Revolut has made self-hosted open-weight GLM models its default for coding across roughly 2,000 engineers, reserving tools such as Claude Code, Codex and Gemini for harder tasks. AI operations lead David Abitbol reports a blended cost of about 30 cents per million tokens - around one-tenth of the company’s previous spending on Anthropic and OpenAI. In the year-long comparison, Gemini improved engineering lead time by 10% and coding time by 20%; self-hosted GLM later improved them by 70% and 60%, respectively. The larger lead-time gain reflects time saved on work around coding, such as searching documentation and handling inboxes or project-management tasks. Low costs also let engineers use AI more routinely instead of rationing expensive requests.
Revolut serves the models on 56 Nvidia B200 GPUs, but improved performance by tuning rather than expanding the cluster: from April to August 2026, time to first token fell from over 10 seconds to about 1.6 seconds as demand rose 76%. Abitbol emphasizes practical configuration - shorter contexts, FP4 precision, high cache-hit rates and optimized inference - and argues that simply adding hardware is not the answer. For regulatory compliance, Revolut downloads model weights and runs inference in approved locations rather than sending company data to Chinese endpoints. Human approval controls, sandboxing, data-retention limits and detailed logging support that approach. The experience suggests self-hosting can reduce costs and provide predictable bills and control, though Revolut’s results come from one company and depend on its workload and infrastructure.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.