AI providers are selling enterprise seats at prices that barely cover compute. OpenAI, Anthropic, Google and others run loss‑leader programs, offering models that cost many times more to run than the subscription fees. A Claude Pro seat costs $20 a month, yet the same workload can consume millions of tokens, translating to $200‑$400 of API cost per seat. Companies that have built critical workflows around these flat fees sit on a financial time bomb.

The shift to agentic AI has broken the subsidized math. Claude Code sessions and GitHub Copilot agents run autonomously, burning tokens far faster than simple chat. Power users of Copilot have reported compute costs of $80 a month on a $10 subscription, and Anthropic’s analysis shows $8 of compute for every $1 of revenue. Providers are now moving to usage‑based billing as the flat‑fee model collapses.

Enterprises that have embedded AI into marketing, engineering, research or finance now face a hidden line item that can eclipse traditional SaaS budgets. KPMG’s Q1 2026 AI Pulse projects U.S. firms will spend $207 million on AI over the next year, nearly double last year’s figure. When providers correct pricing, those firms will see immediate, sizeable bill shocks rather than a gradual rise.