Revolut Cuts AI Coding Costs by 90% via Self-Hosted Models
Fintech giant shifts from commercial APIs to self-hosted GLM models for internal engineering.
1Transition to Self-Hosted Models
Revolut successfully reduced AI coding inference costs by 90 percent by moving away from commercial services like Claude. By adopting self-hosted GLM-5.3 models, the company eliminated the financial unpredictability of token-based billing. This shift allowed 2,000 engineers to utilize AI tools without strict budget constraints, fundamentally changing their daily development workflows.
2Operational Efficiency Gains
Following the rollout of self-hosted AI tools, Revolut reported a 60 percent improvement in coding time and a 70 percent boost in lead time. These gains were primarily driven by offloading repetitive tasks like debugging and managing Jira or Confluence tickets to the AI, rather than just raw code generation.
3Hardware and Model Architecture
The company currently serves GLM-5.3 using 56 Nvidia B200 GPUs, equivalent to seven eight-GPU DGX B200 systems. The model, developed by Z.ai, features 754 billion total parameters. Managing this scale requires significant GPU memory, specifically 380 GB for weights and additional capacity per user to maintain context during active coding sessions.
4Optimization and Performance Tuning
Performance improvements stemmed from tuning hyperparameters and implementing NVFP4 quantization rather than just adding more hardware. Revolut also utilized disaggregated prefill and decode strategies alongside optimized context lengths. These technical refinements enabled the company to achieve sub-3-second time to first token, which is critical for maintaining high developer productivity levels.
5The Context Window Reality
Revolut identified that massive context windows are often inefficient for coding tasks. While the model supports one million tokens, the team found 256,000 tokens to be the practical sweet spot. They prioritize low latency over extreme context capacity, noting that exceeding five seconds for the first token significantly disrupts developer workflow and engagement.
6Strategic Model Routing
GLM does not handle all tasks, as it is a text-only model. Revolut employs a model router to direct vision-based queries to alternatives like Gemini or GPT Luna. This hybrid approach ensures that code generation remains cost-effective on GLM while specialized tasks receive the appropriate model support for UI-related development work.
7Security and Privacy Considerations
While using Chinese open weights models raised initial security questions, Revolut mitigated risks by hosting the models on-premise or within regional cloud resources. By avoiding API calls to servers located in China, the company maintains control over its data, ensuring that the infrastructure remains secure while benefiting from favorable open-source economics.
8Future of Internal AI Infrastructure
Open source is now a core priority for Revolut's engineering strategy. By moving to a flat billing model, the company has removed the risk of monthly bill shock associated with frontier models. This transition demonstrates that for large-scale enterprise operations, self-hosting provides both the necessary cost control and performance consistency required.