Madrid – October 31, 2026 -- Banco Santander cut costs by more than 50% in an experimental customer-service use case while improving conversation quality, and has now signed a three-year AI collaboration framework with NVIDIA.
Smaller models cut serving costs from €5 to €0.03 per interaction
Santander AI Lab has moved progressively from large general-purpose models to smaller, specialised models built for specific banking tasks. In selected experimental use cases, serving costs fell from around €5 per interaction to approximately €0.03, with performance maintained.
The Lab runs inference within Santander's own architecture on NVIDIA Dynamo-Triton, the approach behind the serving-cost reduction.
Model Arena tests open-weight models against banking workloads
With NVIDIA's support, the Lab built Model Arena to evaluate open-weight models and AI agents against banking tasks and match the best model to each workload. It is also being developed to design, simulate, evaluate and supervise banking agents. Santander specialises models with NVIDIA NeMo Customizer.
For core banking workloads, the bank built specialised intelligence on NVIDIA Kumo, a family of open-weight foundation models for structured data. Customer-service comparisons showed that smaller, specialised models can perform strongly on targeted tasks.
The Lab also conducts performance engineering and adversarial testing of conversational AI in controlled environments. Successful capabilities are then meant to scale across the bank.
Three-year framework covers fraud, AML and a banking digital twin
The next phase spans foundation models, AI agents, fraud and anti-money laundering capabilities, real-time decisioning and advanced simulation, including a banking digital twin.
Iñaki Bernal, Santander's chief artificial intelligence officer, said the bank's scale across businesses, customers and geographies leaves it well placed to turn AI advances into practical banking capabilities.
Aser Blanco, Global Banking Lead at NVIDIA, said model size matters less than training and fine-tuning on a bank's own data. "Cost per intelligence output is a key metric," he said, adding that Santander is demonstrating this in production with smaller models evaluated against real banking tasks.