Some Revolut engineers were burning through $10,000 a month on AI coding tools.

David Abitbol of Revolut speaking into a microphone, next to his quote "The billing is flat... and the frontier is the exception."

Revolut moved about 2,000 engineers off Claude Code and Codex and onto an open-weight model it runs on its own GPUs.

Its AI coding bill fell to roughly a tenth, while coding time improved about 60% and lead time about 70%.

The savings are real, but they come from 56 GPUs and a team that tuned them for months, so this piece separates what transfers to a startup from what doesn't.

Below is the full breakdown: how Revolut got there, why its 90% won't be yours, and the middle path that works for most startups. Paid members also get the playbook by team size, a pilot design and the Break-Even Workbook.

1. The $10,000-a-Month Engineer (2025 to Early 2026)

Revolut did not start as an open-source shop. It rolled out Gemini in early 2025 and Claude Code in early 2026, like most fast-moving engineering teams.

The problem showed up on the invoice. David Abitbol, Revolut's ops and infrastructure lead, said at AI Engineer in New York that the company was "highly skeptical about the cost of AI" because some engineers were spending $1,000 to $10,000 a month with little visible return.

Some Revolut engineers spent $10,000 a month on AI coding tools, versus an enterprise average of $150-250 for Claude Code.

Those are outliers. Anthropic's own documentation puts the enterprise average for Claude Code at about $150 to $250 per developer per month, or around $13 per active day.

The most telling detail is what metering did to behavior. With strict budgets in place, engineers saved tokens for the hardest tasks and the end of the quarter. The tool was rationed exactly where it could have saved the most time.

Anthropic itself recommends starting with a small pilot group and measuring a baseline before a wider rollout. It is the step most teams skip, and it is the one that would have exposed the problem early.

The real cost of metered AI was not the bill. It was the rationing.

2. The Bet: An Open Chinese Model on 56 GPUs (April 2026)

In April 2026 Revolut began self-hosting GLM, the open-weight model family from Chinese lab Z.ai. It started on GLM-5.1 and later moved to GLM-5.3, released in August 2026.

The hardware is 56 Nvidia B200 GPUs, the equivalent of seven 8-GPU nodes. Eight is not arbitrary: GLM needs all eight GPUs of a node working together to fit the model and its working memory.

Seven nodes sounds like a lot. Spread across 2,000 engineers it is about one GPU for every 36 people, which only pays off because the cluster stays busy most of the day.

Revolut timeline: Gemini in early 2025, Claude Code in early 2026, self-hosted GLM-5.1 from April 2026, GLM-5.3 in August 2026, results shared in October 2026.

Revolut did not throw the frontier tools away. GLM is text-only, so a model router sends image and screenshot questions to Gemini or a GPT model. Claude Code and Codex stay available for exceptions.

  • Default path: self-hosted GLM-5.3 for code generation

  • Vision path: Gemini or GPT, chosen by the router

  • Exception path: Claude Code and Codex for the hardest work

For a regulated bank, the bet was as much about control as cost. The bill became flat: no per-token surprises at month end, and no data leaving for a third-party model on the default path.

Revolut did not replace frontier models. It demoted them to the exception.

3. Tuning, Not Hardware, Made It Work (2026)

Most people assume self-hosting wins because GPUs are cheap. Revolut's numbers say the opposite: the gains came from tuning the same hardware.

According to Abitbol's talk, time to first token fell from over 10 seconds to 1.6 seconds, and requests per second rose from about 3 to over 7. The team reported 88% better performance while absorbing 76% more demand, without adding GPUs.

After tuning, time to first token fell from over 10 seconds to 1.6 seconds on the same GPUs.
  • Context cap: long context was limited to 256,000 tokens, which covered real coding work

  • Precision: NVFP4 on Blackwell cut the GPU footprint about 40% versus FP8

  • Latency target: under 3 seconds to first token, because past 5 seconds engineers complain

  • Speed: about 200 tokens per second interactive, up to 800 after 5 p.m. when load drops

None of these levers is exotic. They are configuration choices a strong infrastructure engineer can make in weeks, but someone has to own them, watch them and fix them at 2 a.m.

Then behavior changed. Once tokens stopped being rationed, engineers used AI for work they used to skip, like debugging and clearing Jira tickets. That is part of why lead time improved more than coding time.

Cheap tokens change how a team works. That is the upside and the risk.

4. Why Your 90% Won't Be Revolut's 90%

The $0.30 per million tokens is a serving cost. By Revolut's own account it leaves out cloud spend, staffing and observability.

The entry ticket

The bigger issue is the entry ticket. The smallest serious GLM deployment is one 8-GPU node, and it costs money every hour whether anyone is coding or not.

  • At $6.69 per GPU-hour (Lambda's 8x B200 node list price): about $39,000 a month, running 24/7

  • At $8.08 per GPU-hour (GetDeploying median, November 10, 2026): about $47,000 a month

One 8x B200 node costs $39-47k a month, about the Claude Code bill of 200 engineers at $200 each.

At Anthropic's reported $150 to $250 per developer, one node equals the Claude Code spend of roughly 155 to 315 engineers, before anyone is paid to run it.

What an independent benchmark found

An independent benchmark points the same way. imec, the Belgian research center, presented at AI Engineer that owning infrastructure for coding assistance cut total cost of ownership by about 25%, not 90%. Large B200-class systems needed about 15% utilization to beat API prices; small setups needed about 89%, which imec called infeasible.

A half-idle node is the most expensive AI you can buy.

Utilization, not list price, decides who wins.

5. The Middle Path Most Startups Miss

Most write-ups frame this as Claude Code versus self-hosting. There is a third option: call the same open-weight model through an API and pay per token, with no GPUs to run.

A. Frontier tools

B. Open model via API

C. Self-host (Revolut)

Examples

Claude Code, Codex, Cursor

GLM-5.3 on Z.ai or OpenRouter

GLM on your own GPUs

Price anchor

$150-250 per dev per month

$1.40 in, $0.26 cached, $4.40 out per M

$39-47k per node per month

Main risk

Bill growth

Quality on hard tasks

Idle GPUs, unowned ops

To see what that means in practice, take Northwind, a fictional 25-engineer Series A startup. Its numbers are assumptions, but the prices are the published ones.

  • Today: 25 engineers at $220 each = $5,500 a month on Claude Code, plus $200 of admin time

  • Self-hosting check: one node at $39,000 is 7.1x the current bill, so Option C is out

  • Candidate: send the routine 70% of tasks to GLM-5.3 through a gateway and keep Claude Code for the hardest 30%

  • Result: $3,273 a month instead of $5,700. That already assumes routine usage on GLM grows 50% once tokens get cheaper, while Claude Code stays capped at 30% of today's spend. Setup costs $4,000, so payback takes about 1.6 months

The most telling detail is where the savings come from. Northwind does not switch models wholesale. It sends cheap work to a cheap model and keeps expensive work where quality matters most.

Northwind example: monthly savings of $2,427 shrink by about $950 per extra review minute and break even at about 2.5 minutes.

Here is the catch. If the open model adds review or rework time, engineers pay for it. Each extra minute per accepted task costs Northwind about $950 a month at $100 an hour, so the savings disappear at about 2.5 extra minutes.

The price of a token is not the price of a task.

What's next: paid members get the playbook by team size, a 4-week pilot that can't fool you, the gateway setup in 4 steps, prompts to measure your baseline and the Break-Even Workbook to fill in.

logo

Subscribe to PREMIUM to read the rest.

Become a paying subscriber of PREMIUM to get access to this post and other subscriber-only content.

Upgrade

Already a subscriber? Sign in

A subscription gets you:

  • Everything included in Plus.
  • Ongoing access to new AI agent templates and the full library.
  • Complete access to the investor database and pitch deck collection.
  • Save 75% compared with paying monthly.

Reply

Avatar

or to participate