Every AI demo you saw this year was running on someone else’s money.
The pilots in your company that look profitable today look profitable because a venture investor is quietly paying part of your compute bill. It is the business model of the companies selling you tokens. OpenAI is on track to burn roughly 14 billion dollars in 2026, and its own head of ChatGPT has called current pricing “accidental.” The subsidy will eventually end. When it does, the only AI projects that survive are the ones tied to an outcome you would happily pay full price for. Most are not.
How to tell which ones survives
In March 2026 OpenAI shut down Sora, its video product, because it was burning an estimated 15 million dollars a day against 2.1 million dollars in total lifetime revenue. Each ten second clip cost about 1.30 dollars to generate. The head of Sora called the economics “completely unsustainable.” That is a vendor with the cheapest possible access to its own models, and the numbers still did not work. Now ask what happens to a project paying the list rate a tier above that.
The first thing to check is your own business case. It was almost certainly signed at today’s prices, and today’s prices are happy hour. Cheap drinks because someone else is buying the round. Nobody modeled what the same workflow costs once the round ends and the tab arrives at full rates. That is not a knock on your team. The whole market is drinking at the same bar. But a business case that only clears at subsidized prices is not a business case. It is a dependency.
The second thing is that the trap is how a project is priced and scoped, not the price of a token. Most people get this backwards. Token prices are falling, hard. GPT-4 level capability cost about 20 dollars per million tokens in late 2022. By 2026 it is around 40 cents, down roughly 98 percent. And yet enterprise AI bills rose an estimated 320 percent over the same stretch. Both numbers are real. They diverge for one reason: agentic workflows. A single task no longer makes one model call. Gartner puts agentic systems at five to thirty times the tokens of a simple chatbot request. A workflow that cost four cents in 2023 runs about 1.20 dollars as an orchestrated agent today. The token got cheaper. The task got far more expensive.
Uber is the live example. It gave Claude Code to its engineers, ran an internal leaderboard ranking teams by AI usage, and watched adoption jump from 32 to 84 percent of 5,000 engineers in a single month. Cost ran 500 to 2,000 dollars per engineer per month. The company burned its entire 2026 AI coding budget in four months. Its COO put it plainly: if you cannot draw a direct line from the spend to features actually shipped, the trade gets hard to justify. That is a buyer discovering that a per seat mental model breaks the moment usage is uncapped.
The third thing is the part to act on. The projects that survive are tied to an outcome, where the workflow runs end to end and the value is measured, not assumed. The market is already moving there. Intercom charges 99 cents per resolved customer conversation. HubSpot dropped to 50 cents per resolved conversation in April 2026. Zendesk charges per automated resolution. You pay when the problem is actually solved. Contrast that with Salesforce Agentforce, which launched at 2 dollars per conversation whether or not it resolved anything, then had to add a per action option after customers pushed back. That gap is the whole game. When price is tied to a resolved outcome, the vendor absorbs the token volatility. When it is tied to a seat, you do.
A counterargument could be that prices are deflating so fast, that real costs simply stay low and the worry evaporates. Maybe. But do not build a strategy that needs both a venture subsidy and a hardware miracle to land on time. And the deflation case misses the point: track cost per business outcome, not cost per token. A token that got 98 percent cheaper but gets used thirty times per task is not cheaper.
How to go ahead in three questions
- Which of your live projects still clears its business case at three times today’s token prices?
- Which is priced and scoped against a measurable outcome rather than a seat?
- And which one, if you are honest, only works because someone else is paying part of the bill?
Then a next step. Re-run the numbers this quarter at triple the current rate. Kill or reposition anything that only survives at happy hour prices. And put one line into every vendor conversation from here on: show me the cost per outcome, not the cost per seat.
The subsidy is the round someone else bought. Build for the moment the tab lands on your table.
