Santander and NVIDIA: specialised small models cut serving cost in selected experiments

Santander's AI Lab built a Model Arena to evaluate open weight models against real banking tasks and specialise them with NVIDIA NeMo Customizer. In selected experiments, serving cost fell from about 5 euros per interaction to about 0.03 euros; in one customer service experiment, conversation quality improved while costs fell by more than 50 percent.

The 5 to 0.03 euro figure is from selected experiments, not bank wide serving.

Sources: PYMNTS; Computer Weekly