Why token prices are falling but AI bills are rising

According to Ramp data, between January 2025 and April 2026, enterprise token consumption increased 1,000%, while spending increased only 500%. Usage growth is outpacing the rate of price decline by a factor of two.

There are three main reasons AI costs have risen.

• Change in the unit of consumption: In 2024, the basic pattern was one prompt → one response, but by 2026, an agent handles a single request by triggering dozens of model calls.
• Context bloat: RAG pipelines inject documents into every prompt, conversation history is resent with every turn, and reasoning models consume tokens for the thinking process itself.
• Background inference: Monitoring agents, compliance checks, and other processes that run continuously without any human request keep consuming tokens.

This is a textbook case of the Jevons paradox. As inference costs got cheaper, everyone adopted token-heavy approaches because they produce better results.

On top of this, supply-side pressure is also mounting. H100 rental prices bottomed out in October 2025, then surged 40% by March 2026, and GPU capacity through September is already fully booked.

Proposed countermeasures include matching models to the maturity of the use case, or, as Tesla does, setting a $200 per-employee weekly cap that requires manager approval to exceed.