Revolut cuts AI coding costs about 90% with self-hosted models
Revolut moved its coding tools to open-weight models from Z.ai running on 56 Nvidia B200 GPUs, and says tuning raised throughput 76%.
Revolut has disclosed that moving its AI coding tools from external commercial services to self-hosted open-weight models cut its AI usage costs by about 90%, according to digitaltoday.co.kr. The company also reported gains in throughput and response times from the same setup.
David Abitbol, Revolut's head of AI operations and infrastructure, presented the case at the AI Engineer NYC 2026 developer event last month. Revolut began self-hosting Chinese AI company Z.ai's open-weight model GLM-5.1 in April 2026 and later expanded to GLM-5.3.
The fintech serves about 2,000 engineers with 56 Nvidia B200 GPUs. In its self-hosted environment, Revolut averages about $0.3 per million tokens for input and output combined.
Tuning context length, quantization and caching raised requests per second by 76% and cut time to first response by 88% while the GPU count stayed the same. Internal metrics show coding time improved about 60% and time to deployment readiness improved about 70%.
Revolut will keep using Google Gemini and other high-performance commercial AI for some tasks such as image analysis. Last year the company adopted Gemini, and early this year it began using Anthropic's Claude Code in earnest; some engineers' monthly AI usage costs reached $1,000 to $10,000, and after budget caps were introduced staff used AI only for complex coding tasks to save quota.
Quick answers
How much did Revolut save by self-hosting AI models?
Revolut says the switch cut its AI coding costs by about 90%, with its self-hosted environment averaging about $0.3 per million tokens for input and output combined.
Which models does Revolut self-host?
Revolut began self-hosting Z.ai's open-weight model GLM-5.1 in April 2026 and later expanded to GLM-5.3.
What hardware does Revolut use for self-hosting?
Revolut serves about 2,000 engineers with 56 Nvidia B200 GPUs.