The cost per million tokens fell 67% year over year. Enterprise AI bills rose 320% in the same period.
That gap is not an anomaly. It is the story.
This week provided two examples from opposite ends of the adoption curve. Uber burned through its entire 2026 AI coding budget in four months; its COO acknowledged he could not draw a line between rising token consumption and useful consumer features. Starbucks scrapped an AI inventory counting tool nine months after deploying it across North America; the system could not accurately count syrup bottles on shelves. In the same week, Starbucks announced it is tying tech worker bonuses to AI adoption metrics. The tool failed. The metric survived.
People in charge heard “AI adoption” and called it transformation. Eighty percent of Uber’s engineers were using agentic tools. Sixty percent of code commits involved AI. Internal leaderboards tracked usage rates. Starbucks built compensation targets around adoption goals. It looked like progress.
Then the bills came.
This is not a case against AI. Every technology category that produces durable value passes through this moment; the friction is the proof that an infrastructure layer is being built. The organizations that understand cost pressure as a design signal rather than a deployment ceiling are the ones that will carry a structural advantage when the infrastructure matures. That is exactly the bet worth making right now.
The token prices organizations built their models on were not market prices. They were subsidized prices; access priced below cost to drive adoption, the way every major software category has expanded before its economics normalize. The subsidy served its purpose. Adoption happened. Now the real numbers are arriving, and they look nothing like the spreadsheets from 2024.
Gartner projects token prices will fall 90% by 2030. The same research found cheaper tokens will not produce cheaper enterprise AI; agentic models require far more tokens per task, and providers will not fully pass savings through. Goldman Sachs projects token demand will grow 24x as agentic workflows scale. The meter runs faster as the use deepens.
This is the structural reality most organizations have not yet named: there is a tipping point ahead where the question stops being whether AI replaces humans. It becomes whether organizations can afford AI without them.
Not because AI cannot execute. Because execution without human infrastructure is too expensive to sustain.
Agentic workflows trigger 10 to 20 model calls per task. Each call processes context, generates output, reviews its own work, and revises. The cost compounds with every layer of ambiguity the agent has to resolve on its own. Clean data, defined processes, clear objectives; these are not soft organizational values. They are cost controls. Organizations that treat humans as the architects of those conditions will run AI at a fraction of what competitors pay for equivalent output.
One team of three spent $1.3 million in tokens in a single month running agentic tools. That is not a cost problem. That is a process problem with an expensive price tag.
The displacement decisions made before this moment were built on the assumption that AI was cheaper than people for every task. For simple, well-defined work, it is. For complex, agentic, cross-system work running without process design, it is not. The people who designed those processes were the infrastructure. Organizations that eliminated them to fund AI adoption discovered they had removed the scaffolding needed to make AI economical.
The measure of proper implementation is not adoption rate. It is whether the humans still in the system are doing work that makes the AI around them cheaper, faster, and more reliable. That is not a support role. That is architecture.
A token is a unit of compute. A process is a unit of human intention. One without the other is entropy with a budget attached. The organizations that solve that equation first will have a cost structure their competitors cannot match. That is the real prize inside a rising token bill.