
AI & The Subsidy Cliff
Somewhere out there is a company that forgot to set usage limits and woke up to a reported $500 million Claude bill. Not a typo. Half a billion dollars, spent by accident, because nobody was watching the meter. Uber blew through its entire 2026 AI coding budget by April. Microsoft handed its own developers Claude Code licenses and then quietly clawed them back, the corporate equivalent of regifting the present before the recipient finished unwrapping it.
A Priceline renewal came back four to five times more expensive, and the guy in charge of paying it, a man who spent his career in telecom billing and has therefore seen every trick, called it "the crack-cocaine epidemic." His words. "They let you try it to get you hooked on it, and now you're kind of beholden to it."
That is the TechCrunch story. You read it Friday. Good for you. Here is the part nobody connected to your business, which is the only part that matters to you.
The price of AI was never real.
Let me be specific, because specificity is the only thing that survives a board meeting. A token is the unit AI companies bill you by. It is a chunk of text, a word or a piece of one, and every prompt in and every answer out is metered in tokens the way a taxi meters miles. That part is boring. Here is the part that is not boring: per-token prices have fallen roughly 75 percent in a year, from about ten dollars per million tokens to two-fifty, and bills are going up anyway. Up sharply. Up 18.6x per developer in nine months, by one platform's count.
How does the price drop 75 percent while the bill quadruples? Because of agents. A chatbot answers a question and stops. An agent does a task, and to do a task it spins up a small army of itself, each copy burning tokens, checking its own work, calling tools, second-guessing, iterating, and billing you the entire time. A normal chat is two thousand tokens. A single agentic workflow is fifty thousand to five hundred thousand. Stanford clocked agentic tasks at a thousand times the consumption of a simple query, with the cost of the identical task varying by up to thirty times depending on how the agent felt that afternoon. The machine cannot predict its own bill. Neither can you.
So the industry did what industries do. It binged. There is a word for it, "tokenmaxxing," and it means exactly what it sounds like. Meta told employees to burn as many tokens as possible as a productivity metric, which is a sentence that should have ended several careers, and then walked it back when the CTO put in writing that "nobody should be using AI tools just for the sake of using them." Uber's COO went further and said out loud that all this spending showed no noticeable increase in productivity. One CTO told a vendor that a single engineer spent $40,000 in tokens in a month and he genuinely did not know whether to fire the man or clone him.
Now hold that picture, because here is the turn.
The whole thing ran on what one startup founder calls "subsidized intelligence." Investors were footing the bill. The model providers priced AI below what it costs to run, the time-honored Silicon Valley land-grab, get them hooked at a loss and raise the rent once they cannot leave. It is the most predictable plot in technology and it works every time.
And the rent is about to go up. OpenAI and Anthropic are both circling IPOs, which means they are about to acquire several thousand new bosses named "public shareholders" who expect, of all things, a profit. The subsidy ends precisely when the market needs it most. Goldman projects token usage rising 24x by 2030. The meter is speeding up and the discount is expiring in the same breath.
Here is the sentence your AI-native vendor will never say to you.
Half the AI companies in your stack have never once charged you what it costs to serve you. They have been losing money on every login, betting that prices would fall faster than you would use them. Prices did fall. You used more. They bet the wrong direction, on someone else's balance sheet, and they sold you a three-year contract on a cost structure that has never, not for a single quarter, actually closed.
You are not their customer. You are their burn rate.
Subscribe to our premium content at ADOTAT+ to read the rest.
Become a paying subscriber to get access to this post and other subscriber-only content.


