Revolut cut AI coding costs 90%. The catch is GPU infrastructure

The cost signal. Revolut reports cutting AI coding spend to about a tenth of its previous bill with Anthropic and OpenAI by self-hosting an open-weight GLM model on 56 B200 GPUs. The service covers 2,000 engineers at a blended 30 cents per million tokens, and the invoice is flat. That is the strongest number in the talk, and it explains why Revolut has made open source a priority.

The behavior change. Abitbol traces the impact through two lines: coding time and lead time. Gemini gave 10 percent lead-time improvement and 20 percent coding-time improvement. Self-hosted GLM claims roughly 60 percent coding-time improvement and 70 percent lead-time improvement. His explanation is behavioral: unlimited cheap tokens let engineers stop rationing AI for hard tasks and use it for Jira, Confluence, and other context switching, which is where lead time was actually spent.

The infrastructure reality. The cost saving is not free. Revolut runs 56 B200s and had to tune aggressively: time to first token went from over 10 seconds to around 1.6 seconds, while request throughput more than doubled on the same hardware. Abitbol says anything above 5 seconds triggers Slack complaints, and 3 seconds is the target. He treats long context as overrated, 256K tokens is enough, and cache-hit rates for coding should stay above 90 percent.

The compliance argument. Revolut is an EU-regulated bank, so the talk spends time on the stigma around Chinese open-weight models. The position is that using weights on US-based infrastructure avoids data-sharing concerns, as long as data stays within the US, EU, UK, or Israel, and providers offer zero data retention. The bigger control is the harness: coding agents stay human-in-the-loop inside sandboxes with operation-level approval rules, which satisfies regulators that this is not autonomous.

The missing evidence. These are strong claims from one practitioner, and the talk is short on methodology. The before-and-after comparison is not a controlled study: model, budget limits, and engineering habits changed at the same time. The 90 percent cost cut depends on which hosted tier and usage pattern you compare against, and the speaker still uses frontier hosted models for hard tasks, so self-hosting is an exception rather than a complete replacement. That nuance matters for anyone building a decision on this.

The leadership takeaway. The practical shift is organizational, not just technical. AI coding cost stops being a per-seat budget line and becomes a GPU capacity decision that needs an infrastructure team, observability, and compliance review. Engineering leaders should ask what their per-engineer token spend really is before assuming a hosted frontier model is the only option. For a large enough org, self-hosting deserves a bounded pilot, not a default.