The generative imagery stack and where you sit

At the bottom is the physical Graphics Processing Unit (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.). Almost all production generation in 2026 happens on NVIDIA H100s, with H200s and B200s starting to appear at the high end and consumer 4090/5090 cards still being used for hobbyist work. NVIDIA captures roughly $25,000 to $40,000 of margin on each H100 sold. The hardware layer is dominated by NVIDIA so completely that it is almost a tax on the rest of the field. AMD's MI300X is technically competitive but the software ecosystem (CUDA) is not, so almost nobody uses it for diffusion modelDiffusion modelA generative model that learns to produce images by training a neural network to reverse a noise-adding process. Generates by starting from random noise and iteratively denoising to produce a clean image. inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. in production.

The next layer is the providers who buy GPUs from NVIDIA and rent them out by the hour or by the second. The three categories are: hyperscalers (AWS, GCP, Azure) who rent GPUs as part of their broader cloud offering; specialized 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. clouds (CoreWeave, Lambda Labs, RunPod, Together AI) who focus exclusively on GPUs and tend to be cheaper; and emerging players (Crusoe, FluidStack, Vast.ai) who often buy under-utilized GPUs from other sources and resell at a discount. Hourly rates for an H100 in early 2026 range from about $2.50 (cheapest spot pricing) to $8 (on-demand from a hyperscalerHyperscalerOne of the largest general-purpose cloud providers, meaning AWS, Google Cloud, and Azure, which rent GPUs as part of a broader offering.). The difference compounds enormously at scale.

Above the 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. is the 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. layer, companies that take the raw 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. rental and turn it into 'pay per generation' or 'pay per second of model time' APIs. These include fal.ai, Replicate, Modal, Baseten, Beam, and Cerebrium. They host the model weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable., manage the inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. stack, handle autoscaling, batch generation requests, manage cold starts, and expose a clean Application Programming Interface (APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface.) to developers. The value they add is operational, most teams cannot or do not want to manage GPU clusters and inference infrastructure themselves, so they pay a markup to have someone else do it. fal.ai has emerged as the largest player specifically for media generation; Replicate is the broader 'any open-source model' platform; Modal is more developer-tooling-focused.

The more interesting thing happening at this layer in 2026 is that the inference platforms are no longer just serving models, they are building workflow infrastructure that sits between the models and the operators using them. This is where most of the 2025-2026 platform innovation has actually happened, and it is orthogonal to what the model labs are shipping. fal.ai has invested heavily in workflow primitives, their API exposes not just 'run this model' but 'run this chain of models with these parameters and return the final output,' which means developers can build multi-step pipelines as a single API call rather than orchestrating the chain themselves. Replicate has invested in model discoverability and community sharing, making it possible to find and run any of thousands of community-shared models without managing weights directly. Modal has invested in generalized serverlessServerlessRunning a model on infrastructure that spins up on demand and bills per use, so you do not manage an always-on GPU. infrastructure that treats model inference as one of many possible workloads, with the advantage that you can build arbitrary pipelines that mix model inference with other compute without switching platforms.

The specific innovations worth naming, because they are not about models and they are reshaping what operators can build: day-zero model supportDay-zero model supportAn inference platform hosting a newly released model within hours of its launch, so operators can evaluate it immediately. (a new model drops from Black Forest Labs or Alibaba and appears on fal within hours rather than weeks, which means operators can evaluate it immediately instead of waiting for tooling to catch up), unified API surfaces (a single API shape across dozens of models so swapping one model for another is a configuration change rather than a rewrite), queue management and streaming (for long-running video generation jobs where waiting 3 to 8 minutes for a Veo clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. is the norm and you need to return intermediate status to users), managed fine-tuningFine-tuningContinuing to train an existing model on new data to specialize it for a style, subject, or domain, without training from scratch. (upload your dataset, 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. against a supported base model, get an inference endpoint back, without provisioning a training GPU yourself), multi-model chaining (express a workflow as 'generate image with FLUX, remove background with RMBG-2, upscale with SwinIR, apply LoRA X' as a single callable pipeline), and cost-observable routing (automatically route generations to the cheapest inference endpoint that meets quality requirements, across multiple GPU clouds and multiple model variants). None of these are things the model labs themselves provide, they are the operational layer that makes frontier models usable in production, and they are what operators are actually paying for when they use an inference platform instead of running models themselves.

Above the inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. platforms are the model labs themselves, OpenAI, Google, Anthropic, Black Forest Labs, Stability, Alibaba, Kuaishou, ByteDance, MiniMax, Runway, Pika, Luma, Midjourney, Ideogram, Recraft, and so on. Some of them sell direct APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. access to their own models (OpenAI, Google, Anthropic, Midjourney through their API). Some sell only through inference partners (Black Forest Labs primarily through fal, Replicate, and Together AI, plus their own BFL Playground). Some sell only through their own consumer product (Midjourney, Ideogram). The labs capture the value associated with frontier model quality.

Above the labs are the application platforms, Krea, Flora, OpenArt, CivitaiCivitaiA community website for sharing Stable Diffusion and FLUX models, LoRAs, and embeddings., Leonardo, Playground, Magnific, and many others. They wrap the underlying models in user-facing tools, add workflow features (multi-model pipelines, ControlNetControlNetA conditioning technique that lets you provide structural guidance (depth, edges, pose) to a diffusion model to control spatial layout. integration, Low-RankRank (LoRA rank)The inner dimension of a LoRA's two small matrices, which sets how much capacity it has to learn. Higher rank captures more complex behavior but risks overfitting and produces larger files. Adaptation (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.) management, project organization, collaboration), and sell to end users. They are the layer most users actually interact with. They capture the value associated with usability, workflow, and brand. Their margins are determined by the markup they can charge above the underlying inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. cost.

The 2025-2026 wave of innovation at this layer has been about treating generative imagery as a creative medium that needs creative tools, not as an APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. that needs a web front-end. Krea has been the leader on real-time interaction, pioneering the pattern where images update continuously as you type or sketch, which changed what creative iteration looks like in this space. Krea also built a canvas-based interface that lets users combine multiple models, multiple reference images, multiple LoRAs, and multiple adjustment layers in a single visual workspace, which maps onto how working artists actually think about composition. Flora took a different approach, organizing generation around nodes and connections in a way that reads like a lightweight version of ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. aimed at creative professionals rather than technical power users. Magnific specialized in upscalingUpscalingIncreasing an image or video's resolution after generation, often with a model that invents plausible fine detail rather than simply stretching pixels. and refinement, building the best-in-class tool for taking a rough AI-generated image and making it production-quality through multi-pass refinement with detailed control over what gets enhanced.

Civitai plays a different role, it is primarily a distribution and discovery platform for community-trained LoRAs and fine-tunes rather than a generation tool, and its contribution to the ecosystem is making it possible for independent creators to share and find specialized models without running their own infrastructure. Leonardo and Playground built consumer-facing generation tools with stronger branding than the raw API platforms, targeting casual and prosumer creators rather than professionals. OpenArt positioned itself as a workflow-and-sharing hub with unusually deep ComfyUI integration. Each of these platforms makes a bet about what the right level of abstraction is for a specific user segment, and the collective effect is that creators can find a platform that matches how they want to work rather than having to adapt to whatever the model labs happen to ship.

The specific innovations at the application layer that are not just 'a nicer interface on top of the same models' include: collaborative projects (multiple users working on the same creative canvas with shared history and comments, which is table stakes in traditional creative software but was missing from generative tools until Krea and Flora added it), version history with rollback (treating each generation as a commit in a version graph you can branch and merge, which enables exploratory iteration without losing promising earlier directions), LoRA training as a first-class feature (upload 20 reference images, wait a few hours, get a custom model that you can use in any generation, without touching the underlying training infrastructure), multi-model routingModel routingAutomatically sending each request to the best model for it, by quality, speed, or cost, instead of committing to one model for everything. in a single promptPromptThe text description you provide to a model to specify what you want it to generate. (the same prompt runs against FLUX.2, Nano Banana Pro, and Ideogram V3 simultaneously so you can compare outputs without switching tools), real-time video canvas (see video updates as you iterate on the prompt, which Krea pioneered for images and is now extending to video), and brand kit integration (upload your logos, color palette, and typography standards once and have every subsequent generation respect them automatically). These are the things that distinguish application platforms from raw inference access, and they are the reason operators building a creative tool will often choose to build on top of fal or Replicate rather than on top of Krea, you pick the level of abstraction that matches what you want to expose to your users.

At the top are the end users, individual creators, agencies, in-house creative teams, hobbyists, prosumers. They are the source of all the money in the stack. Everything below them is competing for a share of what they will pay.

If you are building a generative imagery company, the first strategic question is: which layer am I building at? Because each layer has different economics, different competitive dynamics, and different paths to a defensible business. The answer determines almost everything else.