Sam Altman wrote in June 2025 that intelligence would converge to near the cost of electricity. One year later, AI inference costs are trending the other direction.
The original article tracks what the author calls the 'token reckoning': the gap between Altman's 'cheap intelligence' thesis and the actual billing reality hitting enterprise users and developers right now. The argument is not that AI got more expensive in absolute terms, but that consumption patterns, context window usage, and agentic workloads are producing costs that scale in ways nobody budgeted for. The AC bill analogy is precise and deliberate.
Read the full piece for the spending breakdown and the specific workload types driving the overage. The conclusion is less important than the mechanics explained in the middle, which reframe how to think about AI cost curves going into late 2025.
[READ ORIGINAL →]