BREAKING: Revolut moved approximately 2,000 engineers onto a self-hosted GLM platform for coding, pushing Claude Code and Codex to an exception list. According to David Abitbol, who leads AI operations, the setup runs on 56 Nvidia B200s at a blended rate of roughly 30 cents per million tokens — approximately a tenth of what the company previously paid Anthropic and OpenAI. This shift has improved engineering lead time by 70%.
The critical insight is that gains emerged when token rationing stopped; engineers shifted from saving AI for "hard tickets" to using it universally. Frontier models became the exception, while open weights became the default. This transition presents a significant pricing challenge for other labs heading toward IPO.
The change was highlighted by @guilleflorvs and @Oliver Molander on X.
Source: Hacker News · Summarized by HeadlinesBriefing