IA en un minutoNewsBusiness and products
Santander and NVIDIA agree a three-year collaboration framework to advance specialized AI models and agents for banking
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
Santander and NVIDIA will work together for three years on foundation models, agents, fraud prevention, real-time decisions and simulation, including a banking digital twin. The bank is testing small, specialized models at the Santander AI Lab and compares them on its Model Arena platform. According to Santander, in tests the cost per interaction fell from about 5 euros to about 0.03 euros.
Banco Santander and NVIDIA have signed a collaboration framework for the next three years, the bank announced from Madrid and Santa Clara, California. It spans foundation models, AI agents, the fight against fraud and money laundering, real-time decisions and advanced simulation, with a banking digital twin as one example.
The agreement builds on the work of the Santander AI Lab, where the bank tries out emerging AI technologies before taking them to the rest of the group. With NVIDIA's support, the lab has created tools to assess models and agents meant for banking and to decide which one suits each task. The main one is Model Arena, a platform that measures open-weight models on banking tasks; the bank is extending it to build, simulate, evaluate and oversee agents, and the lab also stress-tests conversational AI in controlled settings.
Santander is shifting from large generalist models to smaller ones specialized in specific banking tasks. It fine-tunes them with NVIDIA NeMo Customizer and, for core banking processes, works on NVIDIA Kumo, open-weight foundation models designed for structured data. To run them inside its own infrastructure it uses NVIDIA Dynamo-Triton, the technology the bank credits for the lower costs.
According to the bank, in some experimental use cases the running cost has come down from around 5 euros per interaction to roughly 0.03 euros while keeping good results. In a customer service test, also experimental, a more efficient model improved the quality of conversations and cut costs by more than 50%.
Santander is "uniquely positioned" to turn those advances into "practical banking capabilities," said Iñaki Bernal, the bank's chief artificial intelligence officer. Aser Blanco, Global Banking Lead at NVIDIA, argues that getting value from a bank's data "is not about model size, but about training and fine tuning," and that the cost of each AI output has become a key metric.
The release does not give the value of the agreement or a timetable for extending these models across the group.
Why it matters · analysis and opinion
More than the alliance itself, what the release reveals is a change of criteria: the bank picks each model by what it costs to get a useful answer, not by its size. If that approach spreads, the advantage will lie in having well-organized proprietary data and a serious method for comparing models, rather than in hiring the biggest one. It should be read with caution, because the savings come from experimental cases and there is no timetable for bringing them to the group's everyday service. The idea does work on a small scale: before paying for the most powerful model, a company can try several more modest ones on its own tasks and compare cost and quality with data, not hunches.
Official source: Banco Santander · Written with the help of AI: how we make the news
