The economics of AI usage are shifting faster than most companies anticipated, and the early signs of a correction are visible.
Uber has introduced internal caps on AI spending, limiting how much employees can use AI tools inside the company. Microsoft has started charging more per token for GitHub Copilot, an acknowledgment that the service has been subsidised by investor capital rather than priced at its true cost.
Six months ago, unlimited AI usage was a perk. Now it is a budget line that managers are being asked to control.
Subsidy problem
The current pricing of AI services does not reflect the cost of providing them. OpenAI charges $20 a month for ChatGPT Plus. The compute required to serve a heavy user costs considerably more than that. The gap has been filled by investor capital, with AI companies deliberately underpricing their products to drive adoption and capture market share.
That strategy works until it doesn't. The inflection point arrives when companies start passing costs through to end users, either by raising prices or by limiting usage.
Microsoft's decision to increase Copilot pricing is the clearest signal that the subsidy era is ending. If the most widely adopted AI coding tool in the world cannot sustain its pricing, smaller players with thinner margins face a harder reckoning.
Enterprise squeeze
Uber's internal caps illustrate a different dimension of the same problem. Enterprises adopted AI tools with enthusiasm and without budgets. Usage grew faster than anyone modelled. Now CFOs are asking what the AI spend is actually delivering and whether the productivity gains justify the cost.
The answer, for many use cases, is not yet clear. AI tools are useful. Whether they are useful enough to justify uncapped spending at enterprise scale is a question most companies are still answering.
What comes next
AI costs will eventually fall. Models will become more efficient. Inference costs will decline as competition increases and hardware improves. But the timeline between now and that future is where the pain sits.
Companies that built workflows around cheap AI tokens will face a choice: absorb higher costs, reduce usage, or pass the increase to customers. None of those options is painless.
The token bubble was always going to correct. The speed at which it is happening, from unlimited usage to internal caps within six months, suggests the correction will be sharper than the market expected.
