Home / Technology / Revolut Slashes AI Costs by 90% with Open Models
Revolut Slashes AI Costs by 90% with Open Models
4 Nov
Summary
- Revolut reduced AI inference costs by 90% by switching from Claude.
- Self-hosted Z.ai models improved lead time and coding productivity.
- The system reportedly uses 56 Nvidia B200 GPUs with substantial memory.
Revolut has significantly reduced its artificial intelligence expenses, achieving up to a 90% cut in inference costs over the past year. The company has transitioned from using Anthropic's Claude to deploying open-weight models developed by Z.ai, a Chinese firm. This strategic move allows Revolut developers to utilize AI more freely without token limitations, enhancing productivity.
This change has led to substantial gains, with reported improvements of 70% in lead time productivity and 60% in coding time productivity. Revolut's Infrastructure Lead, David Abitbol, praised Z.ai's GLM models for their data privacy and auditability, noting that the move to open source is a priority and has resulted in favorable, predictable economics with no end-of-month "bill shock."
The new self-hosted infrastructure relies on a substantial hardware investment, including 56 Nvidia B200 GPUs, requiring over 380GB of memory per model, plus additional memory for user context. This demonstrates Revolut's commitment to managing its own AI capacity and controlling costs more effectively, moving away from the unpredictability often associated with commercial AI solutions.
