Starling is developing a transactional foundation model (TFM) with support from NVIDIA, with training and hosting carried out on infrastructure operated by UK cloud provider Civo. According to the bank, the TFM will be the first banking model to be trained and hosted within a UK sovereign cloud.
The model will be trained on more than 15 billion anonymised transactions from Starling's proprietary data, with the aim of giving the bank a more detailed view of customer needs and behaviours. Over time, it is set to form a single intelligence layer, named 'Starling Intelligence', which will underpin the bank's full stack of AI capabilities and support the development of new customer experiences.
From separate AI tools to a shared intelligence layer
Starling has already deployed several proprietary AI tools, including Starling Assistant, which the bank describes as the UK's first agentic AI financial assistant, and says it rolls out new features every week. The TFM is intended to connect these tools through one layer trained on the specific needs of the bank's customers.
The bank plans to put the model into production in early 2027. It will first be applied to existing AI services for fraud detection and will then be used to extend Starling Assistant. Once the Starling Intelligence layer is in place, the assistant is expected to offer more personalised functions, such as proactively providing tailored budgeting guidance.
Technology stack
The TFM is being developed with NVIDIA's financial services team using the company's Transaction Foundation Model developer blueprint. This includes NVIDIA CUDA-X libraries for GPU-accelerated data processing and NVIDIA NeMo AutoModel for model training.
Civo's sovereign infrastructure runs on NVIDIA Hopper GPUs, with Vera Rubin deployments upcoming, and uses liquid cooling to provide the low-latency compute required for real-time reasoning, autonomous systems, and agentic AI. The resulting layer is designed to identify patterns in anonymised transaction data and to be reused across multiple applications. Aser Blanco, Global Head of Banking at NVIDIA, said that combining the company's computing and software with Civo's UK cloud allows Starling to deliver personalised services while keeping control of customer data.
Concentration risk and the policy context
The decision to train the model on domestic infrastructure follows recommendations set out in the UK Government's Financial Services AI Adoption Plan, published in July 2026. The plan was co-authored by Starling's Group CIO, who is also a UK Government AI Champion for Financial Services. It flags the financial sector's reliance on a small number of global AI and cloud providers and recommends that government and industry broaden their partnerships to improve optionality and systemic resilience.
Harriet Rees, Group CIO at Starling and a UK Government AI Champion for Financial Services, said the bank also aims to take an active role in developing home-grown cloud infrastructure, describing it as essential to European companies' ability to diversify suppliers and reduce concentration risk. Mark Boost, CEO of Civo, added that pairing NVIDIA compute with a UK sovereign cloud shows banks do not have to give up performance or control of their data, or depend on a handful of global providers.
For the wider UK banking sector, the project is an example of the adoption plan's recommendations being applied in practice: a bank training a proprietary model on its own transaction data within domestic cloud infrastructure, rather than relying solely on the small number of global providers the plan identifies.
