Starling will train a model on more than 15 billion of its own transactions in a UK cloud

Starling's own contribution to the build is the data: more than 15 billion anonymised transactions ^[2]. Of the five technology components the source names (the developer blueprint, the CUDA-X libraries, NeMo AutoModel, Hopper GPUs and the upcoming Vera Rubin deployments), all five are NVIDIA's ^[7][8][13]. Civo operates the infrastructure they run on ^[1]. On this evidence, Starling is training its model from a blueprint, on its own data ^[7][2].

The sovereign hosting answers a specific worry. The Financial Services AI Adoption Plan, published in July 2026 and co-authored by Starling's Group CIO, flags the sector's reliance on a small number of global AI and cloud providers ^[5][6]. That CIO, Harriet Rees, said the bank also wants an active part in building home-grown cloud infrastructure, and called it essential if European companies are to diversify suppliers and reduce concentration risk ^[9]. Mark Boost, Civo's chief executive, said combining NVIDIA's compute with a UK sovereign cloud demonstrates that banks can keep both performance and control of their data without relying on a handful of global providers ^[10]. That holds cleanly for the cloud layer, but the source shows one supplier at the GPU layer ^[13].

The layer may improve fraud detection and then carry Starling Assistant into proactive, tailored budgeting guidance, the sequence the bank lays out ^[4][11]. Or it may improve fraud scoring and stop, leaving Starling Intelligence as a name for one model. Or it may work and prove easy to copy: if the blueprint is open to other banks, the 15 billion transactions are the only input that is Starling's alone ^[2].

We think the first is the likeliest, because fraud detection gives the model an existing production service to be measured against ^[4]. The counter-thesis is that this is one bank, and its CIO helped write the plan it is applying ^[5], so the project says more about Starling than about the sector. Thepaypers calls it an example of the plan's recommendations applied in practice ^[12].

The test arrives in early 2027 ^[4]. If the model does not improve on the fraud services already running, we would call the project a sovereignty exercise with no product gain, and the single intelligence layer would have nothing to sit on. The report does not include the cost, the GPU count or any performance target, so the only measure available then will be whatever Starling chooses to publish.