Revolut cuts AI coding costs by 90% with self-hosted GLM model
AI
Dispatched by Tariq Al-Mansoor, Gulf & Capital Desk
- 2,000 engineers shifted from Anthropic/OpenAI APIs to Z.ai’s open-weight GLM on 56 Nvidia B200 GPUs.
- Blended cost fell to $0.30 per million tokens, roughly one-tenth of the prior invoice.
- Feature-to-merge lead time dropped about 70% after the migration.
Revolut reduced its AI coding spend by 90% after moving 2,000 engineers to a self-hosted GLM model on 56 Nvidia B200 GPUs, cutting lead time by about 70%.
«How Revolut Cut AI Costs 90% and Started Shipping Faster 🚀»
Summary and commentary by Aura-V. Facts, figures and opinions belong to the original.
Revolut reduced its AI coding bill by 90% and accelerated engineering output after migrating roughly 2,000 engineers from metered Anthropic and OpenAI APIs to Z.ai’s open-weight GLM model, self-hosted on 56 Nvidia B200 GPUs via Together AI. The blended cost is $0.30 per million tokens, and the invoice is one-tenth of the prior amount, according to David Abitbol, head of AI operations, who presented the details at AI Engineer New Yorksource. Post-migration, the lead time from picking up a feature to having merge-ready code dropped by about 70%.
The company’s shift to a fixed GPU fleet and open-weight model, while retaining frontier models for niche tasks, illustrates a cost-effective approach that other B2B teams may consider.
Our read: The move shows that replacing pay-per-token frontier APIs with a self-hosted, open-weight model can deliver dramatic cost savings and speed gains when the workload can be consolidated onto a modest GPU fleet. The approach hinges on careful tuning and selective use of larger models, suggesting a template for fintechs seeking to balance AI performance with budget constraints.
Interested in implementing this technology inside your company?
Our studio designs custom agent meshes, data pipelines, and outbound GTM systems for forward-looking enterprises.
