Banks Run Nvidia and OpenAI Models Closer to Cut Cost, Risk

In six days, three banks described running AI closer to their own control. NatWest Group said on 14 October 2026 that speech recognition built on NVIDIA Nemotron and adapted with its own data misheard 20% fewer words across the accents tested. Banco Santander said on 19 October 2026 that smaller specialised models served inside its own architecture cut serving costs, in selected experimental use cases, from around €5 per interaction to approximately €0.03. And Emirates NBD will run OpenAI’s GPT models on servers in the UAE, PCMag Middle East reported the same day.

As we read it, the common thread is control: of the cost of each task, of where the data sits, and of which models do the work. It is the same week JPMorgan Chase and Schwab said they would stop outside AI agents from acting inside customer accounts or moving clients’ money, as CNBC and Bloomberg reported.

Business Pill · OPEN WEIGHTS

A short explainer on open weights: rent the model or run it yourself. It teaches the general idea only and says nothing about any company in this story.

What Each Bank Announced

Santander. The bank said it “has been progressively moving from large general-purpose models towards smaller, specialised models designed for specific banking tasks.” In its words, “In selected experimental use cases, this approach has reduced serving costs from around €5 per interaction to approximately €0.03 while maintaining strong performance.” Its AI Lab “is industrialising inference within Santander’s own architecture using NVIDIA Dynamo-Triton”.

NatWest. The bank “selected NVIDIA Nemotron automatic speech recognition (ASR)” for its Spending Insights tool, and the resulting text “is then interpreted by NatWest’s proprietary Small Language Models”. The work “showed the system to mishear 20% fewer words” across the accents tested. Spending Insights “is now available to 75% of Royal Bank of Scotland customers as part of an ongoing rollout”, NatWest said.

Emirates NBD. PCMag Middle East reported that the bank will run OpenAI’s GPT models on servers in the UAE, “making it the first bank in the MENA region to use in-country inference.” According to the companies, as PCMag reports it, the setup is meant to help the bank meet local rules on data residency, privacy and operational resilience. PCMag says the release does not say when the deployment goes live or how much it costs.

What Was Posted on X

Santander executive chair Ana Botín wrote on 19 October that “being present in the infrastructure layer and having access to compute matters”, and called the bank’s three-year NVIDIA collaboration “another important step in that direction”. As we read it, that is the chair putting the infrastructure argument in her own words, which is the thread this piece follows across three banks.

The FirstSquawk newswire account summarised Santander’s announcement at 05:12 UTC on 19 October as building AI on NVIDIA infrastructure while “Gradually Shifting Toward Smaller Models for Specific Banking Tasks”. It is how the move first reached markets; the details in this piece come from the bank’s own release.

Where the Data Stays

NatWest’s release is the most specific about data. It said its anonymised contact-centre transcripts “remained within NatWest’s secure environment”, that “No NatWest customer data was shared with NVIDIA”, and that “any adaptations made using NatWest data are ringfenced to the bank.”

Santander puts the serving itself inside its own architecture. Emirates NBD’s choice, as PCMag describes it, is about geography: running the inference in the country, so that the bank’s prompts and data are handled there rather than abroad.

The Same NVIDIA Banking Lead, Twice

NVIDIA’s Aser Blanco is quoted in both the Santander and the NatWest releases. In Santander’s, he said: “In this new world, cost per intelligence output is a key metric.” In NatWest’s, he said working with the bank “has allowed us to apply NVIDIA Nemotron open-source models to those real-world banking challenges”.

Santander’s release describes evaluating “open-weight models against banking tasks”, and NatWest’s describes adapting “NVIDIA Nemotron open-source models” with its own transcripts. Emirates NBD’s plan, as reported, uses OpenAI’s models with OpenAI supplying “models, tools, and the UAE-based inference.”

The Other Side of the Same Line

Our earlier reporting covered two US firms drawing a boundary around customer accounts. JPMorgan Chase CEO Jamie Dimon said on 13 October that the bank would not allow outside AI agents to operate inside customer accounts, CNBC reported, and Bloomberg reported on 16 October that Schwab is blocking unauthorized agents that try to move clients’ money.

The Economic Times reported on 19 October that a Goldman Sachs report says “Many enterprises are currently capping AI spend, given elevated AI costs”. As we read it, Santander’s figures are one answer to that bill: a lower cost per task on smaller models it serves itself. NatWest and Emirates NBD, in the sources we read, give accuracy and data-residency reasons, not cost.

The Structural Reading

As we read it, three different routes lead to the same place. Santander tunes smaller open-weight models and serves them itself. NatWest pairs an adapted NVIDIA speech model with its own small language models. Emirates NBD keeps OpenAI’s models but moves the inference into the country.

As we read it, each route keeps more of the work under the bank’s own terms, whether that is where the models run or where the data stays, and the JPMorgan and Schwab statements draw a similar line for outside agents.

Update, 19 October: Huntington Builds a Small-Business AI Suite on Google Cloud

The Huntington National Bank said on 19 October 2026, in a release published by Google Cloud, that it is launching The 1866 Studio, “a new AI-powered operating system built with Gemini platform, Google Cloud’s AI platform,” for aspiring and existing small business owners. The bank said the system “will provide agentic AI workflows across the financial, operational and management needs of business owners.”

The first capabilities Huntington listed are business launch support, from entity selection and registration to licensing; marketing and campaign creation using “multimodal Gemini models”; and loan readiness assistance, in which “AI-powered agents help owners evaluate financing readiness by identifying potential gaps, explaining lender expectations and recommending actions.” Early access opens on 19 October and is “not limited to Huntington customers,” the release said. Huntington describes itself as a “$284 billion asset regional bank holding company.”

“Small business owners shouldn’t have to become experts in every function or in AI to build a successful business,” said Vijay Konduru, head of innovation and AI products at Huntington.

As we read it, Huntington takes a different route from the three banks above: rather than serving models inside its own architecture, it builds a customer-facing product on Google’s platform and brings its own small-business expertise, as the release puts it. The release does not say where customer data is processed.

Update, 19 October: Citi Hires a Google Engineering Director for Its ‘Agent Factory’

Citi has hired Sebastian Gnagnarella from Google as head of applied AI and agent factory, according to an internal memo seen by American Banker, which reported it on 19 October. American Banker says he previously was director of engineering at Google, where he led the deployment of the Gemini Enterprise Platform to Google Cloud clients, and that the memo came from David Griffiths, Citi’s group head of AI.

American Banker reports that, in the memo, Griffiths noted Citi has rolled out enterprise AI tools to more than 180,000 employees in 87 countries, and that Aida Filipovic, Niamh O’Sullivan and Sheetal Chanderkar moved into AI roles as head of responsible AI, head of AI strategy and chief operating officer of the groupwide AI team.

Update, 19 October: FIS Opens a Claude-Built AML Agent to Any US Bank

FIS said in a 15 October release that its FIS AML Investigations Agent, “jointly announced in May 2026 with Anthropic, is now commercially available to any U.S.-based bank, including those that are not FIS clients today.” FIS says it built the agent “with support from Anthropic and design partner Amalgamated Bank,” that the agent “runs in FIS’ private cloud and uses Anthropic’s Claude models,” and that it is now being deployed at Amalgamated. The release gives no price.

The time savings are projections. “Based on Amalgamated’s analysis of its investigation workflows combined with the AML Investigations Agent’s expected abilities,” FIS says the agent is projected to cut Level 1 alert review time by more than 90 percent, Level 2 case investigation time by approximately 98 percent, and SAR narrative preparation time by more than 80 percent. FIS adds that every output is a recommendation and that no case can be closed without explicit analyst confirmation.

As we read it, FIS offers a third route alongside the banks above: a core-banking vendor packaging a Claude-based agent inside its own private cloud and selling it to any US bank, including banks that are not its clients.

The Structural Read

As we read it, cost is only Santander’s stated reason: the bank gives a serving-cost figure, from around €5 to approximately €0.03 per interaction in selected experimental use cases.

As we read it, NatWest’s stated reasons are accuracy and data handling: 20% fewer misheard words across the accents tested, and transcripts that stayed in its own environment.

As we read it, Emirates NBD’s reason, as PCMag Middle East reports the companies’ release, is local rules on data residency, privacy and operational resilience.

Three Implications

FRONTIER-MODEL VENDORS Emirates NBD’s plan, as PCMag Middle East reports it, keeps OpenAI’s models and moves the inference into the UAE, with OpenAI supplying models, tools and the inference.

NVIDIA NVIDIA’s Aser Blanco is quoted in both the Santander and NatWest releases, and both banks name NVIDIA tools or models for the work.

OTHER BANKS JPMorgan Chase and Schwab said in the same week, as CNBC and Bloomberg reported, that outside AI agents would not act inside customer accounts or move clients’ money.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework describes one of its four moats this way: “The Container Moat — built by closed data loops inside a customer’s environment, nascent but the deepest position in the stack.”

As we read it, NatWest’s release describes that kind of loop: contact-centre transcripts that stayed in the bank’s environment and adaptations ringfenced to the bank. The framework was published in July 2026 and does not discuss any of these banks.

What Is Not Established

Santander scopes its cost figures to “selected experimental use cases”; NatWest’s 20% figure is for the accents tested. The Emirates NBD details come from PCMag Middle East’s account of the companies’ release, which we did not read ourselves, and PCMag says it gives no launch date or cost. Neither the Santander nor the NatWest release gives spending figures or the share of the bank’s AI traffic involved, and PCMag says the Emirates NBD release gives no cost. Grouping these three banks is our comparison, not a claim any of them makes.

The Bottom Line

Between 14 and 19 October 2026, NatWest, Banco Santander and Emirates NBD each described running AI closer to their own control. NatWest uses adapted NVIDIA Nemotron speech recognition and its own small language models; Santander serves smaller specialised models in its own architecture; and Emirates NBD will run OpenAI models in the UAE, PCMag Middle East reported.