On October 22, 2026, the OnlyCFO newsletter publishes a guest article by Peter Walker, Head of Insights at OpenRouter, a gateway that provides access to several hundred language models through a single account. The publisher frames AI cost as on track to become companies' largest expense line, and the topic as a matter for the finance function.
OpenRouter's data show a rare combination: the volume processed rises from half a trillion tokens per week in January 2025 to 164 trillion by late September 2026, while the average price paid falls from $0.77 to $0.14 per million tokens. The author concludes that inference behaves like a commodity, not like software purchased from a vendor under an annual contract.
Four forces explain this evolution. Supply is exploding: 862 models launched since 2024, of which 27 exceed one trillion tokens per week. Open-weight models carry about 60% of tokens for 26% of spend; two-thirds come from Chinese labs, but hosting is chosen separately from the model, and 44% of those tokens are served by US-based providers. Agents became the top source of tokens on February 6, 2026; because they resend the same context at every step, prompt caching becomes the decisive price variable, and the cache hit rate for recent models reaches 88%. Finally, the leader keeps changing: five times in eight quarters.
Usage and spend decouple: at AI-native startups, tokens multiply by 18 and spend by 5.1. The decline stems from cheaper, more capable models, cache optimization, and frontier labs' economic models. For the same open-weight model, the price actually paid sits between 43% and 89% below list price, and the cheapest provider on the price list is often not the cheapest in practice.
The author therefore proposes buying at "landed cost": the price after cache discount, plus the cost of latency and the cost of retries. A one-dollar request that fails one time in five costs $1.25 per successful response. Add to that zero data retention, available for 93% of tokens, and availability: 96.9% success after automatic failover.
He concludes with four pieces of advice for CFOs: consuming more does not mean spending more, an open model varies by provider, guardrails are essential, and plans must remain revisable.