Starling Bank claims first banking AI trained on UK sovereign cloud

You have probably seen a bank promise to spot fraud faster or answer questions more naturally. The promise usually rests on a large language model trained somewhere else, on someone else's data, in a jurisdiction the bank does not control. Starling Bank has announced a different path: a transactional foundation model trained on more than 15 billion anonymised events from its own systems, running on Civo's UK cloud infrastructure with Nvidia silicon. The bank calls it the first banking AI trained on a UK sovereign cloud. Production is due in early 2027, first to sharpen fraud detection and then to improve Starling Assistant, the bank's agentic financial assistant. In time the model will form a single Starling Intelligence layer underpinning the bank's AI capabilities.

What a transactional foundation model does

Sebastian Khan, data scientist and technical lead on the TFM at Starling, said the model differs from an off-the-shelf large language model because it is "tailor-made to crunch and understand transactions". Where an LLM learns the meaning of a word from the sentence around it, a TFM learns what a transaction represents from the events around it. Khan offered an example: a regular morning coffee in your hometown at 9am every day, then a morning coffee at 9am while you are on holiday. The transaction is the same, the price similar, but the events surrounding it are probably quite different. The model learns that context.

Khan said the model is Starling's own intellectual property, 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. Starling does not use personally identifiable information. Tokenisation is itself a form of anonymisation, with raw values such as a postcode replaced by an integer embedded into a vector space. Khan said the bank keeps the tokeniser – the map from raw to encoded data – separate. Even if someone gained 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. The move forms part of a multicloud strategy. Mark Boost, chief executive of Civo, said everyone should live in a multicloud strategy to some degree, because they should not be locked into a single provider. Civo wants 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

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, he said, is: can someone turn it off? These are silicon chips that sit in a datacentre in the UK, and Nvidia cannot turn them off as such.

The claim that a model is "sovereign" because it runs on a UK cloud provider in a UK datacentre rests on a narrower definition of sovereignty than many recognise. Computer Weekly's reporting on cloud sovereignty found that many such claims rest on corporate policy and marketing rather than verifiable technical mechanisms. The decisive question is not where the servers sit, but whose jurisdiction can reach them.

Under the US Cloud Act and FISA Section 702, US authorities can compel US companies to hand over data regardless of where it is stored. Under 18 USC 2705(b), courts can attach non-disclosure orders – gag orders – to those demands, forbidding the provider from telling its customer anything has happened.

First, Starling tokenises its data on Google Cloud, a US hyperscaler, before it ever reaches Civo. Second, the model runs on Nvidia silicon and Nvidia's software stack, updated from the US. A global supply chain where code is authored and signed at a US headquarters makes the routine software update a potential invisible vector for state-level intervention.

A growing body of argument holds that a trained model encodes statistical representations of its training data, so sovereignty must be treated as an end-to-end chain: the raw data, the training pipeline, the model weights, the inference outputs and the derived insights. A single weak link breaks the chain, and transformation such as tokenisation does not strip jurisdiction.

Training on a UK-owned cloud reduces dependence on a handful of US hyperscalers and builds domestic capability. But "sovereign cloud" and "all data within UK jurisdiction only" describe a step along a journey rather than a destination.

The argument from Starling and Civo is that sovereignty comes from the cloud provider rather than the silicon. 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 are the top performing – and there is not a lot of choice elsewhere.

What happens next

Khan said the bank was "about to start training on real data in Civo", having spent several months building the codebase in-house. Starling first tested 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. Khan is an expert in drift detection, so knowing when training data and model predictions start to become out of date.

For L&D and digital transformation leaders, the lesson is not the sovereignty claim but the ownership of the model. Starling built the capability in-house, trained it on its own data and chose the infrastructure that matched its risk appetite and regulatory posture. The model is not a black box rented from a foundation-model vendor. It is an asset the bank controls, updates and audits. That is the capability worth building, whether the cloud is sovereign or simply less concentrated.

Sources:

• Computer Weekly, 14 October 2026