Revolut, a European digital bank with about 80 million customers in 45 markets, described its AI coding setup in an October 2026 talk by David Abitbol, who leads AI operations at the company.
Background. Revolut tracks two engineering metrics across a sample of about 2,000 engineers: lead time, from picking up a feature until it is merged and ready to deploy, and coding time. Its first company-wide AI tool, Gemini, introduced in January 2025, improved lead time by about 10% and coding time by about 20%. Claude Code was then rolled out with strict budget limits because some engineers were spending $1,000 to $10,000 a month with little measurable return, and the speaker said it had little impact on the metrics at the time. The goal became giving engineers unlimited tokens at a fraction of the cost.
Self-hosting. In April 2026 Revolut began self-hosting an open-weight GLM model, which it favours for coding. Serving about 2,000 engineers takes roughly 56 NVIDIA B200 GPUs for a model of about 800 billion parameters. The operating targets are a time to first token under 3 seconds (complaints start once it passes 5 seconds) and about 200 tokens per second of interactive throughput; engineers typically run several tasks in parallel rather than watching the token stream, and off-peak throughput reaches about 800 tokens per second. After the switch, the speaker reported coding time down about 60% and lead time down about 70%, and concluded that coding itself is not the main bottleneck: unmetered tokens let engineers also use the agent for Jira, Confluence and inbox work. The blended cost is about $0.30 per million tokens, and the invoice is about one tenth of what Revolut was paying Anthropic and OpenAI, with flat billing. Frontier models from those labs remain in use as the exception.
Tuning. Between April and August 2026, on the same 56 GPUs, request load rose from about 3 to over 7 requests per second (+76%) while time to first token fell from over 10 seconds to about 1.6 seconds, an 88% improvement. The parameters the speaker named: serve a 256K context instead of the maximum 1M; on Blackwell GPUs use 4-bit quantisation (NVFP4), which he said cuts GPU footprint by about 40% versus FP8 without measurable precision loss; consider tensor parallelism of 8 instead of 4 even where the model fits on 4 GPUs, to leave enough KV cache for the workload; keep the prompt cache hit rate above 90% for coding tasks; and tune speculative decoding and prefill settings.
Production setup. Because the GLM model is text-only, a local router sends image-recognition requests to Gemini while the GLM model handles the main task; a smaller GLM vision model is used by the UI design team. The endpoints expose an OpenAI-compatible API, including the Responses API, so they plug into coding harnesses such as OpenCode and Codex.
Compliance. Revolut runs only the model weights on its own GPU capacity, not the model vendor's hosted endpoint in China, and keeps data in jurisdictions covered by data-equivalency frameworks (the speaker named the US, EU, UK and Israel) with zero data retention at its GPU providers. As an EU-regulated bank, it treats the coding agents as human-in-the-loop tools rather than autonomous agents, and enforces this with a sandbox whose rules allow or block specific operations; the sandbox also guards against misuse by insiders and prompt injection. OpenTelemetry logging provides the audit trail for incident investigation and usage analytics.
All performance and cost figures are Revolut's own, reported in the talk without a stated measurement method.