Overview
- Revolut moved off commercial coding APIs and now runs Z.ai GLM-5.x models on 56 Nvidia B200 GPUs inside its own environment to handle developer coding requests.
- The company reports roughly a 90% drop in inference costs, yielding a blended estimate of about $0.30 per million tokens after months of tuning and deployment work.
- Revolut rolled the tools out to roughly 2,000 engineers and says the change removed token rationing and produced measured productivity gains of about 60% faster coding time and 70% faster lead time.
- Getting these gains required large upfront investment in GPUs, platform engineering and months of optimization to push throughput above seven requests per second and cut time-to-first-token to roughly 1.6 seconds.
- To meet regulatory and security needs, Revolut keeps model weights on US-hosted infrastructure and requires human review of code suggestions, and it says the economics of self-hosting mainly favor very large teams because smaller groups face high capital and staffing costs.