Revolut Cut Its AI Bill Tenfold by Buying Machines, Not Tokens
The number everyone repeats is the tenfold saving. The decision behind it was switching engineers from metered usage to owned capacity, which removed the reason they were rationing. What changed in the work after the meter disappeared.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page
David Abitbol runs AI operations at Revolut. He gave a ten-minute talk at a New York conference in mid-October, and the recording only circulated after it was republished to YouTube on 29 October. Numerama read through it and pulled out the numbers, which come from measurements across roughly 2,000 engineers.
The headline is that Revolut says it cut its AI spend to about a tenth of what it was paying Anthropic. The more useful story is how it got there. The bank stopped buying tokens by the million and started buying chips.
Three phases, three different bottlenecks
Revolut's numbers come in stages, and each stage moved a different number.
Early 2025, with Gemini: 20% less time spent writing code, 10% shorter delivery lead time. Delivery lead time is the gap between an engineer picking up a feature and that feature being ready to deploy, so a 10% improvement there is a workflow change, not a typing-speed change.
Then Claude Code, brought in with strict spending caps. Abitbol's description of that period is blunt. Engineers were running up bills of a thousand, ten thousand a month, he said, without much return. Of Claude Code itself in that configuration, he said it did not have much impact at the time.
April 2026, on GLM-5.1 with no token limit: 60% less coding time and 70% shorter delivery lead time, according to the company. Same teams and same product, and the coding-time figure tripled against the capped arrangement.
The model did change between those last two data points, so a clean reading is impossible. But the cap changed at the same time, and Abitbol points at the cap. "Use AI without feeling guilty" is how the French coverage renders his framing of unlimited tokens. That phrase is the whole mechanism. A spend cap turns a tool into a budget line, and a budget line turns every request into a small decision about whether this one is worth it.
What engineers do differently when the meter is gone
Two observations from the talk are worth more than the percentage claims.
First, delivery lead time improved more than coding time. Revolut's own conclusion is that writing the code was not the main bottleneck. If correct, that reframes the entire pitch for coding assistants. The win was not that lines appeared faster; it was that the surrounding work moved, which is exactly the part a per-seat or per-token cap suppresses because it looks like off-task usage.
Second, engineers started handing the assistant tasks that had no visible output: digging through mail, searching internal tools. That is what happens when the marginal cost of a request goes to zero. People stop triaging requests by expected value and start using the tool the way they use a search box.
Anyone who has run a metered image or video pipeline knows this reflex from the other direction. You generate at draft quality because a high-resolution pass costs real money. You try three variations instead of twelve. You keep the reference count down because each one adds tokens. None of those decisions is wrong, and together they quietly standardize the output at "good enough to not re-run."
Owning capacity means running a service
The cost side of the switch is stated plainly: 56 Nvidia B200 chips serving the engineering organization, an average of about 30 euro cents per million tokens, and a total bill around one tenth of the previous vendor spend. Usage is not free, because the bank pays for the machines and has to add more when demand outruns them.
The unglamorous part is the tuning. Abitbol admits the rollout was not smooth. Getting the wait before the first token down from more than ten seconds to 1.6 seconds took long adjustment, on the same number of chips, while demand climbed from three to more than seven requests per second. That is a serving problem that no vendor absorbs for you. Self-hosting open weights converts a procurement contract into an operations team, and the operations team is not free either.
It is also not a clean break from vendors. Open models are now the priority at Revolut, but Claude Code, Codex and Gemini remain available for some uses. The realistic shape is a portfolio, not a replacement.
What the case does not settle
The numbers are Revolut's own, presented by an executive with an interest in the story landing. There is no independent measurement here, and a 2,000-engineer sample from one fast-moving fintech says little about a five-person studio.
The risk answer in the talk is also narrower than the question. Abitbol explained that the weights are downloaded and run on US-based infrastructure rather than routed through a service hosted in China, and that data stays inside the EU or in countries with equivalent protection, inside an isolated environment with pre-approved operations and a human in the loop. That addresses data circulation. It does not address the model's own behavior or the quality of the code it produces. Numerama's article links back to its own earlier reporting on a Z.ai coding assistant that sent users' projects to the cloud without telling them, which is the kind of thing a data-residency answer does not cover.
The same fork, for anyone generating images or video
Revolut's decision maps onto a question that image and video teams hit earlier than they expect: pay per generation, or own the throughput.
Per-generation pricing is the default and it is rational at low volume. It also shapes behavior in ways that show up in the work rather than the invoice: fewer exploratory passes, fewer re-rolls at final resolution, reference images trimmed to keep the count down. The pipeline gets tuned for cost predictability, and cost predictability pulls toward the safest output.
Reserved capacity or your own GPUs flip that. The meter stops arguing with you mid-session, and the cost becomes a utilization question: what percentage of the day are those cards doing anything. Below a certain duty cycle, owned capacity is a more expensive way to be idle, and the latency, queueing and failure behavior all land on your team.
Two things are worth measuring before copying the move. Your monthly spend at current habits tells you the ceiling you can afford to replace. The shape of your demand tells you whether owned capacity can pay for itself, because steady production work utilizes hardware and sporadic client bursts do not. Revolut's numbers look dramatic partly because the before-state was thousands of dollars per engineer per month.
The unit you bill in decides the behavior. Bill per call and the team optimizes calls; pay for machines and the optimizing moves to the pipeline, which is where the interesting decisions were all along.
Related reading: what routing video work by cost per second actually looks like, and why the price of local model weights is not the price of running them.