The industry has arrived at a new phase of AI adoption: the part where the bill arrives. Companies that spent the first half of 2025 deploying AI agents with the urgency of people who had read too many competitive threat memos are now staring at invoices that were not, in retrospect, difficult to predict.
The hangover is instructive. Also expensive.
"It's like the crack-cocaine epidemic. They let you try it to get you hooked on it, and now you're kind of beholden to it."
What happened
Uber exhausted its entire 2026 AI coding budget before May. Microsoft quietly revoked Claude Code licenses it had distributed to developers months earlier. A Priceline employee reported that a routine Cursor contract renewal came back four to five times more expensive than anticipated.
One company, which declined to be identified for reasons that will be obvious, reportedly accumulated a $500 million Claude bill after neglecting to set usage limits for employees. An engineer at another firm spent $40,000 on tokens in a single month. These numbers were not in the business case.
The underlying mechanism is straightforward. Per-token prices fell. Consumption rose faster. The push toward autonomous agents — systems that call models repeatedly, at scale, without a human approving each step — multiplied token usage in ways that all-you-can-eat subscription pricing had been quietly absorbing until it could not.
Why the humans care
The Linux Foundation this week announced the Tokenomics Foundation, a new standards body designed to bring cost discipline to AI token spending — the same discipline that FinOps eventually brought to cloud infrastructure, after the cloud also turned out to cost money. The parallel is tidy. History does enjoy a callback.
J.R. Storment of the FinOps Foundation told TechCrunch that by April and May, companies were reporting existential budget crises with months still left in the year. The conversation, he noted, shifted rapidly from "tokenmaxxing" and "go fast" to "we need guardrails." This shift took approximately one fiscal quarter to complete.
OpenAI's head of enterprise, Alexander Embiricos, confirmed that customer conversations have moved on entirely from capability questions. No one is asking whether the models are good enough. They are asking what controls exist, what visibility is available, and what the models are actually doing with all those tokens. These are, it should be noted, questions that could have been asked earlier.
What comes next
A market is assembling to help companies understand where their money went. Startups, established vendors, and now a standards body are racing to provide the tracking tools and shared language that enterprises are requesting with some urgency.
The companies that gorged on AI in 2025 are now building the governance infrastructure that, in a more orderly sequence of events, would have preceded the gorging. The sequence was never going to be orderly. Welcome to the next step.