The AI subsidy era is ending, and the bills are finally showing up.
Microsoft is reportedly cutting back internal Claude Code licenses and steering developers toward GitHub Copilot CLI instead. It matters because this is Microsoft, not some cash-strapped startup. This is the company that poured billions into OpenAI, owns GitHub, and runs one of the biggest cloud businesses on earth. If even Microsoft is looking at token-based usage and deciding the math does not work, everyone else should probably pay attention.
This is not really an Anthropic problem. By most accounts, Claude Code is popular with developers. The issue is cost. Once AI tools move away from flat-rate pricing and into usage-based billing, companies stop buying the dream and start paying the meter. Every prompt, every code suggestion, every long context window, every agentic workflow starts turning into a line item.
GitHub is making the same shift. Copilot is moving to usage-based billing on June 1, 2026, with AI credits tied to token consumption across input, output, and cached tokens. That is a pretty clear signal from Microsoft’s own house: the flat-rate experiment is getting harder to defend as compute costs pile up.
Uber reportedly ran into the same wall. Its AI spending blew through planned budget assumptions months into the year, which is exactly what happens when adoption grows faster than cost controls. The problem is not that people are ignoring AI. It is almost the opposite. They are using it so much that the budget model breaks.
That is the part investors should care about. A lot of AI valuations assume usage keeps climbing, revenue keeps scaling, and costs eventually fall into a comfortable margin structure. Maybe that happens. But right now, enterprise customers are learning that serious AI usage is not cheap. The more useful these tools become, the more expensive they can be to run.
That creates a nasty fork in the road. Enterprises can scale back usage to keep budgets under control, which slows the revenue ramp AI labs need. Or the labs can cut prices to keep customers hooked, which pushes more pain onto their own margins. Neither path is fatal by itself, but both make the current valuation story harder to sell.
This does not mean AI is fake. It means the subsidy period is ending. The early pitch was simple: pay one flat price and let employees experiment. That phase created adoption, habits, and hype. Now the industry is moving into the metered phase, where usage gets priced closer to reality.
That is when the story gets more complicated.
AI can still be incredibly valuable. It can speed up coding, research, support, analysis, and back-office work. But value and affordability are not the same thing. If companies cannot connect usage to measurable output, the CFO is going to start asking hard questions.
We are not watching the AI boom collapse. Not yet. We are watching the free-sample table get packed up.
