Starling Bank announced on 14 October 2026 that it is developing a Transactional Foundation Model with NVIDIA, trained and hosted on Civo’s UK sovereign cloud infrastructure. The project marks the first claimed banking foundation model built on UK-owned sovereign cloud resources. The initiative aims to create a unified intelligence layer called Starling Intelligence that improves the bank’s existing AI tools.
Development details and technical approach
The model trains on more than 15 billion anonymised transactions and events drawn from Starling’s customer base of approximately 5 million individuals and 500,000 SMEs. Data tokenisation takes place on Google Cloud before the datasets move to Civo for training. NVIDIA supplies the Transaction Foundation Model developer blueprint, CUDA-X libraries, and NeMo AutoModel framework, while Civo provides NVIDIA Hopper GPUs with planned support for future Vera Rubin architectures.
Initial production deployment is scheduled for early 2027 and will first target fraud detection. Subsequent integration will connect the model to Starling Assistant, the bank’s existing agentic AI tool, to deliver proactive budgeting guidance and other personalised services. Starling retains full intellectual property rights over the finished model.
Group CIO Harriet Rees, who also serves as UK Government AI Champion for Financial Services, stated that the project reduces reliance on overseas providers and supports domestic cloud capacity in line with the July 2026 UK Financial Services AI Adoption Plan. Civo CEO Mark Boost and NVIDIA Global Head of Banking Aser Blanco both confirmed the performance and data-control benefits of sovereign infrastructure for financial workloads.
Strategic context and next steps
Starling builds on earlier internal experiments and external research such as Revolut’s Pragma work and Visa foundation-model studies. The choice of Civo addresses regulatory and operational concerns about data residency while maintaining access to high-performance NVIDIA hardware. The bank expects the new layer to improve accuracy of behaviour-pattern detection without increasing exposure to global cloud concentration risks.
Further milestones include expanded use cases beyond fraud and personal finance, with continued evaluation of model performance on Civo’s infrastructure. Starling has indicated it will publish additional technical benchmarks once the first production workloads run in 2027.
Frequently asked questions
What data does the Starling transactional foundation model use?
The model trains on more than 15 billion anonymised transactions and events from Starling’s customer base. Tokenisation occurs on Google Cloud before training moves to Civo’s infrastructure.
When will Starling’s model enter production use?
Initial production deployment is planned for early 2027, beginning with fraud detection before integration with Starling Assistant for personalised customer features.
Why did Starling choose Civo’s sovereign cloud?
Civo provides UK-based sovereign hosting on NVIDIA Hopper GPUs, reducing concentration risk on global providers and aligning with the UK Financial Services AI Adoption Plan.
Does Starling own the intellectual property of the new model?
Yes. Starling states the Transactional Foundation Model remains its own IP, developed in-house while incorporating elements from prior internal work and external research.
By YasTech Editorial
Comments 5