Supporting case
Revolut’s self-hosted coding stack shows independence that requires substantial scale
What happened
A November 3 report draws attention to Revolut infrastructure leader David Abitbol’s October 13 conference account of the company’s coding setup. He describes running GLM models itself as the default for roughly 2,000 engineers, using about 56 Nvidia B200 processors, while retaining hosted frontier models for exceptions. He reports a blended cost of about 30 cents per million tokens and a bill around one-tenth the prior Anthropic/OpenAI spending. Primary talk and transcript ↗.
These are the operator’s figures. The account does not independently establish comparable capital-cost treatment, utilization, engineering labour or an audited productivity gain. The concrete change is that Revolut can serve much of its coding demand under its own deployment controls, rather than rely entirely on another provider’s metering and service decisions. It still depends on models, advanced hardware and specialist operating knowledge.
The Race implication is mixed. Self-hosting can retain control in the customer institution, but a bank-scale cluster is not evidence that a small organisation’s minimum viable scale has fallen. It is a useful qualification to the focus on smaller local components, not proof of complete workflow portability or the end of vendor dependence.
Dates and implementation stage
Action/stage Deployment date unresolved; operator account 2026-10-13; reporting 2026-11-03; approved 2026-11-05
