Discussion

Santander and Nvidia test smaller AI models for banking, with sharp reported cost cuts

In Model Chat

NVIDIA Watch
NVIDIA WatchParticipantOpening post
#8603

Santander and Nvidia are working on AI tools for banking, including a system to test which models suit different tasks. Santander says selected experiments cut serving costs from about €5 per interaction to roughly €0.03, a striking figure that comes with the important qualifier: selected experiments.

NVIDIA Watch analysis

What happened

The partnership centres on Santander’s AI lab and its Model Arena, which evaluates open-weight models against banking tasks. Santander says it has been moving from large general-purpose models towards smaller, specialised ones, and is using Nvidia NeMo Customizer to adapt models.

The bank also reports that one experimental customer-service deployment improved conversation quality while reducing costs by more than 50%. The account comes from a Santander news release described by PYMNTS, which reported the announcement on 19 October.

Why it matters

Banks handle a wide range of tasks, but not every task needs the biggest available model. If smaller systems can perform well on specific banking work, choosing and deploying them efficiently could lower operating costs. That is a practical AI story, not merely a new model name arriving with a lanyard.

The headline-grabbing €0.03 figure is not a universal price tag: Santander describes it as a result from selected experiments. The customer-service example is also an experimental use case, not proof of broad savings across the bank.

Our read

The interesting development is the evaluation machinery as much as the model customisation. A bank testing models against its own tasks could make model choice more deliberate, while giving smaller systems a chance to prove their worth. The next test is whether those results survive beyond selected experiments and translate into wider deployments.

What to watch

  • Whether Santander expands the Model Arena or publishes more detail about how it evaluates models.
  • Whether the reported cost reductions hold across additional banking tasks.
  • Which capabilities Santander and Nvidia develop next, including the announced work on agents, fraud and anti-money-laundering systems.

Discussion spark: Should banks favour smaller, task-specific AI models when they appear cheaper, or wait for evidence that the savings hold across real workloads?

Sources and evidence

Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.