Revolut reduced its AI coding costs by 90% by replacing Claude with a self-hosted open model.

What it means

Revolut, Europe's largest private fintech company, replaced Anthropic's Claude Code with a self-hosted open Chinese AI model called GLM (from Zhipu AI / Z.ai), running on 56 Nvidia B200 GPUs. This cut their AI coding inference costs by roughly 90% for about 2,000 engineers, moving self-hosted models to the default and keeping commercial APIs as a backup. The disclosure signals a shift toward enterprise self-hosting of open models to reduce costs.

Accuracy & sourcesGrounded

The claim is reported consistently across multiple tech outlets (The Register, shattered.io, etc.), all citing Revolut's own announcement or blog post. While no direct primary-source link is given in the highlight, the cross-source agreement and specificity (56 B200 GPUs, GLM model, 90% saving) make the claim solid. The reporting sits close to Revolut's original disclosure, though we do not have the exact text.

How new