Revolut Cut Its AI Coding Bill by 90%. The Talk Leaves Out What the GPUs Cost
Revolut cut its AI coding invoice to a tenth on 56 self-hosted GPUs. The talk never says what they cost. Here is that number, and what it changes.
A European bank stood up at a conference last month and said it had cut its AI coding invoice to a tenth of what it was, by moving 2,000 engineers onto an open-weight model it serves itself, on a fleet of 56 GPUs. The clip went round, and the lesson most people took from it was that the self-hosted AI coding model cost is simply lower, so buy hardware.
The talk is real, it is on the record, and the figures in it are the speaker’s own. What it never mentions, across eleven minutes and forty-seven seconds, is what those GPUs cost. Put a published price list next to the number and the saving turns out to come mostly from somewhere else — somewhere a company a fraction of Revolut’s size can reach without buying anything.
What Revolut actually said
The source is David Abitbol, who leads AI operations at Revolut, speaking at AI Engineer NYC 2026 on 13 October. The session was billed as “Shipping like a startup, regulated like a bank”; the published recording carries the blunter title about the 90%.
The shape of it: Revolut introduced Gemini to engineers in January 2025 and measured a 10% improvement in lead time and 20% in coding time. Claude Code followed, launched at scale with strict budget limits because, in Abitbol’s account, they were watching individual engineers spend “a thousand a month, ten thousand a month” without matching returns. He is explicit that Claude Code under those caps produced little measurable impact. In April 2026 they started self-hosting GLM, and the numbers moved: 60% improvement in coding time, 70% in lead time, measured across the same 2,000-engineer sample.
Serving that takes roughly 56 Nvidia B200 GPUs for GLM-5.3, a model of around 800 billion parameters. The blended cost, in his words, is “thirty cents per million” tokens, and the resulting invoice is “one-tenth of what we were spending” with Anthropic and OpenAI.
One correction worth making before anything else. Several of the most-shared write-ups of this talk, StartupHub’s among them, say Revolut runs GLM 4.5. The talk says GLM-5.1 from April and GLM-5.3 now. If you have been checking these numbers against a model’s capabilities, you were handed the wrong model.
The self-hosted AI coding model cost the talk leaves out
Here is what eleven minutes of a technical talk about GPU economics does not contain: the price of the GPUs.
Abitbol never says what the fleet costs. He never says whether Revolut owns it or rents it. He never mentions the people who tune it — and he spends a third of the talk establishing that tuning is the whole job, which rather implies there are people doing it. And the thirty cents is described as a blended cost of token, which is a unit price, not a total cost of ownership.
So put a number on it. CoreWeave publishes its rates: an eight-GPU NVIDIA HGX B200 node is $68.80 an hour on demand in North America, with spot at $34.11.
Fifty-six GPUs is seven of those nodes. That is $481.60 an hour, about $11,600 a day, roughly $347,000 over a thirty-day month at list. CoreWeave advertises up to 60% off for committed usage without publishing the committed rates; at that ceiling the same fleet lands near $139,000 a month.
Three caveats, all of them load-bearing. CoreWeave is one vendor’s list price and the talk names no provider, so this is a yardstick and not a report of Revolut’s bill. On-demand list is the most anyone pays, and a bank at this scale would not. And it is not established that the thirty cents is a total-cost figure rather than a compute-only one — the talk simply does not say.
What survives all three is the sentence that matters to anyone considering this: the fixed cost of that fleet is six figures a month, and it does not shrink when your usage does. Run the arithmetic backwards and the point gets sharper. For thirty cents a million to hold against $347,000 of monthly hardware, something like 1.16 trillion tokens have to pass through it — about 578 million tokens per engineer per month. At the discounted end it is still 231 million each. The unit price is not a property of self-hosting. It is a property of enormous, sustained, well-cached volume, divided across a fleet that is already paid for.
Where the saving actually came from
The most interesting thirty seconds of the talk is the part nobody quoted.
Abitbol notes that the gain was larger in lead time than in coding time, and draws the obvious inference: that “the bottleneck for engineering is not necessarily coding”. What engineers did with unlimited tokens was stop rationing them — on inbox triage, on Jira, on the work around the work that nobody had been willing to spend a metered token on. His framing is that people no longer had to save their budget for the hardest task of the quarter.
Read that against the Claude Code result, where the tool was present and the budget cap was on, and impact was slight. Two things changed between those two measurements: the model became open-weight, and the cap came off. Only the first required GPUs.
That reframes the whole story. The 90% is a procurement outcome — a cheaper per-token price made an unlimited-usage policy affordable, and the unlimited-usage policy is what produced the productivity numbers. The hardware was how Revolut, at Revolut’s volume, reached that price. It is not the only way to reach it.
What the same move costs a company that is not a bank
Every figure below is from the vendor’s own published page, read on 9 November, in USD.
| Route | Published price |
|---|---|
| Claude Code on a subscription | Pro $20/mo, Max $100/mo, Ultra $200/mo — Claude Code included in all paid plans |
| Claude API, Sonnet 5.5 | $2 input / $10 output per million, $0.10 cache read |
| Claude API, Opus 5.5 | $4 input / $20 output per million, $0.20 cache read |
| GLM-5.3 from Z.ai’s own API | $1.40 input / $4.40 output per million, $0.26 cached input |
| GLM Coding Plan | from $18/mo, works inside Claude Code, Codex and OpenCode |
| 56 B200s at CoreWeave list | $481.60/hour |
The row that matters is the second-to-last. GLM-5.3 — the exact model Revolut put 2,000 engineers on — is sold by its own maker on a coding plan starting at $18 a month, and Z.ai’s documentation names Claude Code, Codex, OpenCode and ZCode among more than twenty supported agent tools. You keep the harness your team already knows and change what is behind it.
That is not the same deal Revolut got. Their blended thirty cents beats Z.ai’s hosted rate by a wide margin, as it should, because they bought the capacity outright. But against a $100-a-month Max seat, an $18 plan on an open-weight model gets you a long way down the same road with no GPUs, no tuning and no one to hire for it. For most teams reading this, that is the whole of the available saving, and it arrives on a Tuesday afternoon.
One detail connects the two ends of this. Z.ai quotes its plan’s token allowances at a 95% cache hit rate; Abitbol names 90% as the floor for coding workloads on his own fleet. Two unrelated primary sources, both saying the economics turn on how much of your context repeats. A workload that does not cache like coding does will not see any of these numbers — a point worth holding onto if you have been reading about how agent memory tools pay for reinjected context on every turn, which is the same arithmetic wearing a different hat.
Running it entirely on your own machines, and the ceiling
If the appeal is that nothing leaves the building, that path is real and it got easier this month. Ollama’s v0.40.14 release on 3 November extends ollama launch claude so a model’s own output token limit is applied rather than Claude Code’s 128k default, which was the kind of papercut that made the local route feel unfinished. Pointing an existing agent harness at a model on your own hardware is now a command, not a project. The full setup for running local models with Ollama covers the rest.
Then the ceiling, stated plainly, because this is where most write-ups stop being useful. Revolut’s model is 800 billion parameters and needs 56 datacentre GPUs to serve two thousand people. What you can realistically run on your own machines is a different class of thing. The same Ollama release added mimo-v2.6, Xiaomi’s 9B model built for code and agent workflows — a genuinely useful tool, and roughly one ninetieth the size of the one in the talk.
That gap is not a reason to skip the local route. It is a reason to be honest about what you are choosing it for. Nothing leaving the building is a real benefit with a real price, and the tradeoffs are the ones covered in the business case for keeping a model in-house. Just do not expect a 9B model to do what the talk describes, any more than you would expect a 1.7B model to replace a 9B one because a benchmark said so.
What to check before you copy any of this
Most of the caveats here come from the primary source, which is the fairest place to get them.
The comparisons are observational. Revolut changed a great deal over the period the chart covers, and the talk does not isolate the model’s effect from everything else. The figures are self-reported by the company that benefits from telling the story, at a conference, with no invoice shown.
The primary source also disagrees with itself in two places, and the AI Engineer page flags both. The lead-time improvement is 70% in the spoken account and 60% on the slide. And the quantisation claim — FP4 rather than FP8 on Blackwell hardware, worth about 40% less GPU footprint for no loss of precision — is stated the other way round in the recording. The intended meaning is clear enough; the fact that a talk this specific contains two unreconciled numbers is itself a reason to treat the headline as an estimate.
Then there is the apparatus. Revolut is an EU-regulated bank, and the controls Abitbol describes are not optional extras: a sandbox that demonstrates a human is in the loop, with rules fine-grained enough to “allow grep but not Gradle”; inference kept inside countries covered by the Data Equivalence Framework; providers on zero data retention; OpenTelemetry deep enough to reconstruct an incident. He also addresses the Chinese-model question directly, and his answer is narrower than the reassurance it gets quoted as: you download the weights and run them on your own servers, so no inference request goes anywhere near the model’s origin. That is true, and it is a statement about network traffic, not about anything else.
The decision
The crossover is not subtle, and you can place yourself on the right side of it with one division. A six-figure monthly fixed cost only beats a metered bill when the metered bill is also six figures. Against the cheapest version of the fleet above, near $139,000 a month, that is somewhere north of a thousand Max seats at $100 each before owning hardware is even worth modelling. If that is not you, the hardware question is answered.
So take an open-weight model on someone else’s infrastructure, put it behind the harness your team already uses, and spend the difference on removing the budget cap. On the published prices that is $18 to $100 a month per seat instead of per-token anxiety, and if the talk is right about where the gains came from, the cap is the part that was costing you.
If you are on the other side of that division, the thing to budget for is still not the GPUs. It is the person who spends a quarter moving time-to-first-token from ten seconds to 1.6 on the same fleet, which is what Abitbol spent most of his talk describing.
Everyone else should take the cheaper lesson. Revolut’s engineers got faster when they stopped counting tokens. You can test that this week, on any model, for the price of raising a limit.
Sources
Every figure in this article traces to one of these. Dates are when the page was read.
- How Revolut Cut AI Coding Costs 90% With Self-Hosted Open ModelsAI EngineerAccessed Nov 9, 2026Official docs
- Shipping like a startup, regulated like a bankAI Engineer NYC 2026Accessed Nov 9, 2026Official docs
- CoreWeave Cloud PricingCoreWeaveAccessed Nov 9, 2026Vendor pricing
- Pricing — Z.AI Developer DocumentationZ.aiAccessed Nov 9, 2026Vendor pricing
- GLM Coding Plan overviewZ.aiAccessed Nov 9, 2026Vendor pricing
- Pricing — Claude Platform DocsAnthropicAccessed Nov 9, 2026Vendor pricing
- Plans & pricingAnthropicAccessed Nov 9, 2026Vendor pricing
- Ollama v0.40.14 and v0.40.15 release notesOllamaAccessed Nov 9, 2026Official docs
- Revolut says self-hosted AI cut coding costs 90%StartupHub.aiAccessed Nov 5, 2026Independent
Written by
Full-stack developer who builds and runs AI automation systems in production. Runs local LLMs on personal hardware, builds N8N pipelines that actually ship, and deploys on Cloudflare Pages. Every guide on Pipeline Monk is tested on real consumer hardware — a Ryzen 7 5800HS with 16GB RAM and a GTX 1650. If it works on that, it works.