The economics of artificial intelligence are usually told as a spending story. Hundreds of billions of dollars in datacenters, clusters measured in gigawatts, a capital-expenditure arms race with the implicit premise that the prize goes to whoever trains the biggest, smartest model. That framing is not wrong about the money. It is wrong about where the money ends up.
Because underneath the capex boom, the price of using a fixed level of machine intelligence is falling faster than almost any input cost in the history of computing. And that collapse quietly dismantles the assumption the whole buildout rests on: that being smartest is where the value is.
The number
Start with a single figure from Stanford’s 2025 AI Index. The cost of running a system at the performance of GPT-3.5 fell from about $20 per million tokens in November 2022 to roughly $0.07 by October 2024 — a decline of more than 280-fold in under two years (Stanford HAI, 2025 AI Index Report).
Read that again. The same quality of output that cost twenty dollars became seven cents. Not cheaper at the margin — cheaper by more than two orders of magnitude, on a timescale that makes Moore’s Law look sedate.
This is not an artifact of one vendor discounting. It shows up directly in published list prices. OpenAI charged $60 per million output tokens for GPT-4 at its March 2023 launch. GPT-4 Turbo, eight months later, was $30. GPT-4o, in May 2024, launched at $15. GPT-4o mini, that July, arrived at $0.60 (OpenAI API pricing). A hundred-fold cut in the price of a capable answer, in sixteen months, from the market leader.
The mechanism
Three forces compound here, and it matters that they are independent. Hardware cost per unit of compute is falling around 30% a year, and energy efficiency is improving roughly 40% a year, per the same AI Index. On top of that sits the algorithmic layer: smaller models trained better now match the frontier of a year or two prior, so the compute you have to buy to hit a given capability keeps shrinking. Cheaper chips, run more efficiently, doing more per parameter. Each lever alone would be significant. Stacked, they produce the 280x.
And capability itself is converging. The gap between the best open-weight model and the best closed model on major benchmarks narrowed from 8% to 1.7% in a single year (Stanford HAI, 2025 AI Index Report). A model good enough for most commercial work is no longer a scarce, proprietary asset. It is available from several vendors, and increasingly free to download and run yourself.
Put the two trends together — price falling 280-fold, quality converging to within two points — and you get a market where the underlying product is racing toward being cheap, abundant, and interchangeable. That is the textbook definition of a commodity.
Where the value goes when the product is a commodity
Here is the part the capex headlines miss. When the marginal cost of intelligence trends toward zero and capability converges across vendors, “we have the smartest model” stops being a durable moat. A lead measured in months, over a rival selling the same capability for a fraction less next quarter, is not a moat. It is a depreciating asset the market keeps repricing downward.
So the value migrates, and it migrates in two directions at once.
It moves downstream, to demand. Whoever owns the user relationship, the workflow, the proprietary data, and the switching costs captures the surplus — because the model underneath is swappable. This is the real reason frontier labs are racing to become product companies rather than living on API margins. An API sells a commodity at a price the competition keeps cutting. A product people build their week around sells a habit. One of those has pricing power.
It also moves down, to the physical layer — not to whoever has the smartest model, but to whoever can serve a good-enough one at the lowest unit cost and largest scale. This is what the datacenter boom actually buys. Not intelligence supremacy; cost leadership in inference. The trillion-dollar buildout and the 280-fold price collapse are the same story read from opposite ends: you spend enormous capital not to have the best model, but to be the cheapest reliable supplier of a commodity whose price is falling under you.
The uncomfortable corollary
If intelligence is becoming near-free and undifferentiated, the businesses that win look less like inventors and more like aggregators of demand with the balance sheets to serve it — hyperscalers and incumbents with distribution already in hand. The pure-play lab whose only asset is a temporary quality lead is in the hardest position: it must convert that lead into either a product customers will not leave or the lowest cost-to-serve in the industry, and it must do so before the lead erodes. Selling raw model access, on its own, is selling a commodity into a deflation.
That is not a prediction that the labs lose. Some will make the conversion; the ones integrating forward into products and infrastructure are visibly trying to. It is a prediction about what they have to become to keep the gains their research creates. Being smartest is the entry ticket, not the moat.
The history of commoditized inputs — bandwidth, storage, raw compute — is consistent on one point. When the thing you produce keeps getting cheaper and better and more abundant, the rent does not usually accrue to the manufacturer. It accrues to whoever sits closest to the customer, or whoever can produce it cheapest at scale. The AI buildout is a bet, ultimately, on being one of those two. Making the smartest model was never the hard part of that bet. Keeping the money is.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.