When founders and technologists deeply involved in AI meet the question we ask is "Where is this all headed?"
I am not talking about societal upheaval. Often the question's boundary is what will happen to AI businesses. When we talk about AI adoption, we talk about how long it takes for industries to adopt. That's diffusion. But there is another kind of diffusion. It's about how long it takes for various parts of AI value to commoditize.
Many moving parts exist, and each will become a commodity eventually. Along the way markets will be made, remade, and decimated.
I summoned the inner P.B.Shelley. It turns out I don't have one. So here is a high-school-level poem on how AI becomes a commodity and how moats will shift. I hope it is sufficiently Ozymandian.
Frontier models are commodity.
Inference is where the moat is.
Inference is commodity.
Apps are where the moat is.
Apps are commodity.
Workflows are where the moat is.
Workflows are commodity.
Verticals are where the moat is.
Verticals are commodity.
Data is where the moat is.
Data is commodity.
Synthetic data is where the moat is.
Synthetic data is commodity.
Evals are where the moat is.
Evals are commodity.
Feedback loops are where the moat is.
Feedback loops are commodity.
Distribution is where the moat is.
Distribution is commodity.
Liability is where the moat is.
Liability is commodity.
Judgment is where the moat is.
Judgment is commodity.
Frontier models are where the moat is.
There is no moat.
There is only cash flow.
The explanation, if you dislike poems:
Frontier models → Inference. For two years the entire industry agreed that whoever trained the biggest model would own the future. Then the capability gap narrowed to about six months and open weights arrived to finish the job. Attention moved to whoever could serve tokens cheapest, which sounded like a moat until we said the words "serve tokens cheapest" out loud.
Inference → Apps. Token prices fell roughly an order of magnitude a year, which is a wonderful thing for humanity and a terrible thing for the labs' gross margin. Serving inference turned out to be a utility business staffed by people who had recently raised at a software multiple. Everyone agreed the real value was closer to the user.
Apps → Workflows. Being closer to the user meant a page, a text box with a send button in the SaaS world. A competitor could ship the identical product over a long weekend and they did. The consensus updated: the durable thing was encoding how the work actually happens, with the approvals and the exceptions and the person who has to sign off.
Workflows → Verticals. Then everybody wrote a blog post about their agentic workflow architecture, complete with diagrams, and the workflows were copied within the quarter. Domain specificity became the answer. You cannot fake knowing what a clean claim is, went the reasoning, from people who had learned what a clean claim was eleven weeks earlier.
Verticals → Data. Domain expertise turned out to be four subject matter experts and a year of listening. Proprietary customer data was declared the true asset, on the theory that nobody else has it. This was the most emotionally satisfying rung and lasted the longest.
Data → Synthetic data. Enterprise data was smaller, dirtier, and more duplicated than the pitch decks implied, and models became sample-efficient enough that example number 10000 contributed roughly nothing. Better to generate exactly what you're missing. The moat became a prompt that says "produce two thousand variations of the following."
Synthetic data → Evals. Since anyone can write that prompt, the scarce skill became knowing whether the resulting model is worse than last Tuesday's. Evaluation harnesses were declared the real intellectual property, and for a brief period this was even true. Then people started publishing them.
Evals → Feedback loops. An eval describes a moment that has already passed. What compounds, we were told, is production signal: users correcting outputs, corrections improving the system, forever. This is a genuinely good mechanism, though it does require users.
Feedback loops → Distribution. It requires quite a lot of users, in fact, and the loop tooling got packaged and sold by six vendors. Whoever has the customers has the loop. Everyone rediscovered that enterprise software has always been a sales business and pretended to be surprised.
Distribution → Liability. Channels are rentable. You can buy the same conference booth, the same partner program, the same outbound sequence, and your competitor will. What isn't purchasable is agreeing to be the one who pays when the agent is wrong.
Liability → Judgment. Liability is a thing insurers price, and once it has a price it has a purchase order. It also only covers the small fraction of work that fails loudly enough to be actionable. The rest of the job (the overwhelming majority of it) is calls where there is no correct answer, no clean error, and nobody to sue. Which invoice to chase first. Which exception to escalate and which to let ride. Whether this customer is annoyed or leaving. So the retreat went there: not to who pays when it breaks, but to who decides when nothing is obviously broken.
Judgment → Frontier models. Good judgment, written down, is a checklist. Checklists are training data. Every quarter another category of expensive human discernment quietly ships as a default behaviour in a model release. Congratulations: you have arrived back at rung one, where a lab you don't control eats the layer you just spent eighteen months defending.
There is no moat. There is only cash flow. Every rung was a real claim, defended by real people, funded at a real premium, and correct for a few months to a few years. The pattern isn't that they were wrong. It's that the half-life keeps shortening.
I wrote a poem and a long-winding explanation to make a simple point: Being in the business of AI is being in the business of selling a commodity. The ones that win play the arbitrage between the inception of a piece of the missing infra and its ultimate loss of value in the chain as it gets prevalent.