Revolut runs GLM itself: what 0.30 dollars per million tokens means

At the AI Engineer conference in New York on October 13, 2026, David Abitbol, who runs Ops and Infrastructure at Revolut, described how the bank cut its AI costs. Around 2,000 engineers there now work with self-hosted GLM models from Z.ai by default. The bill for Anthropic and OpenAI fell to about a tenth, he said. The Register picked up the talk on November 3.

Anyone who wants to base their own decision on figures like these should first see which of them are backed up. We read the sources and recounted.

Self-hosting an AI model means you download open weights, run them on your own or rented GPUs, and pay for the hardware and its operation instead of a price per token. The unit price per million tokens then equals total cost divided by the tokens actually processed, and it falls as utilization rises. A vendor price stays the same whether you use a lot or a little.

What this post rests on

Our main source is the talk itself: Shipping like a startup, regulated like a bank by David Abitbol, October 13, 2026, main stage. The conference publishes a summary and a machine transcript. The video is the day's livestream on YouTube (segment 5:36:54 to 5:48:28). We read the transcript but did not watch the full video, so we report figures from the talk as "per the talk". On top of that comes The Register's report of November 3, 2026. Where the two differ, we say so.

What Revolut reports

• Gemini arrived in January 2025, Claude Code followed later. Per the talk, some engineers spent 1,000 to 10,000 dollars a month with little return, so strict budgets applied.
• Since April 2026 Revolut has run open models itself: GLM-5.1 first, GLM-5.3 later.
• Serving roughly 2,000 engineers takes about 56 Nvidia B200s for GLM-5.3, per the talk. The Register converts that to seven DGX B200 systems with eight cards each.
• Blended price when self-hosting: about 0.30 dollars per million tokens. The invoice is a tenth of what used to go to Anthropic and OpenAI.
• For the switch to GLM-5.1, one slide shows 60 percent better coding time and 70 percent better lead time.
• Frontier models stay as the exception. A router sends images to other models because GLM handles text only.

The conference page notes that the internal comparison does not separate the model change from other causes. Abitbol reads the gap between the two values this way: the bottleneck is not only writing code but also the inbox, Jira and Confluence, and with a generous token budget people delegate that work more readily.

The figures at a glance

The top four rows are the operator's own statements. We have not re-measured them.

Recounting the 0.30 dollars

0.30 dollars per million tokens is not a number you can buy. It is total cost divided by tokens processed, and Revolut names neither side. What we can do is check the order of magnitude.

As an assumption, not as evidence, we set 10,000 dollars of monthly rent per B200. Then 56 cards cost 560,000 dollars a month. To land at 0.30 dollars per million tokens, about 1.87 trillion tokens would have to pass through in that month, roughly 930 million per head across 2,000 engineers. If less flows through, the unit price rises in the same proportion. Assume a lower rent and the required volume drops by the same factor. The calculation makes one point: the 0.30 dollars stand or fall with utilization. A half-empty fleet is expensive per token, a full one is cheap. With a vendor price you pay only for what you use.

Weights you download and weights in production

We summed the files on Hugging Face through the API on November 4, 2026:

The config.json of GLM-5.3 lists 78 layers and a context of 1,048,576 tokens. The Register gives 380 GB for GLM 5.1 (roughly 754 billion parameters). That does not match the GLM-5.3 FP8 checkpoint but does match 4-bit: 753.33 billion parameters at half a byte each come to about 377 GB. In the talk, Abitbol recommends FP4 on Blackwell cards, and The Register names NVFP4 quantization for Revolut. So Revolut likely does not serve the 756 GB you download from Hugging Face; neither source gives the exact configuration.

For our hardware this changes nothing. Our four RTX 5090 offer 127.4 GiB of video memory together. Even the 4-bit variant needs about 350 GiB on paper. The Flash variant at 2 to 3 bits would come to roughly 75 to 112 GiB, which is 80 to 120 GB (our own arithmetic from 321.3 billion parameters, without overhead). Whether quality holds up at that compression, we have not tested.

What applies to us

Our lane runs on four RTX 5090 behind a LiteLLM gateway. The KV-cache pool in the running vLLM measured 1,810,884 tokens on October 3, 2026 (NVFP4 KV, measured on host .180.3). On our box the weights are the scarce resource, and the GLM family is no different.

We take two things from the Revolut report. A budget cap can suppress value nobody sees, because people count tokens and delegate less. Revolut reports that as its own experience; the talk does not prove it. Also, a regulated financial company has moved open weights into regular operation. That is an argument in any internal discussion about self-hosting. The conference page notes that his comments on regulatory classification and acceptable hosting locations describe Revolut's circumstances and are not a general determination for other organizations.

For the model, its MIT license and the release of GLM 5.2, see the AI Radar on z.ai GLM 5.2. This post takes a different angle: it tests what the Revolut figure says about self-hosting. How token prices and consumption add up is broken down in What an AI model really costs. More posts on running AI in a company are collected in the hub AI in production. If you are planning a model in your own house, our offer is described under Corporate LLM in your own house.

What we don't know

The composition of the 0.30 dollars is not documented. We don't know whether power, staff and storage are included pro rata, and we don't know the utilization. The Register ties the 0.30 dollars to GLM-5.1, while the talk says GLM in general. Which version the figure covers stays open.

We set the B200 rent as a calculation assumption. It is neither a documented market price nor a quote. For your own calculation, use the quote you would actually get.

The 60 and 70 percent are self-reported on a slide, without independent measurement and without a definition of coding time. They work as a tendency and not as a target for your own project.

Who should do what now

A team below a hundred developers cannot derive a threshold from the Revolut figures. The useful step is your own calculation: measure token consumption over the last months, check whether cards are already in the building, and set expected utilization honestly. The chart shows what Revolut reached with the same number of cards through tuning. The flow below it is our order of checks.

Sources

• AI Engineer NYC 2026: Revolut talk with transcript
• YouTube: AI Engineer NYC livestream, Revolut segment from 5:36:54
• The Register: British fintech Revolut slashes AI coding costs 10x
• Hugging Face: zai-org/GLM-5.3
• Hugging Face: GLM-5.3 config.json
• Hugging Face: zai-org/GLM-5.3-Flash
• Our measurement: local model speed recounted