48-hour inference run estimated at $13.8M

GregKamradt · x · 2026-08-28

A back-of-the-napkin calculation estimates a 48-hour inference run at $13.824 million, assuming 400M tokens/min throughput and a 50/50 input/output split without prompt caching. Real-world costs would likely be lower due to caching and log optimizations.

Related event: Estimated 48-Hour Inference Cost for OpenAI Event: $13.8M(2 posts)→

Original post →

More from Infra

Infra channel →