Understand this story
British fintech company Revolut has shifted its primary AI programming from commercial services such as Claude Code and Gemini to self-hosted Zhipu GLM-5.1 and GLM-5.3. Its head of operations and infrastructure, David Abitbol, revealed that the comprehensive cost of self-hosted GLM-5.1 is about $0.30 per million tokens; to this end, 56 NVIDIA B200 GPUs were deployed, cutting inference costs to about one-tenth of the original. The coding time of about 2,000 engineers was shortened by about 60%, and the delivery cycle was shortened by about 70%. The team also used NVFP4 quantization, split Prefill and Decode, and optimized caching and context configuration to reduce first-token latency from 14.26 seconds to 1.68 seconds without changing the number of GPUs; visual tasks such as image recognition are still handled by Gemini and GPT.
AI-generated from public reports
Latest development · Report
Revolut Switches to Self-Hosted GLM to Support 2,000 Engineers, Cutting AI Coding Costs by About 90%Developments
1 confirmed developments
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 4
36Kr · AI
Revolut Switches to Self-Hosted GLM to Support 2,000 Engineers, Cutting AI Coding Costs by About 90%
UK fintech company Revolut has shifted its primary AI coding workload from commercial services such as Claude Code and Gemini to self-hosted Zhipu GLM-5.1 and GLM-5.3, reducing inference costs to about one-tenth of the original. For roughly 2,000 engineers, coding time has been cut by about 60% and delivery cycles by about 70%. David Abitbol, its head of operations and infrastructure, revealed that the comprehensive cost of self-hosted GLM-5.1 is about $0.30 per million tokens, for which 56 NVIDIA B200 GPUs were deployed. The team also used NVFP4 quantization, split Prefill and Decode, and optimized caching and context configuration to reduce first-token latency from 14.26 seconds to 1.68 seconds without changing the number of GPUs; visual tasks such as image recognition are still handled by Gemini and GPT.