Starling Bank has announced a major step in UK financial technology by partnering with NVIDIA (NASDAQ: NVDA) to develop a specialized Transactional Foundation Model (TFM).
This initiative marks the first time a banking AI model of this type will be both trained and hosted entirely within a UK sovereign cloud environment, using infrastructure from the British provider Civo.
The move underscores a growing emphasis on data residency and reduced reliance on a limited set of global technology suppliers in the financial sector.
Unlike general-purpose large language models that process text, the TFM is designed specifically to interpret sequences of financial events.
It will learn patterns from Starling’s own proprietary dataset of more than 15 billion anonymised transactions and related events.
Over time, this model is intended to serve as a unified “Starling Intelligence” layer that underpins the bank’s broader suite of AI tools, enabling deeper insights into customer behavior and supporting the creation of new services.
The project draws on NVIDIA’s financial services expertise and its Transaction Foundation Model developer blueprint.
This includes GPU-accelerated processing libraries and tools for model training, allowing efficient handling of high-volume transactional data.
Civo supplies the underlying sovereign cloud platform, featuring NVIDIA Hopper-class GPUs with plans for future Vera Rubin systems, advanced liquid cooling, and low-latency capabilities suited to real-time reasoning and agentic applications.
By locating training and hosting on UK-owned and headquartered infrastructure, Starling aims to keep relevant data within UK jurisdiction.
Group CIO Harriet Rees, who also serves as a UK Government AI Champion for Financial Services, framed the effort as both a customer-focused and strategic resilience measure.
She noted that the bank has already introduced the UK’s first agentic AI financial assistant and regularly adds new features; the new intelligence layer will connect these capabilities more tightly to the specific patterns in Starling’s customer base.
Rees also highlighted the broader goal of expanding domestic cloud options so that European firms can diversify suppliers and mitigate concentration risks arising from dependence on a small number of global providers.
This approach aligns with recommendations in the UK Government’s Financial Services AI Adoption Plan, published in July 2026 and co-authored by Rees.
The plan identifies systemic vulnerabilities from heavy reliance on a handful of overseas AI and cloud platforms and calls for wider partnerships to improve optionality.
Civo CEO Mark Boost described the collaboration as evidence that high-performance AI workloads can be delivered on homegrown infrastructure without sacrificing control or innovation.
NVIDIA’s Global Head of Banking, Aser Blanco, emphasised that combining accelerated computing with a UK sovereign cloud allows banks to generate timely, personalised insights and strengthen fraud protections while retaining oversight of customer data.
Starling plans to bring the TFM into production in early 2027.
Initial deployment will focus on enhancing existing fraud-detection systems. Subsequent work will integrate the model with Starling Assistant, enabling more tailored interactions such as proactive, personalised budgeting guidance.
The announcement reflects wider momentum in the UK toward sovereign AI capabilities, particularly in regulated industries where data location, operational resilience, and supplier diversity carry both regulatory and commercial weight.
By building a domain-specific model on domestic infrastructure, Starling positions itself at the intersection of advanced AI development and national technology sovereignty.