Summary

• Starling is developing a Transactional Foundation Model using more than 15 billion anonymised transactions and Nvidia technology.
• Training and hosting will use UK-owned Civo infrastructure, reducing reliance on major international cloud providers without eliminating overseas technology dependencies.
• The bank targets production in early 2027, beginning with fraud detection before extending the model to Starling Assistant.

Starling Bank is developing an AI model using more than 15 billion anonymised transactions on cloud infrastructure operated by British provider Civo, putting the financial sector’s concerns about dependence on international technology suppliers into a substantial commercial development programme.

The bank is working with Nvidia to build a Transactional Foundation Model that can identify patterns across financial activity and support several applications through a shared system called Starling Intelligence.

Starling intends to put the model into production in early 2027, initially improving fraud detection before extending its capabilities to Starling Assistant, the bank’s conversational financial assistant. The proposed system could eventually provide more personalised budgeting information and other services informed by customer transaction behaviour.

Training and hosting will take place on infrastructure operated by Civo, a UK-owned and headquartered cloud company. Starling describes the project as the first banking transactional foundation model trained and hosted on sovereign UK cloud infrastructure, using a definition that combines British ownership with domestic data jurisdiction.

The bank’s decision follows growing scrutiny of the concentration of cloud and AI infrastructure among a small number of global suppliers, particularly where regulated institutions depend on services that would be difficult to replace following an outage, commercial dispute or change in operating conditions.

Learning from transaction patterns

Foundation models are generally trained on large datasets so their learned representations can support multiple applications. While many well-known models process language or images, Starling’s system is being developed around financial transactions.

Payment records contain information about amounts, timing, merchant categories and sequences of activity. Analysing relationships between those events may help identify changes in spending behaviour or patterns associated with potentially fraudulent transactions.

For example, a payment may be unusual for one customer but entirely consistent with another person’s activity. A model examining broader sequences could potentially provide additional context beyond fixed rules applied to individual transactions.

That flexibility introduces its own limitations. Unusual financial behaviour does not necessarily indicate fraud, and a pattern that was previously normal may change because of travel, employment, household circumstances or business activity.

Fraud detection systems must balance identifying suspicious activity against the disruption caused when legitimate payments are incorrectly flagged or prevented. Their performance consequently needs to be assessed through missed fraud, false alerts and the effects on customers.

Starling has not published independent accuracy comparisons for the proposed model or disclosed how its outputs will be incorporated into final decisions about particular transactions.

The training dataset also requires careful governance. Although the bank says it will use anonymised information, financial transactions can contain combinations of characteristics that are distinctive even when names and direct identifiers have been removed.

The effectiveness of anonymisation depends on the transformation applied to the data and whether individuals could be identified through other information. Access restrictions, data minimisation and safeguards over model outputs remain relevant even when training material is described as anonymous.

UK hosting reduces one form of dependence

Starling has chosen Civo to supply the cloud infrastructure on which the model will be trained and hosted. The platform uses Nvidia GPUs and associated software, providing computing resources for processing large datasets and running AI workloads.

The planned development will use Nvidia’s Transaction Foundation Model developer blueprint, CUDA-X libraries for accelerated data processing and NeMo AutoModel for model training.

These components establish the proposed technical approach, although the announcement does not disclose training costs, final model size or independently verified comparisons against competing financial AI systems.

Harriet Rees, Starling’s group chief information officer, contributed to the UK government’s Financial Services AI Adoption Plan, published in July 2026. That work identified concentration risk arising from financial institutions’ dependence on a limited number of cloud and AI providers.

Choosing a domestic supplier gives Starling another infrastructure option and places the relevant hosting services under a UK-owned company. It may also provide greater control over some contractual and jurisdictional arrangements.

However, UK cloud ownership does not make the computing stack entirely British. Civo’s platform relies on Nvidia hardware and software, and its operation will continue to depend on international semiconductor manufacturing and technology supply chains.

Data location is similarly only one component of operational resilience. Service availability depends on computing capacity, network connectivity, security controls, backup arrangements, incident response and access to replacement hardware.

A smaller domestic provider may offer advantages in flexibility or governance, but a regulated customer must still examine its ability to maintain service during disruption and to support increasing workloads over time.

Those requirements are especially demanding for banks, where customers expect essential services to remain available and where operational incidents can affect payments, account access and regulatory obligations.

Starling has not disclosed the commercial terms of its agreement with Civo or quantified the expected cost difference compared with equivalent infrastructure from international providers.

Production performance remains untested publicly

The model is intended to become a reusable intelligence layer beneath several Starling applications. A shared system could potentially reduce the need to develop separate underlying models for fraud detection, customer insights and other financial services.

Reusing a model across applications does not remove the need for application-specific evaluation. A system that identifies useful patterns in spending may not satisfy the reliability requirements of fraud detection or the safeguards necessary for personalised financial recommendations.

Starling Assistant introduces further considerations because an AI system interacting directly with customers must interpret requests, retrieve appropriate account information and remain within the permissions granted to it.

Personalised guidance could help customers understand spending patterns or identify budgeting opportunities, but the resulting information must be accurate and presented in a way that does not confuse general assistance with regulated financial advice.

Systems capable of taking actions on a customer’s behalf require additional controls over authorisation, confirmation and recovery when an operation fails or a request is misunderstood.

Starling plans to introduce the Transactional Foundation Model into production in early 2027. Its first operational applications will therefore provide evidence about reliability, security and the effectiveness of the underlying data processing architecture.

The deployment will also test whether Civo can support an evolving banking AI workload with sufficient capacity and continuity, while maintaining the domestic hosting and governance arrangements that motivated Starling’s choice.

Results from that programme may help establish where sovereign cloud infrastructure offers a practical alternative to larger providers. The model’s commercial contribution will depend on measurable improvements in banking services, the cost of operating it and the reliability of the infrastructure on which those services depend.