Anthropic revenue +47% but compute bill +77%: cheaper tokens are making GPUs pricier

AccBalanced · x · 2026-11-15

jessiedong argues that cheaper inference is making compute more expensive: labs charge per token but pay per GPU-hour. Token prices fall due to lower model pricing, enterprise discounts, and workloads shifting to cheaper models—but cheaper tokens drive usage up, competing for scarce GPUs and pushing rental rates higher. The data backs it up: Anthropic's revenue grew 47% ($11.6B→$17.0B) last quarter while its compute bill grew 77% ($5.6B→$9.9B); token prices fell 15% ($0.62→$0.53/M) Oct–Nov while usage rose 29%; B200 rental rates climbed 14-15% to $6.87-$7.69/hour. Frontier labs earn less per token while paying more for the GPUs producing them.

Original post →

More from Venture

Venture channel →