Prompt & Done · GLMSWAP
Revolut's talk on moving AI coding to a self-hosted open model
Revolut says it cut its AI coding costs about 90% by moving around 2,000 engineers from metered frontier APIs to GLM, an open-weights model. These are the primary sources, and what the talk does and does not say.
Checked 5 Nov 2026. Figures are Revolut's own, from a conference talk; they are not independently audited.
1 The talk
"How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models", David Abitbol (Ops & Infrastructure Leader, Revolut), AI Engineer New York, uploaded 29 Oct 2026 (about 12 minutes).
What he says on the slides: 2,000 engineers served with GLM 5.3 on 56 × B200 GPUs; under 3 s time to first token; $0.30 per million tokens in/out (cache is free), about 1/10th of the cost; predictable invoice, no bill shock; frontier models kept as the exception for tasks where GPT or Claude do best.
2 The model: GLM on Hugging Face
The weights are public and not gated. GLM-5 and GLM-5.1 are MIT-licensed. GLM-5.3 (the version on Revolut's slide) uses Z.AI's own licence: free to use, modify and sell, with an extra security review only for companies that run a model-as-a-service business above USD 10 billion revenue. Read the LICENSE file before you ship anything.
3 Read this before copying the idea
$0.30 per million tokens is Revolut's blended token cost on dedicated hardware. A StartupHub write-up says the talk leaves cloud spend, staffing and observability outside that figure. The speaker also describes a lot of tuning: the same 56 GPUs went from over 10 s to about 1.6 s time to first token between April and August. The slide mentions Together as the GPU provider.
startuphub.ai: Revolut says self-hosted AI cut coding costs 90%
The kart race in the video is an illustration. Voice: Built with Fish Audio.