Santander and Nvidia Build AI Banking Lab, Cutting Costs From €5 to €0.03
Santander's AI lab with Nvidia tests smaller specialized banking models, cutting per-interaction costs from €5 to about €0.03 in experiments.
What happened
Santander has teamed with Nvidia to build new capabilities for evaluating AI models and agents for banking. Announced Monday (Oct. 19), the work centers on the Santander AI lab, which is designed to determine which AI model is best suited for a given banking task — and to deploy it in the most efficient way possible.
The headline number: by moving from large general-purpose models to smaller, specialized models designed for specific banking tasks, Santander cut serving costs from €5 per interaction to approximately €0.03 in selected experiments — while maintaining strong performance. In one experimental customer-service use case, the bank improved conversation quality while lowering costs by more than 50%.
The "Model Arena"
A core piece of the effort is a "Model Arena" the bank built to evaluate open-weight models against real banking tasks, determining the right model for each workload. Santander is also specializing models using Nvidia's NeMo Customizer — fine-tuning open models on the bank's own data and requirements rather than relying on general-purpose giants.
The partnership extends beyond cost savings. The companies say they will continue working on foundation models, AI agents, fraud and anti-money laundering capabilities, real-time decisioning, and advanced simulation technologies such as banking digital twins.
The cost-per-intelligence metric
Aser Blanco, Nvidia's global banking lead, put the strategic point plainly: in this new world, cost per intelligence output is the key metric. Santander's approach — smaller models, specialized on the bank's own data, evaluated against real tasks — shows what production-scale AI looks like when efficiency, not raw model size, is the goal.
What this means for your business
- Smaller models beat bigger ones on ROI. You do not need the largest frontier model for most business tasks. A smaller, specialized model fine-tuned on your data can match performance at a fraction of the cost — in Santander's case, roughly 1/170th. - Build your own evaluation arena. Before committing to any model, test candidates against your actual tasks the way Santander does. The winning model is the one that handles your workload reliably and cheaply, not the one topping generic leaderboards. - Your data is the moat, not the model. Fine-tuning open models on proprietary data turns that data into competitive advantage. The model becomes a commodity; the training and tuning on your business is what competitors cannot copy. - Attack the data problem first. PYMNTS research found 30% of financial services firms cite data quality and fragmentation as their main obstacle to further AI deployment. Clean, unified data pays for itself before any model investment.
What to watch
Whether the "Model Arena" approach becomes a template for enterprise AI buying, and how far the cost curve keeps bending as specialized models get better.