Revolut cut AI coding costs by 90 percent in under a year. It switched from Claude to self-hosted GLM models. The move gave 2,000 engineers unlimited tokens for debugging and ticket work. Coding time improved by 60 percent and lead time by 70 percent. GLM-5.3 runs on 56 Nvidia B200 GPUs. Time to first token is under 3 seconds on 256,000-token contexts. Generation speed is 200 tokens per second. A model router still sends vision tasks to GPT or Gemini. Open source models are now a priority at Revolut. Chinese models are used on-prem to avoid privacy concerns.
What changed
Revolut relied primarily on commercial AI services like Claude for coding tasks.
What it unlocks
Engineers can now use AI for debugging and handling Jira and Confluence tickets without token limits.
- 90% cost reduction
- $0.30 per million input/output tokens
- 2,000 engineers
- 56 Nvidia B200 GPUs
What you need to act on it
- self-hosted infrastructure
- 56 Nvidia B200 GPUs
- GLM-5.3 model deployment
Sources