How Revolut Cut Its AI Coding Bill 90% by Self-Hosting an Open-Weight Model
How Revolut Cut Its AI Coding Bill 90% by Self-Hosting an Open-Weight Model When British fintech Revolut rolled out AI coding assistants to roughly 2,000 engineers, the bill did what AI bills always do: it exploded. Some engineers were spending $1,000 — even $10,000 — a month on AI tokens with little measurable return. So Revolut did something almost no company of its size has tried: it ditched…
British fintech company Revolut recently managed to slash its AI coding costs by nearly 90% after self-hosting an open-weight model. This decision came after the AI assistants caused costs to skyrocket, with some engineers spending up to $10,000 a month on AI tokens with limited results. Revolut made the unprecedented move of moving its engineers onto a self-hosted, open-weight model, resulting in a significant reduction in AI coding inference costs and speedier coding times.
David Abitbol, Revolut's AI operations and infrastructure manager, detailed this transformation at the AI Engineer conference in New York. Early in 2025, Revolut adopted Gemini, which offered a 20% speedup in coding time. In early 2026, Claude Code was introduced, but its impact was modest due to strict per-engineer budgets. Then, in April 2026, Revolut began self-hosting open-weight models, starting with GLM-5.1 from Chinese lab Z.ai, without any token limits.
Later, they upgraded to GLM-5.3, a 754B model with approximately 40B active parameters per token, which was served from around 56 Nvidia B200 GPUs.
The results were impressive: Revolut reported a 90% reduction in AI coding inference costs, with the blended serving cost amounting to $0.30 per million tokens. The company also experienced a 60% improvement in coding time and a 70% improvement in lead time, with over 2,000 engineers using the system. Additionally, Revolut observed a 76% increase in requests per second and an 88% reduction in time-to-first-token, improving from 14.3 seconds to 1.7 seconds.
One of the key takeaways from Revolut's success story is that cheap, unlimited tokens are more effective than smarter AI models in accelerating coding. By removing token limits, engineers were able to focus on the task at hand without worrying about the bill that comes with every API call. This shift in behavior ultimately led to the significant speed gains observed.
Interestingly, Revolut also decided to self-host the weights on US-based infrastructure, which helped them comply with regulatory requirements and keep data within their legal perimeter. However, it's important to note that Revolut's figures are their own, and the effect of removing token limits should not be isolated from the model change.
Additionally, the 30-cent cost per million tokens calculation does not include cloud spend, staffing, or observability costs, and the 56 B200 GPUs represent a substantial infrastructure investment that a smaller startup might not be able to replicate.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
