Announced 19 Oct 2026 · Sources checked
What Santander and NVIDIA announced
Banco Santander has signed a three-year collaboration framework with NVIDIA to build AI capabilities across the bank, according to a Santander press release reported on Monday, 19 October 2026, by PYMNTS and Spain's Expansión. The agreement covers foundation models, AI agents, fraud and anti-money-laundering tools, real-time decision-making and advanced simulation, including digital twins of banking systems.
The headline number is about cost rather than scale. Santander says it has been moving from large general-purpose models towards smaller models specialised for specific banking tasks, and that in selected experiments this cut serving costs from about 5 euros per interaction to about 0.03 euros while keeping strong performance. In one experimental customer-service case, it says a more efficient model improved conversation quality and cut costs by more than 50 percent.
No financial terms were disclosed. When AI Lookout checked on the morning of 19 October, the release had not yet appeared in Santander's or NVIDIA's online newsrooms, so the details here come from the two outlets that received it and quoted it directly.
From big general models to small specialised ones
The work runs through the Santander AI Lab, the unit where the bank tests emerging technology before rolling it out across the group. With NVIDIA's support, the lab has built what Santander calls a Model Arena, which evaluates open-weight models against real banking tasks and picks the right model for each workload.
"As part of this effort, Santander has been progressively moving from large general-purpose models towards smaller, specialised models designed for specific banking tasks," the bank said in the release quoted by PYMNTS. The logic is that a narrow job, such as classifying a customer request or drafting a reply about a card transaction, does not need the largest available model if a smaller one, tuned on the bank's own data, does it as well at a fraction of the cost.
Santander has not said which open-weight models scored best in its arena, how the 5-euro baseline was measured, or how many interactions the 0.03-euro figure covers. The release frames both numbers as results from experiments rather than from systems already serving all customers.
The NVIDIA software in the stack
According to Expansión's account of the release, Santander is fine-tuning models for specific tasks with NVIDIA NeMo Customizer, part of NVIDIA's NeMo tooling for adapting models to an organisation's data. It has also built specialised intelligence for core banking processes on NVIDIA Kumo, which the release describes as turning data into predictive information for services.
To put those models into production, the lab is using NVIDIA Dynamo-Triton, NVIDIA's inference-serving software, which the bank credits with helping to reduce costs. Santander is also developing models to design, simulate, evaluate and monitor banking agents, and to stress-test conversational AI in controlled environments before it reaches customers.
Aser Blanco, NVIDIA's global banking lead, framed the deal around efficiency. "In this new world, cost per intelligence output is a key metric," he said in the release. "Santander is showing what that looks like in production with smaller models, specialized on the bank's own data, evaluated against real banking tasks and served efficiently enough to run at the scale of a global bank."
What the collaboration will cover
The three-year framework lists five areas: foundation models, AI agents, fraud and anti-money-laundering capabilities, real-time decisioning, and advanced simulation such as banking digital twins, virtual replicas of a product or system that let teams test changes and anticipate problems without real-world risk.
Iñaki Bernal, Santander's head of artificial intelligence, said in the release, as reported by Expansión, that NVIDIA is central to the infrastructure, software and research behind recent AI advances, and that combining this with Santander's scale across businesses, customers and countries puts the bank in a position to turn those advances into useful banking capabilities.
The bank stresses that NVIDIA is one partner among several. Santander's June update on its strategy says it uses Microsoft Copilot for everyday productivity and a multi-provider approach for specialised work, including OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, startups and G42 for AI-enabled banking solutions. The NVIDIA deal appears aimed at the layer underneath: which models the bank runs itself, and how cheaply.
Where this fits in Santander's AI plan
In a 22 June post, Santander's chief data and AI officer, Ricardo Martín Manjón, set a target of more than 1 billion euros in business value from AI between 2026 and 2028, combining new revenue and cost savings. The bank reported 35 million euros in the first quarter and said it was on track to exceed 200 million euros by the end of 2026. Expansión reports that the figure reached 84 million euros in the first half.
The June post also listed AI access for all 185,000 employees, more than 280 process-automation agents in production, 17,000 people using AI in software development with 40 percent of code developed by AI in June, and Openbank models that process 100,000 anti-money-laundering alerts a year. Santander has started using AI on voice channels for some customer queries and uses machine learning on real-time data to decide whether a new customer can get a credit card from the first day.
The AI Lab has also published more than a dozen projects under the Apache 2.0 licence on its GitHub organisation, including a synthetic fraud-graph generator and LLM evaluation tools, which fits the evaluation-first approach behind the Model Arena.
Why banks are watching cost per answer
At the scale of a global bank, the gap between 5 euros and 3 cents per interaction decides whether an AI assistant is a pilot or a default channel. That is why Blanco's phrase about cost per intelligence output matters more than the partnership itself: it signals a shift from asking which model is smartest to asking which model is good enough for each task at the lowest cost.
Smaller in-house models also fit the direction regulators are pushing. The Financial Stability Board's 12 sound practices for AI in banks and the Hong Kong Monetary Authority's new AI rules for banks both emphasise testing, monitoring and control of AI systems, which is easier when a bank owns the models, the evaluation suite and the serving stack.
Banks are also drawing harder lines around outside agents. JPMorgan has moved to block third-party AI agents from Chase accounts, while Santander is investing in agents it designs, simulates and stress-tests itself. The two approaches point the same way: banks want AI on their own terms, inside systems they can audit.
What is still missing
The release leaves several questions open. There is no contract value, no list of the open-weight models that won in the Model Arena, and no detail on how the 0.03-euro cost was measured or whether it includes hardware, engineering and monitoring. The 50 percent customer-service saving comes from an experimental use case, not a group-wide rollout.
It is also unclear how much new work the framework adds to collaborations that, according to Expansión, Santander already had with NVIDIA. Fraud, anti-money-laundering and digital twins are listed as areas of work rather than as products with dates.
Things to watch include Santander's next quarterly results, where it may update the 200-million-euro AI value target, any publication of Model Arena methods or results by the AI Lab, and whether NVIDIA presents Santander as a reference customer at its upcoming events. AI Lookout will update this article if Santander or NVIDIA publish the release or further details.
Common questions
What did Santander and NVIDIA agree?
A three-year collaboration framework covering foundation models, AI agents, fraud and anti-money-laundering capabilities, real-time decisioning and advanced simulation such as banking digital twins. No financial terms were disclosed.
What is Santander's Model Arena?
An evaluation setup built by the Santander AI Lab with NVIDIA support that tests open-weight models against real banking tasks to choose the right model for each workload.
Did Santander really cut AI costs to 3 cents per interaction?
In selected experiments, according to the bank, serving costs fell from about 5 euros to about 0.03 euros per interaction while keeping strong performance. It has not published how this was measured.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. The cover places the companies’ own logos, taken from public-domain or openly licensed files credited in the caption, on a plain background. It is not a product photograph, a screenshot, or an endorsement.
Our editorial standards
