Generative AI: The 2026 Field Guide

We have talked about specific papers and specific labs, but it is worth zooming out and explaining how the entire research-to-product pipeline functions, because the structure of that pipeline shapes everything about which models exist, who builds them, and how fast new ideas propagate. The field looks chaotic from the outside, papers appearing on arXiv every day, models launching every week, benchmarks shifting constantly. From the inside it follows a consistent five-stage pattern, and once you see the pattern it stops feeling like chaos.

Fig.diagram

■The five stages from idea to user. Each stage filters who can participate.

Stage 1: research

New ideas in machine learning almost always start in academic research labs, usually with a graduate student exploring a question their advisor finds interesting. The work is small-scale by frontier standards. Compute is borrowed, from the university cluster, from a generous industry partner, 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.'s academic credits program. Datasets are small. Experiments are toy-sized. The goal is not to build a product but to find out whether an idea is interesting enough to scale up.

This is where the foundational insights come from. Sohl-Dickstein's diffusion paper in 2015 was research-stage work. So was Yang Song's 2019 score-matching paper. So was Peebles and Xie's Diffusion TransformerTransformerThe neural network architecture built on attention that underpins modern language and, increasingly, image and video models. It scales well and handles long-range relationships better than convolutional networks. (DiTDiT (Diffusion Transformer)A diffusion model architecture that uses a Transformer instead of a U-Net. Introduced by Peebles and Xie in 2022. Scales much better than U-Nets and is the basis for almost all current frontier models.) paper in 2022. None of them produced a usable model. All of them produced an idea that, applied at scale by somebody else, would change the field. The output of stage one is an arXiv preprint and usually a small GitHub repository with reference code that works on a toy dataset. Total cost: tens to hundreds of thousands of dollars, mostly compute. Number of groups in the world that can do this work at the frontier: a few hundred. Number of groups that actually publish ideas that scale up successfully: maybe a dozen per year.

Stage 2: scaling

If a research idea attracts enough attention, usually because the empirical results look unusually clean or because the underlying math is elegant, somebody with more compute will try to scale it up. This often happens by hiring the original author. Peebles wrote his DiTDiT (Diffusion Transformer)A diffusion model architecture that uses a Transformer instead of a U-Net. Introduced by Peebles and Xie in 2022. Scales much better than U-Nets and is the basis for almost all current frontier models. paper as a Berkeley graduate student in late 2022. He joined OpenAI shortly after. The Munich CompVis group wrote the latent diffusionLatent diffusionDiffusion that operates in a compressed 'latent' space produced by a VAE, rather than directly in pixel space. Dramatically reduces compute cost. Used by essentially every modern image and video model. paper at LMU in late 2021. They were hired by Stability AI in early 2022. The hiring is the mechanism by which research ideas migrate from academia to industry.

THE PAPER IS A JOB APPLICATION

An idea in this field can be born in a university lab, but it cannot be proven there, because proving it means a training run only a handful of companies can afford. So the moment a paper's scaling curves look clean, the fastest way to see the idea realized is not to fund the academic, it is to hire them. Peebles wrote the Diffusion Transformer at Berkeley and was at OpenAI leading Sora within about a year; the Munich latent-diffusion group was absorbed into Stability. The compute barrier does not merely filter who can build models, it structurally pulls the idea-generators out of the very institutions built to generate ideas. In a field this compute-bound, a landmark paper is less a finished result than a resume, and the field's most important research runs as a hiring pipeline from campuses into a dozen well-funded labs.

The scaling stage takes the original idea and asks the question that academia rarely answers: does this scale? You take the model from a hundred million parameters to a billion. From a billion to ten. From ten to a hundred. You scale the dataset proportionally. You watch the loss curves. If the curves keep going down, you have a winner. If they flatten or get worse, you have an architectural dead end. This is the stage that separates ideas that look pretty in a paper from ideas that actually work in production. Most ideas die here. The ones that survive become the basis for the next generation of frontier models.

The cost is meaningfully higher than stage one, millions to tens of millions of dollars per scaling experiment. The number of organizations in the world that can run it has grown over the past three years but is still limited to maybe thirty labs. The output is typically a follow-up paper showing the scaling results, and an internal demo that the company eventually decides whether to ship as a product.

Stage 3: production

If the scaling experiments work, the lab commits to a full production training run. This is where the budget becomes serious. You assemble a team of research scientists and engineers. You curate a massive dataset (which is itself a multi-month project with its own pipeline of scraping, filtering, captioningCaptioningWriting the text descriptions that accompany training images. In LoRA training, whatever you caption becomes a controllable variable and whatever you omit becomes baked-in identity., deduplication, and safety review). You run the model on thousands of GPUs for weeks or months, watching for instabilities, restarting failed runs, monitoring loss curves, doing intermediate evaluations. You build the inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. infrastructure that will eventually serve the model to users. You do safety review and content moderation. You distill the model into smaller variants. You apply RLHFRLHF (Reinforcement Learning from Human Feedback)Training that tunes a model toward outputs humans prefer, using human ratings as the reward signal. for the final preference tuning. You polish.

The cost ranges from ten million dollars for a frontier image model to over a hundred million for a frontier video model. The number of organizations that can do this stage is in the single digits, OpenAI, Google DeepMind, Meta, Anthropic, Alibaba, ByteDance, Tencent, Kuaishou, MiniMax, Black Forest Labs, Runway. Maybe a few others depending on how strict you are. The output is a production-ready model with weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable., an 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.), and a marketing launch.

Stage 4: platform

Once the model exists, somebody has to make it usable. This is the platform layer, Krea, Flora, fal.ai, Replicate, OpenArt, and dozens of others. They negotiate APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. access with the labs (or run open-weight models on their own infrastructure), build user interfaces, create workflows that combine multiple models, handle billing and rate limiting, and provide the actual experience that end users interact with. The technical work here is conventional Software as a Service (SaaS) engineering plus the additional complexity of managing 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.) inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training., but the value capture can be significant. Krea has raised meaningful venture capital. Flora raised $6.5 million in 2025. fal.ai is one of the better-funded inference providers.

The platform layer is where the field becomes accessible to non-developers. If your interaction with these models is through Krea's canvas or Flora's nodes, you are operating at the platform layer. The model labs do not see you directly, they see API calls coming from Krea's account.

Stage 5: you

And then there is the end user. Type a promptPromptThe text description you provide to a model to specify what you want it to generate., click generate, wait, iterate. Your role in this pipeline is to provide demand, taste signals, and sometimes feedback that flows back upstream to the platform and then to the model labModel labA company that trains and ships the frontier generation models themselves, such as Black Forest Labs, Runway, or Google.. The cost to you per generation is a few cents to a few dollars. The value chain you are participating in cost hundreds of millions of dollars to assemble. The leverage is enormous. You are getting the output of a process that, just five years ago, did not exist at all and that even today only a handful of organizations on Earth can run.

This is the main reason the field feels both abundant and concentrated at the same time. There are millions of users, hundreds of platforms, dozens of models, and only a tiny number of organizations that actually train them. Each layer is much smaller than the one below it. Most of the visible activity happens at the top, new tools, new tutorials, new shared workflows, but the actual capability comes from the bottom, from a small number of well-resourced labs running training jobs that nobody else can afford.