Revolut Cuts AI Coding Inference Costs by Ninety Percent With Self-Hosted GLM Models
Revolut cut its artificial intelligence coding inference costs by roughly 90 percent by moving its default infrastructure to self-hosted open-weight models from Chinese lab Z.ai instead of relying primarily on commercial services like Claude and OpenAI. David Abitbol, ops and infrastructure lead at the British fintech, described the transition at the AI Engineer conference in New York. The setup serves approximately 2,000 software engineers across the company.
The company initially experimented with Google’s Gemini in early 2025, which cut coding time by 20 percent and lead time by 10 percent. Revolut then deployed Claude Code in early 2026, but encountered high costs and strict budget limits that saw engineers spending between $1,000 and $10,000 a month with minimal initial return on investment. By April 2026, the fintech adopted self-hosted GLM 5.1 models before migrating to GLM 5.3, operating on 56 Nvidia B200 GPUs. The blended cost sits at $0.30 per million input and output tokens, amounting to roughly one-tenth of the price charged by OpenAI’s Codex or Anthropic’s Claude Code.
Cheaper tokens shift engineering focus from writing code to debugging
Cheaper tokens altered how engineers approached daily tasks, shifting focus from writing code to debugging and managing Jira and Confluence tickets. When tokens cost $0.30 per million, engineers stopped worrying about saving resources for the end of the quarter. According to Abitbol, writing code was never the primary engineering bottleneck. The rollout resulted in a 60 percent improvement in coding time and a 70 percent improvement in lead time from picking up a task to merging it.
Running GLM 5.3 at scale required 56 Nvidia B200 GPUs, equivalent to seven eight-GPU DGX B200 systems. Out-of-the-box inference performance was initially poor, but applying optimization, quantization, and inference techniques improved performance. Between April and August 2026, Revolut reduced its time to first token from over 10 seconds down to 1.68 seconds while handling a 76 percent increase in requests per second. Abitbol noted that while the models technically support a one-million-token context window, 256,000 tokens functions as the practical sweet spot for engineering tasks.
Revolut deploys open weights regionally to address security concerns
To address data privacy and security concerns regarding Chinese open-weight models, Revolut avoids accessing them via APIs running out of China. Instead, the company deploys the openly available weights on regional cloud resources or on-premise infrastructure within the United States, the European Union, Britain, or Israel, requiring zero data retention by default. Programming agents operate inside a sandboxed environment with strict approval rules and human oversight to prevent unauthorized commands and mitigate prompt injection risks.
GLM remains a text-only model, meaning it is not ideal for user interface design work. Revolut utilizes a model router to direct vision-based queries to multimodal alternatives like Gemini or GPT Luna while keeping code generation running on GLM. Abitbol emphasized that the firm now prioritizes open source for its favorable economics, replacing unpredictable month-end invoices with a flat billing structure.
Revolut details hardware and token costs for self-hosted AI setup
What hardware powers Revolut’s self-hosted AI setup?
Revolut serves GLM 5.3 using 56 Nvidia B200 GPUs, matching the capacity of seven eight-GPU DGX B200 systems.
How much does Revolut pay per token for its self-hosted models?
The company reports an average processing expense of $0.30 per million input and output tokens.
Which models did Revolut use before switching to GLM?
Revolut deployed Google’s Gemini in January 2025 and trialed Claude Code starting in February 2025 before launching it at scale in January 2026.
How are security and privacy managed for Chinese open-weight models?
Revolut runs model weights on infrastructure located in the United States, European Union, Britain, or Israel with default zero data retention, avoiding external API endpoints located in China.
More on Revolut
Worth a look