Artificial Intelligence: Follow the Cost Curve, Not the Benchmarks
Most writing about artificial intelligence is about capability — what a new model can do that the last one could not. That is genuinely interesting, and it is a poor guide to what will happen commercially.
A more useful lens comes from asking a different question: what is getting cheaper, and what stays scarce?
Because the striking thing about this technology is not any single capability. It is that the price of a unit of machine reasoning has been falling steeply and continuously. Things that were too expensive to automate two years ago are routine now. Things that are borderline today will be trivial soon.
When the cost of something falls by an order of magnitude, you do not get a slightly better version of the old products. You get categories of product that were previously impossible, because the economics finally work. That is the story worth following.
The cost curve, and why it matters more than benchmarks
A useful analogy: when computer storage was expensive, software was written to conserve it, and whole product categories — photo libraries, video streaming, continuous backups — simply could not exist. Storage becoming almost free did not make existing software better. It made different software possible.
Machine reasoning is going through the same transition, faster.
This has three practical consequences.
Products should be designed assuming cost keeps falling. A service that is barely viable today at current prices will be comfortably profitable if the underlying cost drops again. Conversely, a business whose advantage is being cheap at running models is standing on ground that is moving.
"Too expensive to automate" is a temporary statement. It describes this year's prices, not a permanent boundary. Anyone planning around it should attach a date.
Capability improvements and cost improvements are different things. A model that is marginally smarter changes little. A model that is equally smart at a tenth the price changes what you can build. The second gets far less attention.
The stack, and who captures the value
Look at the industry as layers: energy and capital at the bottom, then chips and compute, then models, then the applications people actually use.
Each layer would like the layer beneath it to be cheap and interchangeable, and its own layer to be scarce and differentiated. That is commoditize your complement operating at industrial scale, and it explains a lot of behaviour that otherwise looks strange — including large companies giving away capable models for free. A free model is excellent news if you sell the compute it runs on, or the cloud it sits in, or the product it makes better.
Two structural facts shape where value settles.
The bottom of the stack is unusually capital-intensive. Frontier model training requires spending that resembles building infrastructure rather than writing software. This is a genuine change from the previous era of technology, where a small team with good ideas could compete with anyone. Capital intensity favours large balance sheets, which is a different competitive landscape than software has been used to.
The top of the stack is where customer relationships live. And per aggregation theory, whoever owns the demand side has historically been the durable winner, because suppliers below them compete against each other for access.
Which of these dominates is genuinely unsettled. That is the central open question in the industry, and anyone claiming confidence about it is guessing.
What makes an AI product defensible
A great deal of what is currently built is a thin layer over someone else's model: a prompt, an interface, and a subscription. These are often useful and rarely defensible, because the model provider can absorb the same functionality, and so can the next twenty startups.
The things that hold up are the things that always held up.
Owning the customer relationship. If people come to you first, suppliers below you are interchangeable. If you are a feature inside someone else's product, you are the interchangeable one.
A data loop that improves the product. If usage generates data that makes the service measurably better for the next user, you have a network effect that a well-funded competitor cannot copy by spending money. Worth being sceptical here — this is claimed far more often than it is true, and many "data advantages" saturate quickly.
Workflow integration and switching costs. A tool that has absorbed a company's processes and history is hard to leave regardless of whether a rival's underlying model is better.
Being the scarce input. Proprietary data, exclusive distribution, regulatory clearance — the ordinary sources of economic moats.
Notice that none of these is "we have the best model." Model quality is a moving target that every competitor is also chasing, which is exactly the Red Queen situation: run hard, stay in place.
What is genuinely uncertain
Honesty is worth more than confidence here, so the open questions are worth naming plainly.
Whether models become commodities is unresolved. If several are roughly equivalent and cheap, value moves to applications and distribution. If a meaningful capability gap persists, value stays with whoever holds it.
Whether the capital intensity continues is unresolved. Frontier training costs may keep rising, concentrating the industry, or efficiency gains may bring them down and reopen it.
And the returns question is unresolved. Enormous sums are being spent against expectations of future revenue. Both a durable transformation and a painful correction on the way to one remain consistent with what we can currently observe.
The cost curve, though, is not really in doubt. Whatever else happens, reasoning is getting cheaper — and that alone will keep making things possible that are not possible today.
