Revolut says self-hosted GLM models cut AI coding inference costs by roughly 90 percent

Revolut says it has cut 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 such as Claude. David Abitbol, ops and infrastructure lead at the British fintech, described the results at the AI Engineer conference in New York last month. The company also reported 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.

Cost pressure led to self-hosting

Revolut initially adopted GLM-5.1 before moving to GLM-5.3, which it now serves using 56 Nvidia B200 GPUs. Before the switch to GLM, Revolut had been experimenting with AI, deploying Gemini in early 2025 and Claude Code in early 2026. Abitbol said the company was highly skeptical about the cost of AI, with engineers spending $1,000 a month or $10,000 a month with not so much ROI, so it launched the tools with strict budget limits.

Because engineers only had access to a limited number of tokens, Revolut began investigating self-hosting its own models. At $0.30 per million input and output tokens, self-hosting GLM 5.1 cost about a tenth of using OpenAI's Codex or Anthropic's Claude Code. Abitbol said that when tokens are cheap, teams stop worrying about how many they use or what they use them for. The biggest bottleneck was not the code itself, he noted. The largest gains came from letting the AI models deal with debugging and Jira and Confluence tickets.

Serving GLM at scale

GLM 5.1 was developed by Chinese open weights model developer Z.ai. It has roughly 754 billion total parameters, with 40 billion activated for each token generated, and requires roughly 380 GB of GPU memory just to hold the weights plus another 6 to 26 GB per user to maintain context depending on precision. Revolut deployed 56 Nvidia B200 GPUs to serve GLM-5.3 at scale, equivalent in GPU count to seven eight-GPU DGX B200 systems.

Abitbol said the greatest performance gains did not come from adding GPUs, but from tuning the model's hyperparameters along with switching to NVFP4 quantization, implementing disaggregated prefill and decode, and tuning context lengths. He warned that a one-million-token context window does not necessarily make such a large window useful or efficient, and said 256,000 tokens is really what is needed. With the optimizations, Revolut achieved sub 3-second time to first token on context windows of around 256,000 tokens and generation speeds of around 200 tokens a second. Abitbol said engineers who sign in after 5 pm get closer to 800 tokens a second.

Model routing and security

GLM has not replaced GPT or Claude entirely. The Chinese model is text only, so it is not ideal for things such as UI work. Revolut used a model router to direct vision queries to models such as Gemini or GPT Luna while keeping code generation on GLM. Abitbol said open source is now a priority at Revolut and that its economics are favorable, with flat billing and no bill shock at the end of the month, while the frontier is the exception.

Abitbol acknowledged that Chinese open weights models may be a concern for privacy and security reasons, but he emphasized that the major concern is accessing them via APIs running out of China. When run on-prem or on regional cloud resources, there is nothing to worry about, he said.

Sources: The Register