Revolut cut its coding-agent bill 10x by self-hosting an open model on 56 B200s

ccerrato147 · x · 2026-11-05

The wrap-up of Revolut's coding-agent story: after switching to a self-hosted Chinese open-weight model on 56 dedicated B200s at 30 cents per million tokens on a flat invoice, the math works out to roughly a tenth of the old bill — public rental pricing for 56 B200s would be $330K/month, or about $165 per engineer, versus Anthropic's cited $150-$250 per Claude Code developer. The author's core argument: rationing was the real cost — a capped engineer saves tokens for the hardest task, while an uncapped one hands the agent everything.

Related event: Revolut moves 2,000 engineers to self-hosted GLM, cutting AI coding costs ~90%(10 posts)→

Original post →

More from coding & agent

coding & agent channel →