Revolut cuts AI coding inference costs roughly 90% by making self-hosted GLM models its default

Revolut says it cut its AI coding inference costs by roughly 90 percent in less than a year by making self-hosted models its default instead of relying primarily on commercial services like Claude. The British fintech now serves GLM-5.3, from Chinese open-weights developer Z.ai, on 56 Nvidia B200 GPUs, according to a report by Tobias Mann at The Register.

David Abitbol, ops and infrastructure lead at Revolut, described the shift at the AI Engineer conference in New York last month. He said Revolut saw a 60 percent improvement in coding time and a 70 percent improvement in lead time after rolling out AI coding tools to roughly 2,000 engineers.

From Gemini and Claude Code to GLM

Revolut deployed Gemini in early 2025 and Claude Code in early 2026. Abitbol said the company was skeptical about AI costs: "We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI." Revolut launched with strict budget limits, which capped engineers' token access and prompted the company to look at self-hosting.

Revolut adopted GLM-5.1 first, then moved to GLM-5.3. At $0.30 per million input/output tokens, self-hosting GLM 5.1 cost about a tenth of using OpenAI's Codex or Anthropic's Claude Code, according to the report.

Model and hardware specifics

The Register reports that GLM 5.1 has roughly 754 billion total parameters, with 40 billion activated per generated token. It needs roughly 380 GB of GPU memory for weights alone, plus 6-26 GB per user for context depending on precision. The 56 B200 GPUs match the GPU count of seven eight-GPU DGX B200 systems.

Abitbol said the largest performance gains came from tuning rather than adding GPUs:

• Hyperparameter tuning
• NVFP4 quantization
• Disaggregated prefill and decode
• Tuned context lengths

With these changes, Revolut reports sub 3-second time to first token on contexts of around 256,000 tokens, and generation of around 200 tokens per second. Abitbol said time to first token matters more than throughput: "As soon as you exceed 5 seconds, my Slack inbox starts to blow up." Engineers who sign in after 5 pm see closer to 800 tokens per second, he said.

On context length, Abitbol said: "Long context is a scam... Even though you can serve one million tokens, 256 [thousand] is really what you need."

Where frontier models remain

GLM is text only, so Revolut uses a model router to send vision queries to models such as Gemini or GPT Luna, while code generation stays on GLM. "The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception," Abitbol said.

He acknowledged that Chinese open-weights models may raise privacy and security concerns. In his view, the main risk is accessing them through APIs running in China, and running them on-prem or on regional cloud resources removes the concern.

What this means

Revolut's account ties the savings to usage behavior as well as unit price. Abitbol said that once tokens were cheap, engineers stopped rationing them and used models for debugging and for Jira and Confluence tickets, which he described as the biggest gains, ahead of raw coding speed. His point is that the bottleneck was not the code itself.

The deployment also shows what self-hosting a model of this size involves: a 56-GPU B200 cluster and serving-stack work, which Abitbol credits for more of the performance gain than hardware count. The figures come from Revolut's own executive and have not been independently verified. The architecture is a hybrid: a self-hosted default for text and code, with commercial models reserved for vision and cases where the frontier is needed.