Starling Bank has laid claim to be the first bank to train an artificial intelligence (AI) model on a UK sovereign cloud, after it announced it would develop its own transactional foundation model (TFM) with Nvidia and train it on Civo’s UK cloud infrastructure.
The TFM will train on more than 15 billion anonymised transactions and other events from Starling’s own systems, and in time will form a single Starling Intelligence layer underpinning the bank’s AI capabilities. It is due in production in early 2027, first to sharpen fraud detection and then to improve Starling Assistant, the bank’s agentic AI financial assistant.
Sebastian Khan, data scientist and technical lead on the TFM at Starling, said the model differs from an off-the-shelf large language model (LLM) because it is “tailor-made to crunch and understand transactions”. Where an LLM learns the meaning of a word from the sentence around it, he said, a TFM learns what a transaction represents from the events around it.
“Let’s say you have a regular morning coffee in your hometown at 9am every day, but maybe you go on holiday, and you have a morning coffee at 9am while you’re away,” he said. “It’s the same coffee transaction for the same-ish price, but the transactions surrounding it are probably going to be quite different.”
Khan said the model is Starling’s own IP, developed “from the ground up completely in-house” but “standing on the shoulders of giants”, including Revolut’s Pragma model, Visa’s Transaction GPTs research and Nvidia’s financial-services blueprints.
The model trains on tokenised data. Khan said Starling does not use personally identifiable information, and that tokenisation is itself a form of anonymisation, with raw values such as a postcode replaced by an integer embedded into a vector space.
“We keep separate the tokeniser, which is basically the map from the raw to the encoded data,” he said. “Even if someone was to get access to the encoded data, there would be no way of decoding it without the tokeniser.”
That tokenisation happens not on Civo’s UK infrastructure but in Starling’s existing Google Cloud environment, where the bank’s data warehouse lives. Khan said this carried no additional data sovereignty risk. “There’s no elevated risk,” he said. “It’s sort of standard business practice.”
The move forms part of a multicloud strategy, said Mark Boost, chief executive of Civo. “Everyone, to some degree, should live in a multicloud strategy, because they shouldn’t necessarily be locked into a single provider,” he added. “We just want to be one of those cloud providers in a multicloud strategy.”
Starling, he said, was “keen to back a British provider”, and the UK should “be less of a taker and more of a maker, as Keir Starmer said last year”. The move also echoes the UK government’s financial services AI adoption plan, co-authored by Starling group CIO Harriet Rees, a UK government AI champion for financial services, which flags the sector’s reliance on a small number of global cloud and AI providers.
Where the sovereignty line is drawn is the most contested part of the announcement, and Boost’s framing is one of control. “The most important question is, can someone turn it off? Is there a digital kill switch?” he said. “These are silicon chips that sit in a datacentre in the UK, and Nvidia can’t turn it off as such.”
The argument from Starling and Civo is that sovereignty comes from the cloud provider rather than the silicon. So, because Civo is UK-owned and its datacentres sit in the UK, the data sits only under UK jurisdiction.
But that leaves Nvidia, a US company, supplying both the hardware and much of the software stack, with the CUDA ecosystem updated from the US.
Boost acknowledged the supply-chain reality: “In a perfect world, every single thing is kept in the UK, built in the UK. But realistically, the top chips are being made by American companies right now – they’re the top performing – there’s not a lot of choice elsewhere.”
As Computer Weekly reported on the limits of data sovereignty, US authorities can reach data held by US companies wherever it sits.
Khan said the bank was “about to start training on real data in Civo”, having spent several months building the codebase in-house, and was first testing the pipeline – a Kubeflow pipeline on Civo – with synthetic data. The target is more than 15 billion events, “but that figure is preliminary”.
The first planned use case is fraud detection, where existing models rely on what Khan called “traditional feature engineering, which is sort of a dark art”. A model that captures sequential information could deliver “a boost in fraud detection”, he said, before work turns to Starling Assistant. How often the model will need retraining remains open.
“I’m actually an expert in drift detection, so knowing when training data and model predictions start to become out of date,” Khan said.
