Key Takeaways
- Santander and NVIDIA have deepened their AI partnership with a three‑year framework that targets foundation models, AI agents, fraud detection, real‑time decisioning and banking‑digital‑twin technologies.
- The Santander AI Lab is shifting from large, general‑purpose models to smaller, task‑specific models, cutting inference costs from roughly €5 per interaction to about €0.03 in experimental customer‑service scenarios while preserving or improving conversation quality.
- Tools such as Model Arena, NVIDIA NeMo Customizer, NVIDIA Kumo, and NVIDIA Dynamo‑Triton enable the bank to evaluate, specialise, and efficiently serve AI models at scale.
- Executives from both organisations stress that the real value lies in turning proprietary data into “intelligence‑centric” capabilities, where cost per intelligence output—not model size—drives decisions.
Overview of Santander‑NVIDIA AI Collaboration
Santander announced a new milestone in its artificial intelligence collaboration with NVIDIA on October 19, 2026, highlighting that the bank’s AI Lab, with NVIDIA’s support, has developed fresh capabilities to evaluate AI models and agents for banking applications. The partnership aims to identify the most suitable model for each workload and deploy it in the most efficient manner possible. As Iñaki Bernal, Santander’s chief artificial intelligence officer, put it, “NVIDIA is a key player in the infrastructure, software and research that is powering the latest advances in artificial intelligence. Combined with Santander’s scale across businesses, customers and geographies, we believe we are uniquely positioned to succeed in transforming those advances into practical banking capabilities.” This statement captures the strategic rationale: leveraging NVIDIA’s cutting‑edge AI stack alongside Santander’s global footprint to turn research into usable banking solutions.
The Role of Santander AI Lab
The Santander AI Lab serves as the bank’s experimental hub, where emerging AI technologies are explored and tested before broader deployment across the institution. By operating as a proving ground, the Lab allows Santander to de‑risk innovation, ensuring that only robust, vetted capabilities move into production. The Lab’s activities include model evaluation, specialization, performance engineering, adversarial testing, and the design of simulated banking agents. This systematic approach creates a rigorous experimentation path that stresses AI systems against banking‑specific requirements such as security, regulatory compliance, and latency constraints.
From General‑Purpose to Specialized Models
A central thrust of the Lab’s work has been the transition from large, general‑purpose foundation models to smaller, specialised models tailored to discrete banking tasks. In selected experimental use cases, this shift has driven serving costs down from around €5 per interaction to approximately €0.03 while maintaining strong performance. The bank’s engineers argue that specialised models, when fine‑tuned on Santander’s own data, can achieve comparable—or superior—accuracy to their larger counterparts because they focus computational resources on the relevant patterns and eliminate unnecessary overhead. This philosophy aligns with NVIDIA’s view that “cost per intelligence output is a key metric,” as expressed by Aser Blanco, Global Banking Lead at NVIDIA.
Cost Reduction Achievements in Customer Service
One of the most tangible outcomes of the specialised‑model strategy is evident in an experimental customer‑service use case. Santander reported that it improved conversation quality while reducing costs by more than 50% through the deployment of a more efficient AI model. The improvement was measured not only in reduced latency and lower compute expense but also in higher customer satisfaction scores, demonstrating that efficiency need not come at the expense of user experience. The Lab attributed these gains to the combination of model right‑sizing, targeted fine‑tuning, and optimised inference pipelines.
Model Arena and NVIDIA NeMo Customizer
To systematically identify the best model for each workload, Santander built Model Arena, a framework that evaluates open‑weight models against a suite of banking‑specific tasks. Model Arena enables the bank to simulate, benchmark, and supervise AI agents in controlled environments, ensuring that only models meeting performance, safety, and cost thresholds progress to production. Complementing this effort, the Lab employs NVIDIA NeMo Customizer to specialise models. NeMo Customizer streamlines the process of adapting foundation models to Santander’s data, allowing rapid iteration and the creation of task‑specific variants without rebuilding models from scratch.
NVIDIA Kumo and Structured‑Data Intelligence
For workloads that rely heavily on structured data—such as transaction scoring, risk modeling, and fraud detection—Santander has harnessed NVIDIA Kumo, a family of open‑weight foundation models designed for tabular data. Kumo enables the bank to convert raw structured data into predictive insights that power agentic services, such as real‑time credit‑limit adjustments or dynamic fraud alerts. By integrating Kumo into its AI stack, Santander can leverage the same efficiency gains realised in natural‑language scenarios across its data‑intensive back‑office operations.
Industrialising Inference with NVIDIA Dynamo‑Triton
The Lab is also industrialising inference within Santander’s own architecture using NVIDIA Dynamo‑Triton, the inference‑serving technology that underpinned the observed cost reduction from €5 to €0.03 per interaction. Dynamo‑Triton provides dynamic batching, model‑versioning, and GPU‑utilisation optimisation, allowing the bank to serve thousands of concurrent requests with minimal latency. This infrastructure layer is critical for scaling specialised models to the volume demanded by a global bank while preserving the cost advantages demonstrated in early experiments.
Building Governed Banking Agents
Looking ahead, Santander is progressing from open‑weight models toward fully governed banking agents that embed greater control, safety, and lower operating costs at scale. These agents combine specialised language understanding, structured‑data reasoning, and rule‑based safeguards to ensure compliance with regulations such as GDPR, AML, and KYC. The Lab’s performance engineering and adversarial testing pipelines are designed to stress‑test these agents under edge‑case scenarios, thereby reducing the risk of unintended behaviour before wide‑scale rollout.
Three‑Year Framework and Future Focus Areas
As part of the next phase of the initiative, Santander and NVIDIA have formalised a three‑year collaboration framework covering foundation models, AI agents, fraud and anti‑money laundering capabilities, real‑time decisioning, and advanced simulation technologies—including a banking digital twin. The digital twin concept aims to create a virtual replica of Santander’s operational environment, enabling the bank to simulate policy changes, stress‑test AI-driven decisions, and optimise processes without impacting live systems. This holistic approach underscores the partnership’s ambition to move beyond isolated use cases toward an enterprise‑wide AI‑first operating model.
Executive Quotes on Partnership Value
Both leaders reiterated the strategic importance of turning data into actionable intelligence. Aser Blanco noted, “Banking is one of the most data‑centric industries. Data sits at the heart of every process and now banks need to become intelligence‑centric. Transforming years of data into usable and proprietary intelligence is not about model size, but about training and fine tuning. In this new world, cost per intelligence output is a key metric. Santander is showing what that looks like in production with smaller models, specialised on the bank’s own data, evaluated against real banking tasks and served efficiently enough to run at the scale of a global bank.” Bernal echoed this sentiment, emphasizing that the collaboration’s strength lies in combining NVIDIA’s technological leadership with Santander’s extensive customer base and geographic reach.
Conclusion: Implications for Banking AI
The Santander‑NVIDIA partnership exemplifies how a focused, experimental approach—anchored in model evaluation, specialization, and efficient inference—can yield substantial cost savings and performance gains in a heavily regulated industry. By shifting from monolithic, general‑purpose AI to lean, purpose‑built models supported by tools like Model Arena, NeMo Customizer, Kumo, and Dynamo‑Triton, Santander is charting a path that other financial institutions may follow. The three‑year framework promises to deepen these advances, particularly in fraud prevention, real‑time analytics, and the creation of a banking digital twin, ultimately moving the sector toward an intelligence‑centric paradigm where the true measure of success is the cost‑effective generation of actionable insights from data.
https://www.santander.com/en/press-room/press-releases/2026/10/santander-builds-ai-capabilities-that-improve-performance-and-lower-costs-using-nvidia-ai-infrastructure