Starling Bank claims first banking AI trained on UK sovereign cloud
Starling Bank has claimed to be the first bank to train an artificial intelligence model on a UK sovereign cloud. It is developing its own transactional foundation model (TFM) with Nvidia and will train it on infrastructure from UK cloud provider Civo.
The TFM will train on more than 15 billion anonymized transactions and other events from Starling’s systems. In time it will form a single “Starling Intelligence” layer underpinning the bank’s AI capabilities. Production is planned for early 2027, starting with improved fraud detection and later enhancements to Starling Assistant, the bank’s agentic AI financial assistant.
Sebastian Khan, data scientist and technical lead on the TFM, explained that the model is purpose-built to understand transactions rather than language. Where a large language model learns word meaning from surrounding context, the TFM learns what a transaction represents from the events around it. He gave the example of a regular morning coffee purchase at home versus the same purchase while on holiday: the surrounding transactions differ even if the coffee itself looks similar.
The model is Starling’s own intellectual property, built in-house while drawing on prior work such as Revolut’s Pragma model, Visa’s Transaction GPTs research and Nvidia’s financial-services blueprints. Training uses tokenized data with no personally identifiable information. Tokenization replaces raw values (for example a postcode) with integers embedded in vector space; the tokenizer map is kept separate so encoded data cannot be reversed without it.
Tokenization currently occurs in Starling’s existing Google Cloud environment before the data reaches Civo. Khan stated this creates no elevated sovereignty risk and is standard practice. Civo chief executive Mark Boost framed the arrangement as part of a multicloud strategy that avoids lock-in to a single provider and supports a British cloud company. The move also aligns with the UK government’s financial-services AI adoption plan, co-authored by Starling group CIO Harriet Rees.
Computer Weekly notes that claims of “sovereign” status based solely on UK cloud providers and UK datacenters rest on a narrower definition than many accept. US laws such as the Cloud Act and FISA Section 702 can reach data held by US companies regardless of storage location. Starling’s data is first processed on Google Cloud (a US hyperscaler) and the model runs on Nvidia silicon and software updated from the United States. A trained model may also encode statistical representations of its training data, so sovereignty is an end-to-end chain.
Boost acknowledged the supply-chain reality that top-performing chips are currently made by American companies. Khan said the bank is about to begin training on real data in Civo after testing the Kubeflow pipeline with synthetic data; the 15-billion-event figure remains preliminary. The first use case is fraud detection, where sequential modelling is expected to improve on traditional feature engineering. Retraining frequency will be guided by drift detection.
Overall the project reduces dependence on a handful of global hyperscalers and builds domestic capability, even if full technical sovereignty remains a longer journey.