Revolut slashes AI coding costs by self-hosting Chinese open-weight models

Revolut has transitioned its software engineering workforce from commercial AI coding assistants to a self-hosted open-weight model, achieving a reported 90 percent reduction in costs. The fintech giant now runs GLM, developed by the Chinese lab Z.ai, on its own infrastructure for approximately 2,000 engineers. This strategic shift was detailed by David Abitbol, who oversees AI operations at Revolut, during a presentation at the AI Engineer conference in New York.

What happened

The migration to GLM represents the latest step in Revolut’s iterative approach to integrating AI into its development workflow. Initially, the company deployed Google’s Gemini across its engineering teams, which resulted in a 20 percent reduction in coding time and a 10 percent decrease in lead time. Lead time refers to the duration from when a task is picked up until it is merged into the codebase. Following this, Revolut experimented with Anthropic’s Claude Code, but strict per-engineer budget caps limited its utility. While the tool offered benefits, the high costs and usage restrictions meant the overall impact remained modest.

Since the spring of 2026, Revolut has operated GLM on its internal infrastructure, yielding dramatically different results. In a sample group of roughly 2,000 engineers, coding time dropped by 60 percent, while lead time decreased by 70 percent. Abitbol noted that the disproportionate drop in lead time suggests that coding itself is not the primary bottleneck. Engineers spend significant time searching through inboxes, Jira tickets, and Confluence pages. By removing token limits, Revolut allowed engineers to use AI freely for these search and synthesis tasks without rationing usage near the end of financial quarters.

How it works

Revolut runs the GLM model, which contains approximately 800 billion parameters, on a cluster of roughly 56 Nvidia B200 graphics processing units. This setup generates tokens at a cost of about 30 cents per million, replacing unpredictable monthly invoices with a flat operational cost. However, the initial out-of-the-box performance was poor. Through rigorous optimization, the team reduced the time to first token from over 10 seconds to 1.6 seconds, even as demand increased by 76 percent. Abitbol emphasized that latency above five seconds triggers immediate complaints from users.

The system employs specific constraints to maintain efficiency and compliance. The context window is capped at 256,000 tokens, a decision Abitbol defended by stating that excessive long-context capabilities are often unnecessary for coding tasks. Since GLM handles only text, a router directs image-related tasks to Google’s Gemini. While Claude Code and OpenAI’s Codex remain available for exceptional cases, they are no longer the default tools. This hybrid approach ensures that the most suitable model is used for each specific type of request.

Key details

• Revolut reports a 90 percent reduction in AI coding costs compared to previous spending on Anthropic and OpenAI services.
• The infrastructure relies on approximately 56 Nvidia B200 chips to serve 2,000 engineers.
• Coding time decreased by 60 percent, and engineering lead time dropped by 70 percent after switching to GLM.
• Data is processed exclusively in jurisdictions with equivalent data protection laws, including the US, EU, Britain, and Israel.
• The model weights are open-source, but inference runs on Western servers to avoid using Chinese endpoints.
• Token costs average 30 cents per million, with time-to-first-token optimized to 1.6 seconds.

Why it matters

For technical leaders, this case study highlights the potential economic advantages of self-hosting open-weight models over relying on proprietary APIs. The removal of token limits changes how engineers interact with AI tools, encouraging broader usage for non-coding tasks like information retrieval. This shift suggests that the value of AI in software development extends beyond code generation to include workflow acceleration. Companies that can manage the infrastructure complexity may find significant cost savings and productivity gains that closed-source providers cannot match due to their pricing structures.

However, the approach requires substantial upfront investment and specialized expertise. The hardware costs for 56 B200 chips could range from $1.7 million to $3.5 million, depending on whether the units are purchased or rented. Additionally, the rapid release cycle of new GPU architectures means these assets may depreciate quickly. Smaller teams without dedicated AI infrastructure staff may struggle to replicate these results, as the savings exclude cloud, staffing, and monitoring expenses. The success of this model depends heavily on continuous tuning and strict governance protocols.

What you can do

• Evaluate the total cost of ownership for self-hosting versus API usage, including hardware, power, and engineering time.
• Implement strict latency budgets, aiming for sub-two-second response times to maintain developer adoption.
• Use routers to direct specific task types, such as image processing, to specialized models while keeping text tasks on cheaper open-weight models.
• Cap context windows to prevent unnecessary resource consumption and reduce inference costs.
• Ensure data residency compliance by running open-weight models on infrastructure within approved legal jurisdictions.
• Maintain human-in-the-loop safeguards and sandbox environments to satisfy regulatory requirements for autonomous agents.