Revolut says self-hosted GLM cut its AI coding inference costs about 90%

Revolut says it has cut its AI coding inference costs by roughly 90% in under a year by making self-hosted open-weight models the default for engineers instead of relying mainly on commercial services such as Claude. The claim comes from David Abitbol, the British fintech's ops and infrastructure lead, who spoke at the AI Engineer conference in New York last month. The Register reported it on 3 November 2026. The figures are his, and the report does not show an audit.

What Revolut says it did

According to The Register, Revolut deployed Gemini in early 2025 and Claude Code in early 2026, then moved coders to GLM-5.1 and later GLM-5.3, an open-weight model from the Chinese developer Z.ai. Abitbol said engineers on commercial tools were spending $1,000 a month and sometimes $10,000 a month 'with not so much ROI', so Revolut launched with strict budget limits. Those limits left engineers with few tokens, which led the company to look at self-hosting.

The Register reports that self-hosting GLM-5.1 at $0.30 per million input and output tokens cost about a tenth of using OpenAI's Codex or Anthropic's Claude Code. Abitbol said Revolut saw a 60% improvement in coding time and a 70% improvement in lead time after rolling AI coding tools out to roughly 2,000 engineers. The report does not say how those were measured.

The infrastructure behind it

GLM-5.3 runs on 56 Nvidia B200 GPUs, which The Register equates to seven eight-GPU DGX B200 systems. The Register puts GLM-5.1 at roughly 754 billion parameters, 40 billion active per token, needing about 380 GB of GPU memory for weights alone. Abitbol said the biggest gains came from tuning hyperparameters, switching to NVFP4 quantization, separating prefill from decode and tuning context lengths, not from adding GPUs.

On context, Abitbol said 'long context is a scam' and that although a million tokens can be served, 256,000 is what teams need. He said Revolut reaches under 3 seconds to first token at about 256,000 tokens and about 200 tokens a second, and that once first-token time passes 5 seconds 'my Slack inbox starts to blow up'. Engineers who sign in after 5 pm, he said, see closer to 800 tokens a second.

What changed in how people work

Abitbol said that flat billing changed behaviour: when tokens are cheap, engineers stop rationing them and hand the model debugging and Jira and Confluence tickets, which he described as the biggest gains. He said the bottleneck was not the code itself. He also said GLM has not replaced the commercial models: it is text only, so a router sends vision queries to Gemini or GPT Luna, and he described the frontier models as 'the exception'.

What the report leaves out

The Register's piece gives no total cost for the 56 GPUs, power, hosting or the engineers who tuned and run the cluster, so '90%' compares inference prices, not total cost of ownership. It does not say whether output quality matched Claude Code on Revolut's tasks, or whether the 60% and 70% gains hold against a control group. Abitbol acknowledged that Chinese open-weight models can raise privacy and security concerns, and said the real risk is reaching them through APIs hosted in China, not running them on-premises or on regional cloud.

What to do with it

If your coding-assistant bill is the CFO's question, build the comparison on your own usage: measure tokens per engineer per month, then price a self-hosted cluster including GPUs, an owner and the tuning time. Test whether the cheaper model passes your own tasks before moving a default, and decide in advance which work, such as vision, stays on a commercial model.