Revolut's self-hosted GLM playbook: 56 B200s, TTFT cut from 10s to 1.6s, 1/10 the cost
AccBalanced · x · 2026-11-03
At AI Engineer NYC, Revolut's David Abitbol detailed the bank's self-hosted GLM setup serving 2,000 engineers:
- 56 B200 GPUs carry the entire load
- blended cost 30 cents/million tokens, invoice 1/10 of prior Anthropic/OpenAI spend
- since the April switch: 60% better coding time, 70% better lead time
- same GPUs, April to August: TTFT from 10s+ to 1.6s while req/s went from 3 to 7+
- Codex and Claude Code kept as exceptions
Transferable practices: set a TTFT budget first (target <3s; past 5s Slack fills up); cap context at 256K ('long context is a scam'); keep cache hit rate above 90% for coding. Lead time beating coding time because once tokens aren't rationed, engineers point agents at everything.
Related event: Revolut switches 2000 engineers' default coding model to self-hosted GLM(5 posts)→
More from coding & agent
- Dev floats idea of using agents to optimize Unreal Engine's source code — gajesh · 2026-11-03
- CleanWeb v2.8.4: SPA state parsing 95% faster, official LangChain/LlamaIndex packages — EstablishmentTough18 · 2026-11-03
- LangChain launches Managed Deep Agents: define agents as a directory of files — hwchase17 · 2026-11-03
- Designer builds AI website UX auditor on $1,500 budget, $7 spent so far — noya1979 · 2026-11-03
- Geiru Context Suite adds degradation protection for long MiniMax H3 video chains — roychodraws · 2026-11-03
- AI coding is 10–20x faster for the first 1–2 months, then complexity drags it back down — LadyAshBorg · 2026-11-03