In the last two weeks of April, my AI tooling consumed 17.7 billion tokens. At Anthropic's API list prices, that fortnight of compute is worth a hair over $15,000 - call it $30,000 a month at the same pace. I paid $800.
Every "stop renting intelligence, buy your own GPUs" thread I read falls apart on contact with this workload. So I ran the numbers properly: my measured usage profile against GPU rental prices, on-premise TCO, and the best open-weight models available in June 2026. The hardware loses, and not by a little. The reasons it loses are more interesting than the price tag.
The workload
All figures below are measured from complete local session logs, April 16-30, 2026. Personal projects only, employer work excluded. Monthly numbers are that pace doubled - and the weeks after this window ran hotter, not cooler, so the extrapolation is the conservative read. Pricing basis: Anthropic list as of June 2026 ($5/$25 per million input/output tokens, $6.25 cache writes, $0.50 cache reads). External prices and benchmarks throughout are as of June 12, 2026.
Metric
