Banco Santander said on 19 October 2026 that, in selected experimental use cases, moving to smaller specialised models served on NVIDIA infrastructure “has reduced serving costs from around โฌ5 per interaction to approximately โฌ0.03 while maintaining strong performance.” By our arithmetic, that is a cut of about 99.4% per interaction, in those experiments only.
In a separate experimental customer-service use case, the bank said it “improved conversation quality while reducing costs by more than 50%”. Santander and NVIDIA have also “established a three-year collaboration framework” for the next phase of their work.
Business Pill · COST PER TASK
A short explainer on cost per task: paying for the job, not the word. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, Santander’s experiment puts a price on the case for small, tuned models: in its selected tests, an interaction cost around โฌ5 on the old approach and approximately โฌ0.03 on a smaller specialised model served inside its own architecture.
The work runs through the Santander AI Lab, which the bank says “explores and tests emerging AI technologies before broader deployment across the bank.” With support from NVIDIA, the Lab “has developed new capabilities to evaluate AI models and agents for banking applications, helping identify which model is best suited for each task and deploy it in the most efficient way.”
The bank names the tools. It built Model Arena “to evaluate open-weight models against banking tasks and identify the right model for each workload”, and it is “specialising models using NVIDIA NeMo Customizer.” It also built “specialised intelligence for core banking workloads on NVIDIA Kumo, a family of open-weight foundation models for structured data”.
The release says the Lab “is industrialising inference within Santander’s own architecture using NVIDIA Dynamo-Triton, the approach behind the serving-cost reduction above”.
Serving cost per interaction in Santander’s selected experimental use cases: from around โฌ5 to approximately โฌ0.03, after moving to smaller specialised models served on NVIDIA infrastructure (Banco Santander, 19 October 2026). Experiments only; not bank-wide figures.
What Was Posted on X
Santander to Build AI Capabilities on Nvidia Infrastructure, Gradually Shifting Toward Smaller Models for Specific Banking Tasks
First Squawk, a markets headline feed, posted the news at 05:12 UTC on 19 October, before Santander’s release went up on its press room: Santander to build AI capabilities on Nvidia infrastructure, shifting toward smaller models for specific banking tasks. The post carries no cost figures; the โฌ5 and โฌ0.03 numbers in this piece come from Santander’s release.
From General-Purpose to Specialised Models
Santander describes the shift in its own words: it “has been progressively moving from large general-purpose models towards smaller, specialised models designed for specific banking tasks.” The โฌ5-to-โฌ0.03 figure is the result it gives for that approach in selected experimental use cases.
The release adds that “Customer-service comparisons delivered evidence that smaller, specialised models can deliver strong performance on targeted banking tasks.” The next step it describes is “progressing from open-weight models towards specialised models and governed banking agents designed for greater control, safety and lower operating costs at scale.”
In one experimental customer-service use case, Santander says it improved conversation quality while reducing costs by more than 50%; the three-year framework with NVIDIA covers foundation models, AI agents, fraud and anti-money laundering, real-time decisioning and a banking digital twin (Banco Santander, 19 October 2026).
The Three-Year Framework
The two companies’ new framework covers “areas such as foundation models, AI agents, fraud and anti-money laundering capabilities, real-time decisioning and advanced simulation technologies, including a banking digital twin.” The release gives no financial terms for it.
What the Two Companies Said
Santander’s chief artificial intelligence officer, Iรฑaki Bernal, said: “NVIDIA is a key player in the infrastructure, software and research that is powering the latest advances in artificial intelligence.”
Aser Blanco, NVIDIA’s global banking lead, said that “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.”
The Structural Reading
As we read it, the release puts a price on a choice every large company now faces: rent a large general-purpose model per call, or tune smaller open-weight models and serve them on its own stack. Santander’s own numbers, from experiments, describe the gap between those two options.
Our earlier reporting covered two other companies working on the same cost question: Latham & Watkins fine-tuning Nemotron in-house on NVIDIA H200s it bought, and Accenture capping Claude funding at $29 a month per employee. As we read it, Santander’s release puts a per-interaction price on that question.
The Structural Read
As we read it, the saving comes from two moves the release names together: picking a smaller specialised model for each task through Model Arena, and serving it inside the bank’s own architecture with NVIDIA Dynamo-Triton.
As we read it, NVIDIA’s role here is tools and infrastructure rather than a frontier model: the release lists NeMo Customizer, Kumo open-weight models and Dynamo-Triton, and NVIDIA’s Aser Blanco calls cost per intelligence output a key metric.
As we read it, the open question is scale: the release scopes both cost figures to experiments, and the three-year framework is where any wider rollout would have to show up.
Aser Blanco, Global Banking Lead at NVIDIA, in Santander’s press release, 19 October 2026
“In this new world, cost per intelligence output is a key metric.”
Three Implications
FRONTIER-MODEL VENDORS Santander says it has been moving from large general-purpose models towards smaller, specialised models designed for specific banking tasks; the release does not name the models it moved away from.
NVIDIA The release names NVIDIA NeMo Customizer, NVIDIA Kumo and NVIDIA Dynamo-Triton as the tools behind the Lab’s work, plus a three-year collaboration framework with no financial terms given.
OTHER BANKS NVIDIA’s Aser Blanco said banks “need to become intelligence-centric”, and Santander’s three-year framework lists fraud, anti-money laundering and real-time decisioning among its areas.
The framework puts it this way: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”
As we read it, Santander’s โฌ5-to-โฌ0.03 figure is that marginal cost per query, measured inside one bank’s experiments. The framework was published in March 2026 and does not discuss Santander or this collaboration.
What Is Not Established
The โฌ5 and โฌ0.03 figures are for “selected experimental use cases”, and the more-than-50% figure is for one “experimental customer-service use case”, in Santander’s words. The release does not say how many interactions were measured, which models were compared, what share of the bank’s traffic the approach covers, or what the three-year framework costs. The 99.4% figure is our arithmetic on the bank’s two numbers, not a figure Santander gives.
Business Engineer Framework
The State of The Inference Economy
A Business Engineer analysis of how inference became the main cost and revenue line of the AI economy.
Banco Santander said on 19 October 2026 that smaller specialised models on NVIDIA infrastructure cut serving costs in selected experimental use cases from around โฌ5 to approximately โฌ0.03 per interaction, and cut costs by more than 50% in an experimental customer-service case while improving conversation quality. The two companies set a three-year collaboration framework.
A note on sourcing. This piece is based on Banco Santander’s press release of 19 October 2026, read in full on the bank’s press room. The cost figures are Santander’s and come from experimental use cases; the 99.4% figure is our arithmetic on them. Nothing here is a forecast, and nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
