Revolut self-hosts GLM on 56 B200s, cutting coding token cost to ~$0.30/M

alex_verem · x · 2026-11-05

An article by alexverem breaks down how Revolut cut its AI coding bill by self-hosting an open model:

2,000 Revolut engineers use AI coding tools; on Claude Code some spent $1,000-$10,000/month, forcing the company to impose budget caps.
In April Revolut made Zhipu's open-weight GLM its default coding model, self-hosted on 56 B200 GPUs.
Cost per million tokens: Claude Opus 5.5 API $20 in / $4 out; GLM-5.3 API $4.40 / $1.40; Revolut's self-hosted blended cost $0.30.
Break-even math: a rented 2-GPU B200 server costs $10,059/month, so the API is cheaper for 10 engineers; 4-GPU breaks even around 20 engineers; an 8-GPU server at $39,070/month still wins at 40 engineers — provided GPUs stay busy.

Takeaway: self-hosting open models takes pricing power back from frontier labs; utilization determines per-token cost.

More from coding & agent