The Cheapest Tokens We've Ever Had. The Biggest Bill.
Token prices are down 98%. Enterprise AI bills are up 320%.
That pair of numbers ran on Metrics that Measure Up this week and it is the whole software story right now. GPT-4 cost roughly $20 per million tokens in late 2022. The average is about 40 cents today. Over the same stretch the typical enterprise AI budget went from $1.2 million to $7 million. Cheaper per unit, far more expensive in total, and by the read of Benchmarkit founder Ray Rike and New Plan's Peter Buchanan, almost nobody is tracking it properly.
Where the money comes from is the part worth noting. Not a new budget line. Attrition. The roles that quietly do not get backfilled. Nearly half of this year's tech cuts are being attributed to AI by the companies making them, at a pace running about a third faster than 2025. The layoffs get the headline. The unbackfilled req funds the token spend.
Which is why the org chart is now the AI strategy. ClickUp's Dan Zhang, on the same show the next day: "We don't think the AI transformation is a technology problem. We think it's a fundamental business design problem." He got there by failing twice. First a top-down mandate, every all-hands about AI. Then bottom-up, hackathons everywhere and, in his words, "tokens burning." Both failed. What worked was redesigning the function before approving the pilot.
The result shows up in next year's plan. Over half of ClickUp's functions are budgeted for no new headcount, and the functional leaders volunteered for it. "I've never seen that kind of confidence," Zhang said. Their revenue per account doubled in twelve months.
The demand side ran the other way. Martin Eriksson brought the Jevons paradox to ProductTank SF with the radiology numbers. AI mammography screening detects 29% more cancers and cuts workload 44%. There are also 16% more radiologists in the US than a decade ago, their salaries are up 44%, and the Mayo Clinic went from 200 to 400. "When you reduce the cost of solving one problem, you don't run out of problems," he said. "You just unlock the next layer."
The shape of the week: AI stopped behaving like a software line item and started behaving like compensation. It scales with people, it gets funded out of headcount, and it grows faster than its unit price falls. That is why it landed on the CFO's desk and not the CIO's.
In our corpus, "token" appeared in 5.4% of business podcast episodes in July 2025. This month, 14.9%.
Sources: Ray Rike (Benchmarkit) and Peter Buchanan (New Plan), Metrics that Measure Up, "AI is a Compensation Scale Expense," Jul 21, 2026. Dan Zhang (ClickUp), Metrics that Measure Up, "Building the 100x Org: The CFO as AI Architect," Jul 22, 2026. Martin Eriksson (Mind the Product), The Product Experience, "The Jevons paradox: why AI creates more demand," Jul 20, 2026.
Enterprise AI spend is scaling with headcount rather than seats, funded largely through roles that go unbackfilled, which is moving AI budget ownership from the CIO to the CFO.