Some people burn through $10,000 per month and consider Claude too expensive. This company switched to GLM: 56 B200 GPUs support 2,000 engineers, directly cutting AI costs by 90%.

Someone spends $10,000 a month on AI programming, and when the boss checks the bill: Is this money well spent? — This is exactly the awkward situation that UK fintech company Revolut once encountered.

To enable engineers to use AI, the company successively tried Gemini and Claude Code, but some engineers could burn through $1,000 or even $10,000 per month. However, after the money was spent, the efficiency improvement did not meet expectations. So Revolut made a bold decision: instead of relying mainly on commercial models for AI programming, it chose to deploy Zhipu's open-source model GLM on its own.

The result is quite staggering: in less than a year, the cost of AI programming inference has dropped by about 90%, the coding time of the company's approximately 2,000 engineers has been shortened by 60%, and the delivery cycle from receiving requirements to code merging has also been shortened by about 70%!

Claude Is Too Expensive! Some Engineers Burn $10,000 Per Month

At the very beginning, Revolut actually had no plan to train or deploy large models on its own.

In early 2025, the company started using Gemini to help engineers look up information and read technical documents; in early 2026, it introduced Claude Code, hoping to further improve development efficiency. But as the scale of use expanded, cost quickly became a problem.

David Abitbol, Head of Operations and Infrastructure at Revolut, revealed that some engineers spend as much as $10,000 per month on AI services, but the actual return on investment is not obvious. The company had to set strict Token budgets when promoting AI programming tools.

This in turn created a new problem: the more expensive the Token is, the more reluctant engineers are to use it. They will leave their limited quota for complex code generation tasks. For work such as debugging bugs, processing Jira tickets, and sorting out Confluence documents, even if AI can help, they will consider whether it is worth consuming Tokens.

Therefore, Revolut began to look for cheaper alternatives, and finally set its sights on Zhipu's GLM series models. The company initially adopted GLM-5.1, then upgraded to GLM-5.3, and set the self-hosted model as the main AI programming service.

The cost gap is very obvious: according to the data given by Abitbol, the combined cost of input and output Tokens for self-hosted GLM-5.1 is about $0.30 per million Tokens, which is only about one-tenth of the cost of commercial services such as OpenAI Codex or Anthropic Claude Code; the overall inference expenditure has also dropped to about one-tenth of the previous commercial service bill.

With cheaper Tokens, engineers' usage habits have also changed. Abitbol said that when Tokens are no longer so expensive, engineers no longer have to count every penny, do not have to save their quota until the end of the quarter, and do not have to call AI only for the most difficult tasks — in the past, engineers were thinking about how to save Tokens, but now they are thinking about what other work can be handed over to AI.

What Really Slows Down Development Progress May Not Be Coding At All

Revolut also shared a particularly noteworthy detail. They counted two indicators before and after the company's approximately 2,000 engineers used AI:

Coding time: the time engineers actually spend writing code, calculated in hours.

Delivery cycle: the time required from receiving a functional requirement to merging the code and preparing for deployment, calculated in days.

The internal data shown by Abitbol shows that after adopting GLM, coding time has been shortened by about 60%, and the delivery cycle has also been shortened by about 70%. However, he did not simply attribute these improvements to the fact that the model writes code faster: in his view, the things that engineers actually spend time on every day are far more than coding.

For example, troubleshooting bugs, replying to messages, processing Jira tickets, updating Confluence documents, and following up on various development processes. These tasks do not seem complicated, but they will gradually take up the time that engineers actually spend writing code. When the cost of using AI is low enough, engineers can also hand over these scattered tasks to agents, instead of remembering to use AI only when writing code.

This is why Revolut's delivery cycle is shortened even more than coding time: faster code writing only solves part of the problems in the development process; letting AI take over the trivial work outside of code creates the opportunity to further shorten the entire delivery cycle.

56 NVIDIA B200 GPUs Support 2,000 Engineers to Use AI

Letting nearly 2,000 engineers use the self-hosted large model at the same time is obviously not an easy task.

According to the introduction, the total parameter size of GLM-5.1 selected by Revolut is about 754 billion, and about 40 billion parameters are actually activated for each Token generated. Even with the Mixture of Experts (MoE) architecture, the model weights and context cache still require a large amount of GPU video memory.

According to Abitbol's estimation, storing the model weights alone requires about 380GB of video memory, and each user needs an additional 6GB to 26GB of video memory to maintain the context, depending on the precision settings. To support the large-scale inference of GLM-5.3, Revolut finally deployed 56 NVIDIA B200 GPUs.

However, Abitbol does not believe that performance improvement can only be achieved by buying more GPUs.

At the AI Engineer Conference held in New York last month, Abitbol showed a very interesting set of data: with the number of GPUs remaining unchanged, after the inference system is optimized, the number of requests processed per second has increased by 76%, and the first Token latency has dropped from 14.26 seconds during the migration period to 1.68 seconds after stable operation, a drop of about 88%.

Abitbol shared that the team mainly did the following things during this period:

Adjust model hyperparameters: Optimize model operation configuration according to real business load.

Adopt NVFP4 quantization: Use the low-precision computing scheme supported by Blackwell architecture GPUs to reduce video memory occupation and computing overhead.

Split Prefill and Decode: Separate input context processing from subsequent Token generation, and optimize for different computing stages.

Optimize cache and context configuration: Avoid reserving too much video memory for unnecessary ultra-long contexts.

Especially for the last point, Abitbol's point of view is quite straightforward.

Is the Million-Token Context a Gimmick? Revolut Thinks 256,000 Tokens Are Enough

Nowadays, large model manufacturers are working hard to expand the context window, and 1 million Tokens or even longer contexts have become a marketing selling point. But Abitbol does not agree with the blind pursuit of ultra-long context:

Long context is a scam. Even if you can provide a 1 million Token context window, 256,000 Tokens are actually enough.

Why? Because the longer the context, the larger the KV Cache (key-value cache) that the model needs to maintain, and the more GPU video memory it occupies. To put it simply, KV Cache can be understood as the model's temporary memory area, which is used to store the Key and Value tensors in the attention mechanism. In this way, when the model continues to generate content, it does not need to repeatedly calculate the information that has been processed before.

But if engineers actually only need a context of tens of thousands of Tokens, and the enterprise reserves resources required for a million Tokens for each request, it will obviously cause waste. Therefore, Revolut pays more attention to how long the context is needed for actual tasks and whether the cache can be effectively reused.

The team also found that tensor parallelism configuration is also critical: distributing model weights to 4 GPUs can hold the model, but may not leave enough space for KV Cache; using 8-way tensor parallelism can sometimes achieve better overall performance.

After a series of optimizations, Revolut has now controlled the first Token latency within 3 seconds, and the generation speed is maintained at about 200 Tokens per second. Abitbol even joked that if the first Token latency exceeds 5 seconds, his Slack inbox will start to be flooded with complaint messages.

As for the generation speed, Abitbol believes that 200 Tokens per second is already sufficient. Now engineers are more accustomed to starting multiple AI tasks at the same time, and coming back to check after the tasks are completed, instead of staring at the model output word by word — he also shared an interesting point: for engineers who log in after 5 p.m., the generation speed can even approach 800 Tokens per second.

Do Not Let GLM Take Over Everything, Visual Tasks Are Still Handed Over to GPT and Gemini

Although GLM-5.3 has become the main AI programming model, Revolut has no intention of completely saying goodbye to commercial models such as Claude and GPT — the reason is very simple: different models have their own areas of expertise.

For example, when engineers develop UI, they often need AI to read page screenshots, identify the interface layout, and then modify the code according to the visual effect. For these tasks, the performance of Gemini and GPT is more in line with Revolut's expectations. To this end, Revolut introduced Model Router, which automatically assigns models according to different tasks:

Tasks such as code generation and text processing are preferentially handed over to the self-hosted GLM.

Tasks such as image recognition and visual understanding are handed over to commercial models such as Gemini and GPT.

For this solution, Abitbol said: "Open source has now become a priority for Revolut, and from an economic point of view, our solution is very advantageous. Now the bill is fixed, we will not suddenly receive a huge bill at the end of the month. Cutting-edge models have become the exception instead."

However, Revolut's path cannot be copied by all companies. The hardware investment of 56 B200 GPUs, model deployment, inference optimization and daily operation and maintenance are not small expenses — whether the saved API bills can really cover the GPU and operation and maintenance costs is the most important part of the account that needs to be calculated clearly.

Reference link: https://www.theregister.com/ai-and-ml/2026/11/03/british-fintech-revolut-slashes-ai-coding-costs-10x-by-cutting-out-claude-and-going-self-hosted/5307345

This article is from the WeChat Official Account "CSDN", compiled by Zheng Liyuan, and published with authorization from 36Kr.