Revolut’s David Abitbol describes making self-hosted GLM the default for AI coding across roughly 2,000 engineers, while retaining frontier services for exceptions.

The deployment uses 56 Nvidia B200 GPUs. He reports $0.30 per million tokens, blended, and spending around one-tenth of the previous Anthropic/OpenAI invoice.

The interesting part is usage: removing token budgets let engineers bring AI into inbox searches, Jira and Confluence instead of saving it for difficult coding tasks. Revolut’s internal chart reports 60% improvement in coding time and 70% in lead time—from picking up a feature to having it merged and ready to deploy.

This is a self-hosted GPU service. These are internal before-and-after figures: models, serving optimizations and usage policies changed together. The talk supplies no itemized total-cost breakdown.

October talk and transcript:
ai.engineer/nyc/2026/talks/revolut