Revolut Cuts Coding AI Costs With Self-Hosted GLM
Revolut made self-hosted GLM-5.3 the default for about 2,000 engineers on 56 B200 GPUs. Inference costs fell about 90%, in company figures.
Revolut cut the inference bill for AI coding tools by about 90% after making a self-hosted open-weights model the default for its engineers. Day-to-day coding now runs on infrastructure the company manages. Commercial models still handle some jobs. For a regulated fintech watching every token invoice, that is a sovereignty story as much as a cost story.
Ops and infrastructure lead David Abitbol described the shift at the AI Engineer conference in New York. Revolut rolled AI coding tools out to about 2,000 engineers. Coding time improved about 60%. Lead time improved about 70%. The cost and speed figures are company claims from that talk. They have not been independently verified.
What changed
Revolut first tried Gemini in early 2025 and Claude Code in early 2026. Some engineers burned through large monthly budgets with unclear return. The company put strict token limits on, then looked for a cheaper path. It adopted GLM-5.1 from Chinese lab Z.ai, later moved to GLM-5.3, and now serves that model on 56 Nvidia B200 GPUs.
Self-hosted GLM ran at about $0.30 per million input and output tokens, roughly a tenth of OpenAI's Codex or Anthropic's Claude Code on a token basis. Once tokens stopped feeling scarce, engineers used the models for debugging and for Jira and Confluence tickets, not only for the hardest coding tasks. The team also reduced the model's memory requirements and separated processing incoming prompts from generating answers. It capped useful context around 256,000 tokens rather than chasing a one-million-token window. The target is under three seconds before an answer starts and about 200 tokens per second during the workday.
What this does not prove
Inference cost is not total cost. The public figures do not establish the full cost of GPUs, power, staffing and the tools around the model. "Self-hosted" also does not mean Revolut owns every machine. The stack can run on rented GPUs in a regional cloud. That point is unsettled in the public record.
GLM is text-only in this setup, so vision work such as UI tasks still goes to models like Gemini or GPT Luna. There is no published third-party review of quality against Claude or Codex on Revolut's own codebase. Abitbol argued that Chinese open-weights models raise privacy questions mainly when teams call APIs that run in China. Running them on-premise or in a regional cloud presents a different risk. That is a company view, not a privacy guarantee.
Why it matters
Banks and fintechs face a choice about sending source code and tickets to external model providers. Revolut's answer is pragmatic: keep high-volume text work on a model you serve yourself, and treat frontier APIs as the exception. Cheaper tokens change behavior. People try routine chores they would ration on a metered plan.
The open question for every security team is simpler. If the model runs on your GPUs, who still sees the prompts, the tickets and the code that flow through the coding tools?
Sources
- British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted (The Register)
- Revolut moves to self-hosted GLM (Hacker News discussion)
Photo: "Datacenter Server Racks" by Carl Lender, via Wikimedia Commons, CC BY 2.0. Graded and overlaid by PSI.