TL;DR
Starling Bank is training a Transactional Foundation Model with Nvidia on infrastructure run by Civo, a UK-owned cloud provider, using over 15 billion anonymised customer transactions and keeping the data inside UK jurisdiction. The bank expects the model in production in early 2027, starting with fraud detection before it feeds Starling Assistant. The move is framed as a response to the Treasury’s July adoption plan on cloud concentration, and it arrives the same week the Financial Stability Board told banks to manage dependence on third-party AI suppliers.
No other bank, Starling claims, has trained and hosted a model of this kind on cloud owned by a company headquartered in the UK. It describes the resulting system, “Starling Intelligence”, as a shared intelligence layer across its AI products, with personalised budgeting guidance among the later uses.
Who does what
Nvidia supplies the transaction-model developer blueprint and its data-processing and training software, with support from its financial services team. Civo supplies the GPUs and the UK-based hosting for both training and serving. Starling brings the data and will own the model. All three partners pitch the arrangement as evidence that heavy AI workloads need not default to the biggest global clouds.
Civo’s chief executive Mark Boost argued that AI “will only deliver lasting value for the UK economy if it’s built on secure, resilient, sovereign infrastructure”. Harriet Rees, who is Starling Group’s CIO and also the government’s financial-services AI champion, said the bank wanted a “leadership role in the development of home-grown cloud infrastructure” so that European firms could diversify suppliers and cut concentration risk.
The policy backdrop
Starling says the approach follows the Treasury’s AI adoption plan for financial services, published in July, which Computing summarises as flagging the sector’s reliance on a handful of global cloud and AI suppliers and calling for broader supplier relationships. The same theme runs through the FSB’s final sound practices, published on 15 October, whose twelfth practice covers dependence on third-party AI providers.
Computing’s own reading is that the project tests whether smaller sovereign providers can carry enterprise-scale AI compute while giving firms more control over where data sits.
Looking forward
In our view the interesting question is not whether Civo can host a training run, but whether a model built on one bank’s 15 billion transactions outperforms the general-purpose fraud tooling that the hyperscalers sell. If it does, the sovereign-cloud argument gains a commercial case to go with the regulatory one, and other UK banks will have a reason beyond concentration risk to follow. The early-2027 production date gives a clear point at which to judge it.