it competes with were retrenching. The money is telling you where the value moved.

None of this means the models do not matter. They are the engine. But you do not sell an engine to someone who wants to arrive somewhere, and increasingly the business, the craft, and the jobs are in building the car around it.