Santander teams with Nvidia to match AI models to banking tasks

Santander is working with Nvidia to answer a practical banking question: which AI model fits each job, and how cheaply can it run? The collaboration, announced October 19, centers on evaluating models and AI agents. It accompanies Santander’s shift from large, general-purpose models toward smaller systems specialized for particular banking tasks.

A model-selection arena inside the bank

Santander’s AI lab has built Model Arena to evaluate open-weight models against banking tasks. Its purpose is to identify the right model for each workload, rather than make model size the deciding factor. The collaboration also covers capabilities for evaluating AI agents, alongside the models themselves.

The bank is also using Nvidia NeMo Customizer to specialize models. These are two related parts of its approach: Model Arena evaluates candidates against the work, while customization adapts models for banking. Santander says its lab’s work is designed to help it select suitable systems and deploy them efficiently.

Aser Blanco, Nvidia’s global banking lead, described Santander’s approach as using smaller models specialized on the bank’s own data and evaluated against real banking tasks. His account emphasizes training and fine-tuning—the work of further adapting a model—rather than relying on size alone to deliver useful banking capabilities.

The savings come from selected experiments

The most striking cost figure comes from a Santander news release provided to PYMNTS. The bank said selected experiments with its smaller, specialized-model approach reduced serving costs—the cost of running a model to handle an interaction—from €5 to approximately €0.03 per interaction, while maintaining what it described as strong performance.

The figures describe selected experimental results, not a bank-wide measure of savings. Santander also reported an experimental customer-service use case in which deploying a more efficient model improved conversation quality and lowered costs by more than 50%. That example pairs a claimed quality improvement with lower spending, rather than presenting cost reduction as the only outcome.

Blanco framed the economic goal as “cost per intelligence output.” His statement places the emphasis on what a system produces for the money spent, with specialization, banking-task evaluation and efficient serving all contributing to that goal. The performance and savings figures remain Santander’s claims.

Fraud detection and simulation are next areas of work

Santander and Nvidia say they will continue working together beyond model selection and specialization. Their stated areas of work span the underlying models, systems that act as agents, banking decisions and simulation. The companies identify several directions for that continued collaboration:

• Foundation models and AI agents, extending the collaboration’s work on the systems used for banking tasks.
• Fraud and anti-money-laundering capabilities, alongside real-time decision-making.
• Advanced simulation technologies, including banking digital twins.

Those are planned areas of collaboration, not completed product launches. The concrete work described in the announcement is the bank’s evaluation environment, its use of Nvidia’s customization tools and the experimental results it attributes to choosing more efficient models.