Revolut chooses z.AI: what businesses can learn about private AI
The Revolut case connects open-weight models, predictable costs and infrastructure control. What it means for UTOMAI private AI in dedicated cloud or on-premise.
Editorial analysis and illustrative technical scenarios. They do not define the features or integrations included in a UTOMAI subscription. See Private AI On-prem pre-launch information.
Photo: Revolut — official press kit.
Revolut’s use of z.AI models offers a useful starting point for companies deciding how to adopt AI. Our reading is that the important business decision concerns the whole service: model quality, where information is processed, operating costs and responsibility for keeping it useful.
What Revolut reported
At AI Engineer NYC on 13 October 2026, David Abitbol described self-hosted GLM models from z.AI for roughly 2,000 Revolut engineers, initially GLM-5.1 and subsequently GLM-5.3. He reported a blended cost of $0.30 per million tokens and a bill one-tenth of previous Anthropic and OpenAI spending. Those tools remain available as exceptions. The conference transcript is the primary source for these statements.
These are company-reported figures, not a universal savings benchmark. The presentation does not provide a full, itemised total cost of ownership or a controlled comparison separating model choice from changes in usage. They cannot establish what a smaller company would save.
Open weights and a hosted API are different choices
Open-weight means that model weights can be obtained and run on compatible infrastructure, within the applicable licence. The official GLM repository lists downloadable models alongside hosted API access. Choosing one route or the other changes who operates inference and which service receives prompts and documents.
A model’s country of origin does not, by itself, identify where a request is processed. Equally, local inference does not settle every security question. A company still needs to check the model and its licence, software dependencies, access controls, logs, updates and any external connections. The processing boundary must be established for the whole configuration.
Predictability depends on capacity and usage
Our business takeaway is to compare the cost of a useful, checked result. API charges are one part of that cost. Dedicated infrastructure also brings capacity costs, maintenance and operating work. Light or intermittent usage can favour a consumption-based service; sustained demand can make reserved capacity worth evaluating. Both cases require measurements.
The vLLM deployment guide for GLM illustrates the operational work involved: hardware, precision, parallelism and serving configuration. Downloading weights is only the beginning. Before choosing an architecture, test representative tasks, simultaneous requests, response times and the effort needed to correct answers.
What this means for UTOMAI
For UTOMAI, this case supports private AI on dedicated infrastructure, with a shared workspace, configured models and agreed management. A support colleague consulting manuals, a sales team preparing a draft or an office searching procedures can use that environment. These are applications to evaluate, rather than results measured at Revolut.
Two deployment paths to assess
With UTOMAI Dedicated Cloud, private AI runs on dedicated cloud servers in Italy. The proposal defines resources, operating window, available functions and service responsibilities, without installing a machine at the customer site.
With UTOMAI Private AI On-prem, private AI runs on a dedicated machine at the customer’s premises, with local model execution and agreed management. This offer is in pre-launch; models, hardware, availability and concurrency need validation. External integrations can introduce additional data flows; fully local operation requires a validated local alternative for every required component.
Neither path implies that Revolut’s GLM deployment fits a UTOMAI appliance. GLM compatibility, availability and capacity are not confirmed by this article. Custom ERP/CRM connections, bespoke agents and additional workflows are assessed and quoted separately. Commercial terms are defined in a proposal after the deployment and service scope have been assessed.
Start with a task, a boundary and a measure
Choose one recurring task, identify the documents the team can use and agree how a person will check the result. Measure completion time, answer quality, correction effort and expected usage. That provides a basis for comparing deployment options and deciding whether an external model is needed for a particular task.
The decision for your company is which AI service you can use consistently, at an understood cost and within an agreed data boundary. Discuss your team’s requirements with UTOMAI.
Sources and further reading
Primary evidence: David Abitbol’s talk at AI Engineer NYC. Model documentation: z.AI’s official GLM repository. Deployment reference: vLLM GLM guide.
The Trending Topics article prompted this analysis. Statements about Revolut above are attributed to the conference transcript; the implications for UTOMAI are our editorial assessment. Revolut is not presented as a UTOMAI customer or partner.
