Alex Dudov
Models improved, context windows expanded, reasoning capabilities advanced, and enterprises rushed to deploy copilots across entire organizations. The assumption was simple: AI was valuable enough that the economics would eventually work themselves out. Recent events suggest that assumption is facing its first serious test.
Microsoft reportedly canceled internal Claude Code licenses this week after token-based billing made the service too expensive to justify at scale. Around the same time, Uber's CTO warned employees that the company had burned through its entire 2026 AI budget in just four months. Meanwhile, AI software prices across the U.S. have reportedly increased between 20% and 37%, while GitHub is moving away from flat-rate subscriptions toward usage-based billing across its product portfolio.
Viewed individually, these are operational decisions. Viewed together, they look like the beginning of a broader repricing of AI consumption.
The End of the Unlimited Era
The shift away from flat-rate pricing may be the clearest signal. Unlimited plans work well when usage remains relatively predictable. They become much harder to sustain when a small percentage of users consume disproportionately large amounts of compute. AI products are particularly vulnerable because their costs scale directly with token consumption, inference complexity, and model size.
As reasoning models become more capable, they often become more expensive to run. The result is a growing mismatch between subscription-based pricing and the actual economics of serving enterprise customers.
GitHub's move toward usage-based billing reflects a reality that much of the industry has quietly acknowledged for months: somebody ultimately has to pay for the compute. That does not mean AI adoption is slowing. It means vendors are becoming less willing to subsidize heavy usage.
The Budget Constraint Is Back
For most of the current AI cycle, discussion focused on supply-side bottlenecks. Investors worried about Nvidia GPUs. Cloud providers worried about data-center capacity. Utilities worried about power demand. Semiconductor manufacturers worried about advanced packaging and HBM supply. A different bottleneck is now emerging.
When organizations begin tracking token consumption as closely as cloud spending, behavior changes. Teams become more selective about which models they use. Managers start questioning whether every workflow requires frontier-level reasoning. Finance departments begin asking how much productivity is actually being generated per dollar spent.
Those questions were largely absent during the experimentation phase of AI adoption. They become unavoidable once usage reaches enterprise scale.
| Signal | What It Suggests |
| Microsoft cancels internal Claude Code licenses | AI usage costs are being scrutinized even inside Big Tech |
| Uber exhausts annual AI budget in 4 months | Enterprise adoption can outpace budget planning |
| AI software prices rise 20–37% | Vendors are passing infrastructure costs to customers |
| GitHub shifts to usage-based billing | Unlimited plans are becoming harder to sustain |
| Token billing becomes standard | AI is increasingly priced like cloud infrastructure |
The signals are emerging from different parts of the market, but they point in the same direction. AI is becoming a metered resource. What began as a race to maximize adoption is increasingly becoming a race to control consumption. Vendors are adjusting pricing models, enterprises are revisiting budgets, and finance teams are beginning to treat AI usage the way they once treated cloud spending.
AI Is Following the Cloud Playbook
The pattern is familiar. Cloud computing began with a promise of flexibility and abundance. Over time, companies discovered that convenience often produced unexpectedly large bills. Entire disciplines emerged around cloud cost optimization. FinOps became a profession. Infrastructure efficiency became a competitive advantage.
AI appears to be entering the same stage. Organizations that once measured success by deployment numbers are increasingly measuring utilization, return on investment, and cost per task completed. Model quality still matters, but it is no longer the only variable.
The New AI Trade-Off
The industry's first chapter rewarded capability at almost any price. The next chapter may reward efficiency. That creates opportunities for smaller models, open-source alternatives, specialized systems, and vendors focused on reducing inference costs rather than maximizing benchmark scores. It also introduces a new competitive pressure for frontier-model providers, whose products must increasingly justify their expense in measurable business terms.
Microsoft canceling licenses, Uber exhausting budgets, and GitHub abandoning flat-rate pricing are not isolated anecdotes. They are early signs of a market becoming economically disciplined. The AI boom is not ending.
But the era of treating compute as effectively unlimited may be. And for the first time since generative AI entered the enterprise, cost is starting to matter as much as capability.
Alex Dudov
