Revolut cut AI coding costs 90% by ending usage rationing. A sub can copy the pricing lesson, not the 56 GPUs
Revolut says self-hosting an open model dropped its AI coding costs about 90% and let engineers stop rationing use. A contractor can't run that hardware, but the flat-billing lesson applies to every AI renewal.
A fintech says it cut AI coding costs about 90% by hosting an open model itself, and the lesson for a contractor isn't the hardware. It's that metered pricing makes people ration AI, and the rationed tasks are often the ones that pay. Before your next AI renewal, ask for flat or capped pricing.
What did Revolut actually do?
According to The Register, Revolut's infrastructure lead David Abitbol told an October AI Engineer conference in New York that the company made self-hosted open models its default for coding. Engineers had reportedly been spending $1,000 to $10,000 a month each on commercial tools with uneven return. Self-hosting GLM, an open model from Z.ai, came to about $0.30 per million tokens, roughly a tenth of the commercial cost.
The setup runs on 56 Nvidia B200 GPUs for about 2,000 engineers. Revolut reports a 60% improvement in coding time and 70% in lead time. StartupHub notes these are self-reported and the $0.30 figure leaves out cloud spend, staffing and observability.
Can a contractor do the same?
No, and it shouldn't try. Compare the scale:
| Revolut | Typical 50-person sub or mid-size GC | |
|---|---|---|
| Users | About 2,000 engineers | A handful of estimators, PMs, coordinators |
| Hardware | 56 B200 GPUs | None |
| Staff to tune and run it | Dedicated infrastructure team | Usually none |
| Realistic move | Self-host | Negotiate pricing, pilot cheaper hosted tools |
Our earlier piece on hosted versus self-hosted for bid files covers when private hosting makes sense, mostly for data control, not cost.
What's the lesson that does transfer?
Abitbol said flat billing meant no bill shock, and that cheap tokens meant engineers stopped rationing. He also said the biggest gains came from debugging and ticket handling, not generating code.
Translate that to a project office:
- A PM who worries about per-query charges won't run the AI over the whole RFI log, only the one that's urgent.
- An estimator who is watching a usage meter won't re-run bid leveling after each addendum.
- A coordinator won't check every submittal against the spec section if each check costs money.
The unglamorous tasks are where the hours are. Metering discourages exactly those.
What should you ask your AI vendors?
- Is there a flat or capped plan? Seat pricing with a hard monthly ceiling beats open-ended usage billing.
- What happens at the limit? Throttling is fine. A surprise overage invoice is not.
- Can we see usage by task? Our per-task pricing piece explains why that matters for judging a pilot.
- Is the price tied to one model? Cheaper models improve fast, and your contract shouldn't lock in last year's price.
What are the limits of this story?
It's one company's account, at a conference, with unaudited numbers. Revolut still sends image work to other models, and Abitbol himself called long context "a scam" past what you need, which is useful if you're feeding in spec books. He also acknowledged security concerns about Chinese open-weight models and argued they mostly disappear when you host them yourself. A contractor with bid files and owner data should weigh that before touching any open-weight tool.
Takeaway
This week, list every AI tool your office pays for and note how each one bills. For the metered ones, ask whether anyone is skipping useful work to save money. If so, request a capped plan at renewal.
Forward this to whoever approves your software spend. Construction AI Brief publishes three times a week at constructionaibrief.com.
- Should a mid-size contractor self-host an AI model to cut costs?
- Almost certainly not. Revolut's reported setup runs on 56 Nvidia B200 GPUs for about 2,000 engineers. A contractor's better move is negotiating flat or capped pricing with the vendors it already uses.
- Why does metered AI pricing change how people use it?
- When every query adds to a monthly bill, staff ration use and skip the tasks that would pay off. Revolut's infrastructure lead said flat billing removed bill shock and let engineers stop counting tokens.
- What AI tasks produced the biggest gains at Revolut?
- Debugging and handling Jira and Confluence tickets, not writing new code, according to The Register's report of the talk. The construction parallel is routine admin work: RFI logs, submittal tracking and meeting follow-ups.
- Are Revolut's savings figures verified?
- No. They come from a company talk at a conference and exclude cloud spend, staffing and monitoring costs, per StartupHub's coverage. Treat 90% as a claim, not a benchmark.