Revolut Moved 2,000 Engineers to Self-Hosted Chinese Open-Weights GLM and Cut Its AI Coding Bill to a Tenth

Revolut, the European fintech giant, has quietly done what many engineering leaders talk about and few finish: it replaced a commercial AI coding stack with self-hosted open-weights models as the default for its roughly 2,000 software engineers. David Abitbol, the company's head of AI infrastructure, laid out the details at the AI Engineer conference, in an account reported by The Register.

The default model is now Z.ai's GLM, deployed by Revolut itself on a cluster of about 56 Nvidia B200 GPUs, at a blended cost Abitbol put at around $0.30 per million tokens. The commercial services are not gone: Claude Code and OpenAI's Codex remain available for the hardest problems, and image-related work goes to Gemini. Coding agents run inside sandboxes governed by approval rules.

The economics are the headline. Revolut's bill for AI coding is now roughly a tenth of what it paid Anthropic and OpenAI, per Abitbol. Two caveats belong next to that number: it is company-reported and unaudited, and the often-quoted 90 percent saving excludes cloud, headcount and monitoring costs, which the open-source path does not eliminate.

The internal metrics Abitbol showed are similarly directional rather than rigorous: coding-time indicators improved about 60 percent and lead time from task assignment to merged, deployable code improved about 70 percent across the roughly 2,000-engineer sample. Treat those as one company's self-measurement, not a benchmark.

The more interesting confession is what it took to get there. Out of the box, Abitbol said, GLM's inference performance was poor for their workload; the team spent heavily on tuning and pushed first-token latency down from 14 seconds to 1.7 seconds. The timeline is also instructive: Revolut adopted Claude Code early this year with a $300 monthly cap per engineer, began downloading open-weights models in April, and moved through GLM-5.1 before landing on GLM-5.3. Once models were self-hosted, individual token budgets stopped being a constraint.

There is a risk ledger too. Open-weights Chinese models are now embedded in real financial infrastructure, and quality control is on the deployer: Cloudflare separately confirmed this year that some GLM-5.3 responses returned fragments of other customers' data through its own deployment. Self-hosting moves the trust boundary in-house, which is the point - but it also means every regression is yours to catch.

The pattern Revolut crystallizes is a two-tier market: open-weights models absorb the bulk of routine engineering at a fraction of the cost, while frontier APIs become an escalation tier for the tasks that actually need them. If that division holds, the pricing power of commercial coding assistants erodes from the bottom, and the moat for frontier labs moves further up the difficulty curve.