Revolut self-hosts GLM-5.3 for 2,000 engineers: the GPUs, the settings and what the $0.30 really means

Revolut is a European digital bank with, in the speaker's words, 80 million customers in 45 markets. On 29 October 2026 the AI Engineer conference channel published an 11-minute talk by David Abitbol, who leads AI operations there, called "How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models". It reached Hacker News on 3 November with 193 points.

His story runs like this. Revolut first gave engineers an AI model with access to company data in January 2025 (Google's Gemini), then rolled out Claude Code, Anthropic's coding agent, with strict budget limits, because some engineers were spending $1,000 or $10,000 a month without a clear return. To give engineers unlimited tokens (the small pieces of text models read and write) at a fraction of the cost, Revolut began running open models on its own GPUs in April 2026, starting with GLM-5.1 from Z.ai, and now runs GLM-5.3.

On his chart of 2,000 engineers, the self-hosted model came with a 60% improvement in coding time and 70% in lead time (from picking up a feature to having it merged), against the months before it. His reading: the bigger gain was in lead time, so the slow part of engineering is not only writing code, and engineers with unlimited tokens use the model for email, Jira and documentation too. These are Revolut's internal numbers, not an outside study.

Revolut has not dropped the closed models. Claude Code, Codex and Gemini are still available as exceptions, and GLM-5.3 is the default. GLM-5.3 reads text only. His advice is to put a router in front that sends image questions to a model that reads images, such as Gemini.

In short: the GPU setup Revolut describes checks out on paper, and the saving is believable, but its $0.30 cannot be compared directly with a list price. The rest of this page shows why.

GLM-5.3 is an open-weight model from Z.ai (Zhipu AI) in Beijing, released on Hugging Face on 25 August 2026. Open weights means anyone can download the model files and run them. Z.ai says GLM-5.3 uses the same base model as GLM-5.2, and that all of its gains came from post-training, the later rounds of training that teach a model to code, follow instructions and use tools. On Z.ai's own table, it scores 66.9 on DeepSWE (fixing real software issues) and 28.3 on Terminal Bench 3.0 (tasks in a command line), against 69.7 and 33.7 for Anthropic's Fable 5.

Its licence lets you use, change and sell it freely. The one condition: a company that sells model access as a service and earns over $10 billion a year must pass a Z.ai security review first. Revolut runs the model for its own staff, which is not selling it as a service.

The Hugging Face API counts 753 billion parameters (the numbers a model learns); the talk rounds this to 800 billion. It is a mixture of experts: each of its 78 layers, apart from the first 3, holds 256 small sub-networks called experts, and a small gating network picks 8 of them, plus 1 shared expert, for each token. That keeps each token fast, but every expert must still sit in GPU memory.

For attention, the step where each token looks back at the earlier ones, it keeps a compressed record of every earlier token, 576 numbers per layer, and a small scorer called the indexer picks the 2,048 most useful earlier tokens to read in full. The indexer runs on 21 of the 78 layers; the others reuse the last choice. That record is the KV cache, the attention memory kept for each conversation, and it is what decides how many conversations a server can hold at once.

Read next: Reading a model name: what 27B, FP8, A3B and Instruct mean

A B200 GPU has 180 GB of memory (NVIDIA's DGX B200 has 1,440 GB across 8). Tensor parallelism cuts the model into slices, one per GPU, and all the GPUs in a copy work on every token together. So the weights per GPU are the file size divided by the number of GPUs: 116 GB each for the 4-bit copy on 4 GPUs, 58 GB each on 8.

Now the attention cache. From GLM-5.3's config file, in 16-bit numbers, each token needs 96,384 bytes: 89,856 for the compressed attention record across 78 layers, 5,376 for the indexer's keys on the 21 layers that run their own indexer, and 1,152 for the extra layer that guesses tokens ahead. One 256K conversation therefore needs about 25.3 GB.

Why does the talk say 4 GPUs can leave too little room? With this kind of compressed attention and plain tensor parallelism, the cache is not split across the GPUs: every GPU in a copy keeps the full cache. So what a copy can hold is set by what is left on ONE GPU. On 4 GPUs that is 64 GB, room for 2 conversations of 256K. On 8 GPUs it is 122 GB, room for 4.

Across all 56 GPUs the totals come out about the same: 14 copies of 2 or 7 copies of 4, 28 long conversations either way. At 32K tokens, TP4 even holds slightly more in total (280 against 266). What changes is that each TP8 copy holds twice as many. A plausible reason this helps, my reading rather than the talk's: long sessions and the shared start of many prompts stay cached on the copy that serves them, which keeps cache hits high. Each GPU also reads half as many weights per token, which speeds up producing each output token, though the 8 GPUs must then exchange more data on every layer. The full 8-bit model does not fit on 4 GPUs at all. All of this assumes a 16-bit cache and plain tensor parallelism: engines that store the cache in 8 bits fit about twice as many conversations, and modes that split the cache across GPUs (data-parallel attention, or decode context parallelism) change the TP4 against TP8 trade.

Read next: Scaling and GPU infrastructure: serving models without burning money

The Revolut facts come from the talk itself, published by AI Engineer on YouTube on 29 October 2026, read through YouTube's English captions. The model facts come from Z.ai's model card, config file and licence, NVIDIA's NVFP4 model card and the Hugging Face API. Prices come from OpenRouter's public lists, the B200 memory from NVIDIA and the GPU price from Lambda, all read on 9 November 2026 and linked below.

I did not run GLM-5.3 or rent B200 GPUs. Revolut's speed, quality and productivity numbers are its own, from its own data. The memory arithmetic is an upper bound in 16-bit numbers and leaves out the serving software's own memory. The token mix is from this site's own coding agents, which run many jobs on one Mac, not from Revolut.