London-based British neobank Revolut has made a self-hosted AI its default coding solution, in place of Claude Code, Anthropic's tool. It runs the GLM models of Z.ai, a Chinese developer of open-weight models, on its own machines. David Abitbol, its head of operations and infrastructure, claims a 90% drop in inference costs in less than a year. He said so at the AI Engineer conference in New York in October 2026, according to The Register of 3 November 2026.

What Revolut announces and what remains to be verified

Revolut's figures are statements from its executive, not audited results. The Register and TechRadar Pro, which picked them up on 9 November 2026, report them based on his public talks. They cover a specific scope: the coding tools of around 2,000 engineers, not all of the bank's AI.

FigureWhat it measuresNature
90%Drop in inference costs for coding, in less than a yearStatement by David Abitbol
$0.30Cost per million input and output tokens with self-hosted GLM-5.1Statement, cost scope not detailed
60% and 70%Gains in coding time and in development lead timeStatement, measurement method not published
56 Nvidia B200 GPUsHardware currently serving GLM-5.3Statement on infrastructure
380 GBGraphics memory needed for the GLM-5.1 weights aloneModel characteristic, according to The Register

The $0.30 price applies to GLM-5.1, the first version adopted. Revolut now serves GLM-5.3. According to The Register, this cost was about one tenth of that of Codex, OpenAI's tool, or of Claude Code.

The starting point sheds light on the decision. Revolut had deployed Gemini, Google's model, in early 2025, then Claude Code in early 2026. "We were seeing engineers spend $1,000 a month, $10,000 a month, without much return on investment," David Abitbol explained, quoted by The Register. Engineers were therefore working with a strict token budget.

Claude has not disappeared for all that. GLM only handles text. A router sends requests involving images to Gemini or GPT Luna. "Frontier models remain the exception," the executive sums up.

External API or self-hosted AI: what trade-off?

The trade-off sets a variable bill against a fixed investment. With an external API, the company pays for every token consumed. With self-hosting, it pays for the hardware, its operation and the team that tunes it. "Billing is fixed, the nasty surprises at the end of the month have disappeared," says David Abitbol.

The hardware is heavy. According to The Register, GLM-5.1 has around 754 billion parameters, of which 40 billion are activated for each token produced. Its weights take up some 380 GB of graphics memory. Each user requires a further 6 to 26 GB to hold their context, depending on the chosen precision. Revolut's 56 B200 GPUs are equivalent, in number of cards, to seven DGX B200 systems with eight GPUs each.

Skills matter as much as cards. According to David Abitbol, the clearest gains did not come from additional GPUs. They come from tuning the model:

  • quantisation to the NVFP4 format, which lightens computation;
  • separating the phases of reading the request and generating the response;
  • a context window limited to 256,000 tokens, whereas the model can serve one million.

The team aims for under 3 seconds to the first token and around 200 tokens per second. Beyond 5 seconds of waiting, complaints pour in, the executive recounted.

TechRadar Pro qualifies the gain. GPUs, their maintenance and electricity may have pushed up other expense items, the magazine notes. The 90% figure covers inference, not the total cost of ownership.

The same calculation is being made in Switzerland, on a smaller scale. The canton of Valais runs its sovereign AI of Valais, AVA, on dedicated GPUs located in Switzerland. It puts the cost at 2 francs per month per daily active user. As with Revolut, nothing indicates whether this amount includes the purchase of the hardware.

A model of Chinese origin in a bank: what risk?

The question of the model's origin is real, and Revolut does not dismiss it. According to The Register, David Abitbol acknowledges that Chinese open-weight models can raise concerns about privacy and security. In his view, the main danger lies in using them through APIs operated from China. On internal servers or in a regional cloud, there would be "really nothing to fear". That is his assessment, not an independent evaluation.

He also highlights data privacy and the auditability of the GLM models, according to TechRadar Pro. Self-hosting solves part of the problem: requests do not leave the infrastructure chosen by the bank.

Our analysis: it does not solve everything. An open-weight model publishes its weights, not necessarily its data or its training method. The Swiss language model Apertus, released by EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS), goes further. According to EPFL, its architecture, weights, data and training recipes are documented. Lausanne start-up Giotto also offers on-premises installation of its Solo and Alto models, presented as sovereign AI.

AI in Swiss banking: what FINMA expects

In Switzerland, no AI-specific rule targets banks. The Swiss Financial Market Supervisory Authority (FINMA) published its guidance on supervision 08/2024 on 18 December 2024. It covers governance and risk management related to the use of AI. According to FINMA, this type of text is not intended to have any legal effect: it clarifies how to apply existing rules, as detailed on our page FINMA and artificial intelligence.

FINMA's press release of 18 December 2024 lists the risks to monitor:

  • model risks: lack of robustness, accuracy or explainability, and bias;
  • data-related risks: security, quality and availability;
  • IT risks and cyber risks;
  • growing dependence on third parties;
  • legal and reputational risks.

FINMA asks institutions to identify, assess, manage and monitor these risks. Our analysis: Revolut's choice touches several lines at once. It reduces dependence on an API provider, but leaves model risks and infrastructure risks fully in place.

Outsourcing is a sensitive point. In its survey published on 24 April 2025, FINMA questioned around 400 institutions. About half were using AI or developing their first applications. Among them, 91% were using generative AI, which increases dependence on BigTech providers. Smaller institutions often rely on third-party applications. FINMA points out that outsourcing critical functions remains a central source of operational risk.

Its principle remains "same business, same risks, same rules": the same activity calls for the same rules, whatever the technology. FINMA also recommends that institutions contact it early if they plan to use AI in critical processes.

The AI-assisted programmer is already a reality in Swiss banks. UBS, the country's largest bank, has more than 18,000 developers using AI tools, according to figures published with its results of 28 October 2026. Our article on AI in banking at UBS details its 608 use cases in production.

What this changes for Revolut's Swiss customers

Revolut has more than 1.3 million users in Switzerland, according to finews.ch in October 2026. It does not hold its own Swiss banking licence there and applied to FINMA for one in September 2026. Our analysis: if this licence were granted, its uses of AI would fall within the scope of FINMA's expectations of supervised institutions.

Revolut's technical dependence has already had a visible effect in Switzerland. On 20 October 2026, an outage in Google Cloud's London region coincided with the blocking of its card payments and its app. The link with Google was based on a source cited by the Reuters news agency. According to The Register, the change of model, for its part, concerns the engineers' coding tools.

What we don't know yet

  • The cost of buying or renting the 56 B200 GPUs: David Abitbol did not put a figure on it in the reported talks.
  • Where the models are hosted, on Revolut's servers or in a regional cloud.
  • The method used to measure the 60% and 70% gains, and the comparison period for the 90% figure.

Sources