Starling trains its own AI model and keeps it off the hyperscalers

Starling Bank said on 14 October that it is building a transactional foundation model with support from Nvidia. The model will be trained on more than 15 billion anonymized transactions from Starling’s own customers and hosted by Civo, a cloud provider that keeps both its data centers and its ownership in the same country as the bank. Starling intends to put it into production in early 2027.

The release describes a reusable layer, which Starling calls Starling Intelligence, built to find patterns in anonymized transaction data and to support several applications. Its first job is to improve the bank’s existing fraud detection. After that it is meant to augment Starling Assistant, the agentic AI tool Starling launched for personal accounts in March 2026 and for business customers on 24 August.

“Now we want to connect these tools with a single intelligence layer that’s trained entirely on the specific needs of our customers,” said Harriet Rees, group chief information officer at Starling, who joined the bank in 2018 and founded its data science practice.

Starling Assistant is built with Google Gemini running on Google Cloud, according to the bank’s August release. For the new layer, Starling’s own payment history is the training set, and the compute sits with a supplier that is not Amazon, Microsoft or Google. What its customers get from that, and what it costs the bank, are the parts the release does not yet show.

What 15 billion transactions are for

The business version of Starling Assistant is free in the app. The same release says it can move a share of earnings into a tax pot on command and talk business owners through invoices from suppliers, flagging signs of a scam such as changed payment details.

A general chat model can explain what invoice fraud looks like. A model trained on a bank’s own transactions can, in principle, see what normal looks like for one account and spot the payment that breaks the pattern. Starling’s release puts its aim more broadly, as a layer that identifies patterns in anonymized transaction data and first improves the bank’s existing fraud detection.

Aser Blanco, global head of banking at Nvidia, said banks are at the forefront of turning large amounts of data into timely, personalized insights while also protecting customers from fraud.

The 14 October release says the new layer will let the assistant offer more personalized help, “such as proactively serving bespoke budgeting guidance”. It does not say whether the business version of the assistant will get the same features, or when.

Behavox went to Civo a week earlier

Starling is not the only company in the finance sector to choose Civo this month. On 7 October Behavox, a London-based compliance software company that serves banks and hedge funds, signed an $8m (£6.1m) commitment for dedicated GPU capacity with Civo so that it could process its customers’ data locally.

“Pairing NVIDIA compute with Civo’s UK sovereign cloud shows that banks don’t have to trade performance or innovation for control of their data, or depend on a handful of global providers,” said Mark Boost, chief executive of Civo, in the Starling release.

Civo has its own numbers on how rare that move still is. Its survey of 1,000 UK IT leaders, published on 30 June, found that 66% would consider switching cloud providers, but that successful migration to a domestic alternative had stalled at 15%. The research is Civo’s own, and Civo sells the alternative it measures. The same survey found the share expecting to become more deeply tied to US hyperscalers had risen from 12% to 28% in a year.

Starling’s deal is a new workload placed with a smaller supplier. The release does not say whether any existing Starling systems are moving, which is the harder step the survey describes.

What the release leaves out

Starling has published no cost for the project, no size for the model and no target for how much more fraud it expects to catch. The “first bank” claim rests on a footnote, which narrows it to the first bank to train such a model on cloud infrastructure owned by a company that is owned and headquartered in the bank’s home country, so that all the data stays under one jurisdiction.

The control Starling gains is over where the data sits and who owns the infrastructure. The model itself is being developed with Nvidia’s financial services team, using Nvidia’s Transaction Foundation Model developer blueprint, its CUDA-X libraries and its NeMo AutoModel training software, on Civo infrastructure that runs Nvidia Hopper GPUs, with Vera Rubin deployments to come. The assistant the model is meant to augment still runs on Google Gemini.

Rees framed the cloud choice as one about suppliers. “We also want to take a leadership role in the development of home-grown cloud infrastructure, which is essential to European companies’ ability to diversify their suppliers and reduce concentration risk,” she said.

The test comes in 2027, when the model is due in production: whether fraud flags trained on Starling’s own ledger stop more bad payments than the tools the bank uses today. Until Starling publishes that figure, the case for the model rests on the bank’s own description of it.