Revolut Cuts AI Coding Costs by 90 Percent by Moving to Self-Hosted Models

British fintech company Revolut has reduced its artificial intelligence coding expenses by roughly 90 percent over the past year by switching from commercial services like Claude to self-hosted models.

David Abitbol, operations and infrastructure lead at Revolut, shared the details at the AI Engineer conference in New York, according to the publication. The firm initially adopted GLM-5.1 before upgrading to GLM-5.3, serving the software across 56 Nvidia B200 graphics processing units.

Before making the switch, Revolut tested Gemini in early 2025 and Claude Code in early 2026. Abitbol explained that the company worried about high expenses, noting that engineers were spending between $1,000 and $10,000 a month with limited return on investment.

To control spending, Revolut imposed strict budget limits that restricted engineers to a small number of tokens. That limitation prompted the company to explore self-hosting. Using self-hosted GLM-5.1 cost $0.30 per million input and output tokens, which was about one-tenth the price of commercial tools like OpenAI's Codex or Anthropic's Claude Code.

With cheaper tokens available, engineers stopped worrying about how many tokens they used. Abitbol noted that the biggest productivity gains came from letting AI models handle routine tasks like debugging, alongside Jira and Confluence tickets, rather than just writing code.

Developed by Chinese open weights model creator Z.ai, GLM-5.1 is a large system with roughly 754 billion total parameters. Running the model requires about 380 gigabytes of memory just to hold the weights, plus another 6 to 26 gigabytes per user for context.

To operate GLM-5.3 efficiently at scale, Revolut set up 56 Nvidia B200 GPUs. Abitbol explained that performance improved most through tuning hyperparameters, applying NVFP4 quantization, and adjusting context lengths rather than simply adding more hardware.

The company found that supporting massive context windows is not always practical. While the model can handle one million tokens, Abitbol called very long contexts a scam and stated that 256,000 tokens is the ideal size.

These adjustments allowed Revolut to achieve a response time of under three seconds for the first token and a generation speed of roughly 200 tokens per second. Abitbol noted that engineers working after 5 p.m. can see speeds closer to 800 tokens per second.

Because GLM is a text-only model, it cannot handle visual tasks like user interface design. Revolut uses a model router to direct vision tasks to tools like Gemini or GPT Luna while keeping code generation on GLM.

While acknowledging that some organizations worry about the security of Chinese open-weights models, Abitbol said those risks usually stem from using APIs based in China. He emphasized that running the models on-premise or through regional cloud resources removes those security concerns.

Summary

By shifting from expensive commercial AI services to self-hosted open models, Revolut lowered its coding expenses and removed usage limits for its engineering staff. The company's approach shows how large firms can optimize hardware and hyperparameters to run massive AI systems locally at a fraction of the cost.