Revolut's tenfold AI coding saving is an inference bill, not a verdict
Revolut moved 2,000 engineers onto self-hosted GLM and cut AI coding costs by about 90%. The productivity claims behind that headline are harder to audit.
Vlad Makarovreviewed and published
Revolut says it moved roughly 2,000 software engineers off Claude Code and Codex and onto a self-hosted Chinese open-weight model, and that its AI coding bill is now about a tenth of what it was. The claim comes from David Abitbol, who runs AI operations and infrastructure at the London fintech, in a talk titled "Shipping like a startup, regulated like a bank" at the AI Engineer conference in New York last month. The talk was technical and unusually specific about the serving stack. It was also a company reporting on itself, and the numbers that travel furthest are the ones nobody outside Revolut can check.
The bill that shrank is the inference bill
The mechanism is easy to state. At roughly $0.30 per million input and output tokens, serving GLM on Revolut's own hardware cost about a tenth of buying the same work from Anthropic or OpenAI, The Register reports. Before that, Claude Code had arrived with strict per-engineer budget caps, because the early bills made no sense. "We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI," Abitbol said.
| Measure | Revolut's figure |
|---|---|
| Blended inference cost | about $0.30 per million tokens |
| Prior Anthropic/OpenAI spend | roughly ten times higher |
| Serving hardware | 56 Nvidia B200 GPUs, seven eight-GPU nodes |
| Model size | about 754B total parameters, 40B active per token |
| Weights and context memory | about 380 GB of weights, 6-26 GB per user |
That table is where the honest accounting starts, and where it stops. A 90 percent saving on tokens is not a 90 percent saving. An analysis by StartupHub.ai cited by Trending Topics notes that the figure excludes cloud, staff and monitoring. It also excludes the hardware, and Abitbol did not say whether Revolut bought its 56 B200s or rents them. When Nvidia launched the chip, Jensen Huang put it at $30,000 to $40,000 each, which is $1.7 million to $2.2 million for the raw silicon; complete eight-GPU servers run about $450,000 to $500,000, so seven of them land near $3.2 million to $3.5 million before power and floor space.
Two productivity numbers with no control group
The efficiency chart is the part everyone quotes: on a sample of about 2,000 engineers, coding time fell 60 percent and lead time, from picking up a feature to merging it, fell 70 percent. The earlier Gemini rollout, by contrast, bought 20 percent on coding time and 10 percent on lead time. The session notes on the talk page are more careful than the coverage: "The internal comparison does not isolate the model change from other causes." That is not a small caveat. The same year brought a new agent harness, a cheaper default, the removal of budget caps and a reorganisation of how work is queued. Any of those could move a lead-time chart.
Nor is there a published quality metric anywhere in the story. Nothing on defect density, rework, review cycles or incident rates attributable to machine-written code. Revolut is a regulated bank whose engineers ship payments software, so the absence is conspicuous rather than merely inconvenient. Faster merges are a throughput measure. They are not evidence that the code merged is as good, and the two only look interchangeable if you never measure the second.
Tuning, not more silicon, is the transferable lesson
The most useful part of the talk is not the price. It is that out-of-the-box inference was, in Abitbol's word, terrible, and that the fix was configuration rather than procurement. Hyperparameter tuning, NVFP4 quantization on Blackwell, disaggregated prefill and decode, and context windows sized to actual work took time to first token from 14.26 seconds during migration to 1.68 seconds on the same 56 GPUs, while requests per second rose 76 percent. The session page records that slide directly.
What engineers feel is the first token. "As soon as you exceed 5 seconds, my Slack inbox starts to blow up," Abitbol said, and he argues the target is under three seconds at around 256,000 tokens of context. Throughput he treats as nearly irrelevant at 200 tokens per second, because people now launch tasks in parallel and come back when they finish. Engineers signing in after 5pm get closer to 800 tokens per second, which says something about how much headroom is left in off-peak capacity.
Long context is a scam, and the router admits the limits
Abitbol's line about context windows is the one that will outlive the talk: "Long context is a scam. You really need to be mindful of that. Even though you can serve one million tokens, 256 [thousand] is really what you need." The reasoning is mechanical. Every reserved context slot competes with concurrent requests for GPU memory, and a bank paying for its own rack has a direct interest in not reserving capacity nobody uses.
The architecture that emerges is a portfolio, not a replacement. GLM is text-only, so a router sends vision and image work to Gemini or GPT Luna. Claude Code and Codex stay switched on for the cases that need them. "The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception," Abitbol said. Read carefully, that sentence concedes the argument: the frontier labs still win the hard cases, and Revolut has simply stopped paying frontier prices for the easy ones, which is most of them.
Chinese weights on American GPUs
A European bank running a model trained in China is a compliance question before it is an engineering one, and Abitbol did not duck it. "There's been a lot of stigma in using Chinese open source model," he acknowledged, per Trending Topics. Revolut's answer is jurisdictional rather than rhetorical: it downloads only the open weights and serves them on infrastructure in the United States, never through Z.ai's own inference endpoint. Data is processed only in countries with equivalent protection, meaning the US, EU, UK and Israel, and providers must default to zero retention. Coding agents run in a sandbox with approval rules, "always with a human in the loop," and the team keeps telemetry deep enough to reconstruct an incident.
That is a coherent position, and it is worth naming what it actually rests on. The objection to a Chinese model is answered by where the weights execute and who holds the logs, not by who trained the model. That distinction holds only while the weights stay downloadable and the serving stack stays in the building. It is a procurement and legal argument, and it evaporates the moment the model is consumed as an API.
Who can copy this, and who cannot
The economics here are not a pricing menu. They are open weights plus owned or rented GPUs plus roughly 2,000 engineers to amortise them across. Fifty-six B200s are a rounding error against a $115 billion valuation and a payroll of thousands; the same rack is a seed round for a twenty-person startup, which would also have to hire the person who knows why prefill and decode should be separated. The conference notes make the point bluntly: the $0.30 figure describes "Revolut's deployment and utilization, not a portable price for any self-hosted model."
Revolut has 80 million customers across 45 markets and an IPO targeted as early as 2028 at up to $200 billion, which is the context in which a multi-million-dollar inference rack is a cost-control measure rather than a bet. The narrower, defensible version of this week's headline is that a large, well-staffed engineering org cut its per-token bill by about tenfold, got measurably faster serving, and reports faster merges on an internal chart. Whether the cheaper model writes code a bank should ship is the question the talk did not answer, and the one that will decide whether anyone else follows.
