FinanceHigh importance

Revolut cuts coding inference cost ~90% with self-hosted GLM-5.3 on 56×B200

At AI Engineer NY, David Abitbol (Revolut ops/infra) described the pivot: ~2,000 engineers, coding time +60% and lead time +70% after the coding-AI rollout; the default became self-hosted GLM (from 5.1 to 5.3) on 56 Nvidia B200 GPUs. Cost ~US$0.30 / million tokens — about 1/10 of Codex/Claude Code — with ~90% cut in coding inference spend. Stack: NVFP4, disaggregated prefill/decode; sub-3s TTFT at ~256k context (Abitbol: “long context is a scam”); ~200 tok/s. A router sends vision to Gemini/GPT Luna; commercial frontier stays the exception.