Revolut cuts AI costs 10x by self-hosting GLM on 56 B200s, boosting lead time 70%
SumitGup · x · 2026-11-02
Revolut, Europe's largest private company, has made self-hosted Chinese open-weight models its coding default and moved frontier APIs like Claude Code and Codex to an exception list. Per AI ops lead David Abitbol, running GLM on 56 Nvidia B200s yields a blended cost of $0.30 per million tokens — roughly a tenth of what Revolut paid Anthropic and OpenAI — while improving engineering lead time by 70%.
The tweet also cites Decagon running 90% of workloads on open-weight models, and argues the future isn't binary: use the smartest general model for new use cases, but once workflows mature, smaller, faster specialized models increasingly win on cost and latency.
More from coding & agent
- Agents Make the One-to-One World Economically Feasible, Beyond Feed Ranking — signulll · 2026-11-02
- Musk promotes Grok in Cursor; dev says UX beats Claude Code and Codex — elonmusk · 2026-11-02
- Comic nails the agent economy: humans just approving approvals, forever — sierracatalina · 2026-11-02
- Claude Opus as project manager: a fully automated AI video production pipeline — Familiar_Self5368 · 2026-11-02
- Back up your Grok bot daily to a private repo: 20+ agents of skills and memories — 0xkarasy · 2026-11-02
- New study examines how developers actually supervise autonomous coding agents — burkov · 2026-11-02