Banco Santander said Oct. 19 that using smaller models for banking tasks cut AI serving costs in selected tests, as it established a three-year framework for further work with NVIDIA.
The bank reported a cost of about €0.03 per interaction, down from roughly €5, and said performance remained strong. It described those figures as results from selected experiments.
The announcement also cites customer service. In an experimental use case, Santander reported more than 50% lower costs alongside improved conversation quality after using a more efficient model.
Santander’s AI lab built Model Arena to compare open-weight models on banking tasks. The bank is using NVIDIA NeMo Customizer to adapt models for specific work and tests its systems before considering wider deployment. It linked the lower serving cost to both the shift toward smaller, task-specific models and NVIDIA Dynamo-Triton within its infrastructure.
Under the three-year framework, Santander and NVIDIA plan work on foundation models and AI agents, alongside fraud and anti-money laundering capabilities. The collaboration also covers real-time decisions and simulation tools, including a banking digital twin.
