Skip to content
audiotale.net
Sign in
NewsNews

Revolut says self-hosted GLM cut AI coding costs 90% for 2,000 engineers on 56 B200s

A Revolut infrastructure lead told AI Engineer that self-hosting open-weight models cut coding inference costs by about 90 percent; the reports disagree on the model.

audiotale.net team

· 6 min read

Written by AI from the sources linked below. How we use AI

On this page
  1. Key points
  2. What happened
  3. Background
  4. How it works
  5. What it means
  6. What we don't know yet
  7. What to watch next
  8. Sources

Revolut says it has cut the cost of AI coding assistance by roughly 90 percent in under a year by making self-hosted open-weight models its default instead of relying mainly on commercial services such as Claude Code. The figures come from a conference talk by one of the fintech's infrastructure leads, reported by The Register and by StartupHub, and they have not been published by Revolut as a case study. That matters, because the two reports disagree on a basic detail: which model Revolut runs.

Key points

  • Revolut's David Abitbol told the AI Engineer conference in New York in October 2026 that self-hosting cut its AI coding inference costs by about 90 percent, according to The Register.
  • The setup serves roughly 2,000 engineers on 56 Nvidia B200 GPUs, the equivalent of seven eight-GPU DGX B200 systems, per The Register.
  • The cost quoted is about $0.30 per million tokens, around a tenth of what the company paid for commercial coding services, according to both reports.
  • Revolut says coding time improved 60 percent and lead time 70 percent after the rollout; these are the company's own figures from a talk, not an independent measurement.
  • The Register says the model is Z.ai's GLM-5.1, later GLM-5.3; StartupHub says GLM 4.5. We could not resolve the difference.

What happened

According to The Register, which published on 3 November 2026, Abitbol, described as ops and infrastructure lead at Revolut, spoke at the AI Engineer conference in New York in October. He said the company had rolled out AI coding tools to roughly 2,000 engineers and saw a 60 percent improvement in coding time and a 70 percent improvement in lead time.

Revolut had tried commercial tools first. The Register reports that it deployed Gemini in early 2025 and Claude Code in early 2026. Abitbol said the company was "highly skeptical about the cost of AI" after seeing engineers spend $1,000 or even $10,000 a month with little return, so it launched with strict budget limits. Limited token budgets pushed the company to look at running models itself.

The Register says Revolut began with GLM-5.1 and moved to GLM-5.3, serving it on 56 Nvidia B200 GPUs. At $0.30 per million input and output tokens, self-hosting cost about a tenth of using OpenAI's Codex or Anthropic's Claude Code, the report says. StartupHub, which published on 29 October and links the talk video, gives the same price and the same GPU count, but names the model as GLM 4.5.

Background

Metered per-token pricing makes coding assistants a variable cost that grows with use, which is why Revolut capped spending at the start. The alternative is to rent or buy the hardware and run an open-weight model, which turns the bill into a fixed cost. Revolut's reported answer is that, for roughly 2,000 engineers, the fixed-cost route paid off.

Abitbol framed that as the main benefit. The Register quotes him: "The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception." The Register adds that cheap tokens changed how engineers worked. With budgets no longer rationed, they used models for debugging and for handling Jira and Confluence tickets, not just writing code. StartupHub attributes the larger lead-time gain to those tasks outside pure coding, such as inbox triage.

The Register notes that Revolut still uses commercial models where it must. GLM is text only, so a model router sends vision queries, such as UI work, to models like Gemini. Code generation stays on GLM.

How it works

The Register says GLM-5.1 has roughly 754 billion total parameters, with 40 billion active per generated token, and needs about 380 GB of GPU memory just for the weights, plus 6 to 26 GB per user for context depending on precision. That is why the deployment needs a rack of B200s rather than a single server.

Abitbol said the biggest gains came from tuning, not adding GPUs. The Register lists hyperparameter tuning, NVFP4 quantization, disaggregated prefill and decode, and tuned context lengths. StartupHub adds details from the talk: performance improved 88 percent for 76 percent more demand without extra GPUs, NVFP4 on Blackwell saves about 40 percent of GPU footprint compared with FP8 with no precision loss for this use, and GLM needs eight-way tensor parallelism rather than four-way to fit its KV cache. We have not verified these against the talk itself.

On context length, Abitbol was blunt. "Long context is a scam," The Register quotes him as saying, arguing that although a million tokens can be served, about 256,000 is what engineers actually need. At that size the company reports a time to first token under 3 seconds and about 200 tokens per second of generation. He told the audience that beyond 5 seconds to first token, complaints arrive. StartupHub reports first-token time fell from over 10 seconds to 1.6 seconds, with requests per second rising from 3 to over 7. Engineers who sign in after 5 p.m. see closer to 800 tokens per second because load drops, The Register says.

Improvement Revolut says it measured, by period and metric (company-reported)%
Improvement Revolut says it measured, by period and metric (company-reported) (%)
%
Self-hosted GLM, lead time70
Self-hosted GLM, coding time60
Gemini 2025, coding time20
Gemini 2025, lead time10

Source: StartupHub, summarising the AI Engineer talk; The Register

What it means

For engineering leaders. The claim is that a flat-cost model changes behaviour, since engineers stop rationing tokens. If that holds, the saving is not only the 90 percent on the invoice but the extra work people are willing to hand to the model. It is also self-reported, and Revolut chose both the metrics and the sample.

For smaller teams. Abitbol's setup needs 56 high-end GPUs, steady tuning and a security posture a regulated bank can enforce. StartupHub's reading is that smaller teams without that leverage may still face large bills or slow first tokens. The talk, as reported, does not put a number on staffing, cloud or observability costs in the $0.30 figure.

For the model vendors. Abitbol's stated position is that commercial models become the exception for hard or visual tasks rather than the default. A single customer's talk does not show a market shift, but it shows how one large buyer is thinking.

Security is part of the picture. StartupHub reports that Abitbol framed the threat as local: the system affected is the coding-agent harness used by employees, with a sandbox, human in the loop, and allow lists for commands. The Register adds that he acknowledged concerns about Chinese open-weight models but said the real risk is calling them through APIs running in China; run on-premises or in regional clouds, he said, there is nothing to worry about. That is his view, not an independent assessment.

What we don't know yet

  • Which model. The Register says GLM-5.1 then GLM-5.3. StartupHub says GLM 4.5. We did not watch the talk and cannot say which is right.
  • The full cost. The $0.30 per million tokens figure does not obviously include hardware purchase or rental, staff and monitoring. Neither report gives a total cost of ownership.
  • How the productivity numbers were measured. Neither report says how coding time and lead time were defined or whether a control group was used.
  • Revolut's own confirmation. We found no Revolut press release or blog post.

What to watch next

  • A published write-up or slides from Revolut with the model name and a cost breakdown.
  • Whether other large engineering teams report similar savings from self-hosting, and at what scale.
  • Whether commercial vendors respond with flat-rate or committed-use pricing for coding agents.

Sources

SaveStart the discussion

Written by

audiotale.net team

The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.

Read next