Revolut self-hosts GLM 5.3 for ~2,000 engineers on 56 B200s

Revolut made a self-hosted open model the default for coding agents across a ~2,000-engineer sample. The talk is David Abitbol (Ops & Infrastructure / AI ops) at AI Engineer NYC, Oct 13, 2026 — "Shipping like a startup, regulated like a bank." Speaker framing: Revolut ~80M customers, 45 markets.

Why they capped Claude Code. Some engineers were spending $1k–$10k/month on Claude Code with weak ROI. Revolut put strict budget caps on frontier coding tools.

What replaced the default. Self-hosting open models started around April 2026 with GLM 5.1. The default now is GLM 5.3 on about 56 Nvidia B200 GPUs for those ~2,000 engineers. Frontier tools (Codex, Claude Code, Gemini) stay as exceptions, not the default.

Cost (their utilization, not a portable price). Blended ~$0.30 per million tokens. Invoice roughly 1/10 of prior Anthropic + OpenAI coding spend — Revolut's mix and utilization, not a list price you can copy.

The chart — and the catch. Abitbol showed ~60% better coding time and ~70% better lead time after GLM 5.1. The conference write-up is explicit: that chart does not isolate the model change from other causes. Abitbol's own explanation for lead-time gains: removing token rationing so engineers could use agents on inbox / Jira / Confluence work, not "open models beat Claude."

Same 56 GPUs, better infra (April → August). RPS +76%. TTFT from ~14.26s at migration / over 10s in the talk down to ~1.68s / ~1.6s — about an 88% TTFT cut. Target TTFT under 3s (Slack blows up over 5s). Interactive ~200 tok/s (up to 800 after 5pm).

Tuning notes from the talk. "Long context is a scam" for their coding path — 256K enough. FP4 on Blackwell. TP8 often better than TP4 for KV cache. Cache hit >90% for coding. Speculative decoding / PD disaggregation. Text-only GLM; route images elsewhere. OpenAI-compatible API so Codex / OpenCode-style harnesses keep working.

EU bank constraints. Human-in-the-loop sandbox (e.g. allow grep, deny gradle). Weights on a US-based service — not China-hosted inference. OpenTelemetry for observability.

For builders. Cap spend early. Prefer a cheap default + frontier exceptions. Measure TTFT and cache hit rate, not just model brand. And do not treat a before/after chart as a clean A/B on the model alone.

Not tested by MODEL. Not investment advice.

MODEL Editors