How Revolut Cut Its AI Coding Bill 90% by Self-Hosting an Open-Weight Model

Description and use

Revolut cut AI coding costs by 90 % and sped up coding by 60 % by self-hosting the GLM-5.3 open-weight model on 56 Nvidia B200 GPUs.

Problem

Revolut’s AI coding assistants cost engineers up to $10,000/month with little measurable return.

Outcome

Revolut reduced AI coding inference costs by ~90% (to $0.30/million tokens) and improved coding time by ~60%, lead time by ~70%, while maintaining 56 GPUs and achieving 76% throughput gain.

Workflow

• Deploy Gemini company-wide (early 2025)
• Introduce Claude Code with per-engineer budgets (early 2026)
• Begin self-hosting GLM-5.1 (April 2026)
• Upgrade to GLM-5.3 (~754B params) and tune inference
• Maintain 56 B200 GPUs while usage grows
• Use frontier models only for exceptions

Evidence

Revolut did something almost no company of its size has tried: it ditched the frontier APIs as its default and moved its engineers onto a self-hosted, open-weight model.

Source

The result, per Revolut's own numbers, was a roughly 90% cut in AI coding inference costs — plus coding that got about 60% faster.

Source

GLM-5.3 (~754B total parameters, ~40B active per token), served from about 56 Nvidia B200 GPUs.

Source