Revolut has cut its AI coding costs by roughly 90%, a tenfold drop in under a year, by making self-hosted models the default instead of commercial services such as Claude, David Abitbol, who heads operations and infrastructure at the British fintech, said at the AI Engineer conference in New York, as covered by The Register in a November 3 piece. About 2,000 of its engineers use AI-assisted coding tools. The model at the center is GLM, an open-weight family built by Z.ai (also known as Zhipu) whose weights are downloadable so anyone with enough GPUs can run them in-house without sending code to an outside API. Revolut uses 56 Nvidia B200 GPUs in production, with self-hosting costing $0.30 per million tokens, input and output. Abitbol reported a 60% improvement in time spent writing code, a 70% improvement in lead time, time to first token under 3 seconds with 256,000 tokens of context, and generation speed of about 200 tokens per second, up to 800 after 5 p.m. Before GLM, Revolut had tried Gemini in early 2025 and Claude Code in early 2026, with engineers burning through $1,000 or $10,000 a month with no proportional payoff, prompting strict spending caps.