Revolut moves 2,000 engineers onto self-hosted AI, cutting coding costs 90%
What's the deal? Revolut has shifted roughly 2,000 engineers off Anthropic's Claude Code and OpenAI's Codex onto self-hosted GLM models, cutting its AI coding costs by about 90%. Ops and infrastructure leader David Abitbol detailed the move at the AI Engineer NYC conference on October 13, 2026.
Why now? Early rollout of Claude Code proved shaky economically — some engineers burned through large token budgets without a clear return, so Revolut imposed limits. That created a second problem: when tokens feel scarce, engineers hoard them for hard tasks instead of using AI across the surrounding work.
What's the endgame? Self-hosting GLM gives engineers freedom to use AI without unpredictable bills. Abitbol says the bottleneck was never just coding speed; delays also come from the inbox, Jira, and other work around the code. Cheaper access widens which tasks people delegate.
The numbers. Revolut runs 56 B200 GPUs to serve GLM 5.3 to its engineers, at a blended cost near $0.30 per million tokens — roughly a tenth of its prior Anthropic and OpenAI invoices. The company reports about 60% better coding time and 70% better lead time, though those figures are self-reported and the gains may partly reflect removing usage limits rather than the model itself.
Tune before you buy. Loading weights onto GPUs alone did not deliver the performance Revolut wanted, Abitbol said. After tuning its serving system against real demand, the team reported 76% higher requests per second and an 88% cut in time to first token — from 14.26 seconds during migration to 1.68 seconds once stable. His advice: resist adding GPUs until the existing setup is tuned.
What could go wrong? Self-hosting is not a free win. GPU investment and operating expertise can erase savings, and Revolut's internal comparison does not isolate the model change from other causes. Its scale and infrastructure make the case hard to replicate at smaller firms.
The signal: Revolut puts open-weight AI into a scaled, daily workflow at a regulated fintech — not a pilot. It reflects a broader shift toward cheaper open models for routine, high-volume work, with frontier services held in reserve for tasks where they still perform better.
Read more: ai.engineer
