Revolut’s AI operations lead says 56 B200 GPUs serve GLM to roughly 2,000 engineers, while Claude Code and Codex remain available as exceptions. He reports a blended cost of $0.30 per million tokens and an invoice around one-tenth of earlier service spending. Tuning context length and KV cache helped handle more requests with less latency on the same GPUs. Related talk notes flag unclear comparison periods and inconsistencies in some productivity figures. These are internal observations, not an independently controlled study that includes operating staff costs.