What the industry data actually says

fal's data shows that enterprise production deployments use a median of 14 different models per company. This is a striking contrast to the LLM market, where a16z's own growth analysis shows that three providers (OpenAI, Google, and Anthropic) command 89 percent of enterprise wallet share. In generative media there is nothing like that concentration. The reason is that every model is strong in specific areas and weak in others, a photorealism leader is not usually the best option for anime aesthetics, a strong text-to-image model is not usually the strongest image editor, a model that excels at hero assets is rarely the right choice for high-volume thumbnail generation. The operator implication is that your competitive position is almost never about picking the right single model. It is about the workflow you build around a rotating set of models.

Producing a single polished asset is almost never a single inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. call. In practice, developers chain multiple models together: generate an image with one model, remove the background with another, upscale with a third, recolor with a fourth, then apply a style-consistent LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. for brand alignment. fal's framing is that 'the unit of work isn't one model, it's a workflow.' This has direct implications for your infrastructure decisions. Serving individual models quickly is table-stakes. The thing that actually matters is orchestrating multi-step pipelines with low cumulative latencyLatencyThe delay between asking for a generation and getting the first result. Distinct from throughput., managing dependencies between steps, and making it easy to swap in new models as the frontier moves every few weeks. If you have built your product around a single-model-per-request architecture, you are already behind the infrastructure curve.

Operators have gotten savvy about matching model choice to generation context. The rule that has emerged, per fal's analysis, is that the right model depends on what you are generating and at what scale. For high-volume utilitarian images (product thumbnails, feed assets, batch variations), you bias toward fast cheap models like FLUX because the marginal value of perfection is low and the marginal cost compounds fast across millions of generations. For hero assets where polish is the priority (ad campaigns, logos, brand imagery), you pay for something like Nano Banana Pro because small imperfections look unprofessional at that level of scrutiny. The interesting consequence is that infrastructure cost matters more than model cost at the high-volume end, while model quality matters more than infrastructure cost at the hero-asset end. Most operators need both at once.

fal's joint survey with Artificial Analysis found that 58 percent of organizations identify cost optimization as their primary criterion when selecting model infrastructure, ahead of model availability and generation speed. That 58 percent is the single most operator-revealing number in the entire report. It tells you that despite all the marketing focus on quality, capability, and features, when operators sit down to actually pick where to run their inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training., the first question they ask is 'how much does this cost.' Competition is happening at two layers simultaneously: between infrastructure providers racing to offer the most cost-effective run of a given model, and between models along the cost-quality frontier where the right choice depends on scale and tolerance for imperfection.

Generative imagery adoption is showing up across every industry, but three verticals are visibly ahead: gaming, advertising, and e-commerce. In gaming, studios are using generative models for concept art, environment population, and in-game asset generation at a pace traditional art pipelines cannot match. In advertising, campaigns that once took weeks of production now spin up hundreds of personalized variations in hours, which has changed the economics of creative testing and spawned entire new startup categories. In e-commerce, the case almost makes itself, when you need product shots, lifestyle imagery, and seasonal creative across thousands of SKUs, generative media turns what used to require a team of photographers, weeks of shoots, and long editing cycles into a few prompts and a library of production-ready assets.

The operator implication is that if you are building in one of these three verticals you are in the front of the adoption curve and should expect intense competition and fast-moving buyer expectations. If you are building in a vertical that is not one of these three (healthcare visualization, scientific imaging, educational content, real estate, architecture), you have a slower-moving market but also less benchmarked competition, and your biggest challenge is probably buyer education rather than product differentiation.

The closed-versus-open debate in generative media is not playing out the way it did for language models. For LLMs the open-source case has mostly been about cost and sovereignty. For generative imagery the open-source case is about customizability. Enterprises that need brand consistency, character persistence, or product fidelity across millions of generated assets cannot get that from closed APIs, because fine-tuningFine-tuningContinuing to train an existing model on new data to specialize it for a style, subject, or domain, without training from scratch. on your own proprietary data is either not offered or offered in severely constrained ways. For these customers, fine-tuning is not an optimization, it is the entire game. An open-weight model like FLUX.2 [dev] or Qwen-Image-Edit lets you train a custom LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. on your catalog, your brand assets, your specific characters, and run inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. against that customized model as many times as you want. A closed model like Midjourney or GPT-Image does not. This is why open-source has been closing the quality gap much faster than anyone expected, the labs shipping 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. are getting compounding benefit from enterprises fine-tuning those weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable. for their own use cases.

These five findings are not the whole operator playbook, they are the factual ground under it. The rest of this section, the stack, the build-vs-buy decisions, 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., the moats, the decision framework, is built on top of these findings. If any of the advice in the sections below feels wrong to you, go back to the five findings and check your instinct against the data.