Revolut moves 2000 engineers to self-hosted GLM: 90% cost cut

British fintech giant Revolut has successfully reduced its AI coding assistant expenses by approximately 90% over the past year. By abandoning third-party services like Claude Code and Codex in favor of the open-weight GLM model from China's Z.ai, the bank has moved to a self-hosted architecture on its own GPUs.

• By hosting GLM-5.3 on internal hardware, Revolut reduced the cost per million tokens to just $0.30.
• The infrastructure relies on 56 NVIDIA B200 accelerators serving 2000 engineers.
• Fine-tuning inference hyperparameters and capping context at 256k tokens ensured low-latency performance.
• Developer productivity increased by 60% in coding speed and 70% in lead time due to unrestricted parallel task execution.

Technical Architecture and Optimization

Instead of merely scaling hardware, Revolut's infrastructure team focused on inference optimization. Techniques such as NVFP4 quantization and separating prefill and decode pools were instrumental in achieving efficiency. Currently, seven standard DGX B200 systems are sufficient to handle the entire load for 2000 engineers.

David Abitbol, head of operations and infrastructure, argued that massive context windows are often overkill. By limiting the window to 256,000 tokens, the team maintains a time-to-first-token of under three seconds, which is more than enough to handle complex microservices while keeping the KV-cache affordable and efficient.

Behavioral Shifts in Development

The most significant qualitative change occurred when tokens stopped being a scarce resource. With unlimited access, engineers began automating peripheral tasks such as debugging, Jira ticket management, and Confluence updates. This shift effectively eliminated the bottleneck, allowing developers to focus on higher-level problem-solving.

Implications for the Tech Industry

The Revolut case study demonstrates that for organizations with hundreds of active users, self-hosting agentic models is a viable and highly profitable strategy. Beyond cost savings, it provides greater control over data privacy—a critical concern for fintech. As the quality of open-weight models continues to rival proprietary APIs, we expect more large-scale engineering teams to follow this path.