Banks have spent the last decade stacking up AI tools the way offices stack up software subscriptions. One model for fraud. Another for credit. A third for product recommendations. A fourth for risk. Each one built by a different team, trained on different data, and unable to talk to the others.
The problem is not that any individual tool is bad. The problem is that real customer behavior does not separate neatly into categories. A customer who travels frequently, pays bills early, and uses three currencies is a different risk profile than one who does not, but a fraud model that only sees transaction amounts and merchant categories cannot know that. The credit model that approved the loan did not know about the device change last Tuesday. The walls between systems mean that every new insight has to be rebuilt from scratch.
Revolut was one of the first to publish detailed results on what happens when you tear those walls down. Their model, called PRAGMA, was trained on 24 billion banking events covering transactions, app navigation, and account behavior from 26 million users across 111 countries. The result: credit scoring accuracy improved by 130% compared to their previous best system. Fraud detection recall went up 65%. Product recommendation accuracy improved by more than 40%. All from a single model, trained once, applied everywhere.
The credit scoring improvement is the one that matters most commercially. Lending is where banks make money. A model that can identify creditworthy customers that a previous model would have declined means more loans approved at the right price, more revenue, and fewer defaults. Revolut's team also reported that the weeks of manual work previously needed to prepare data before training dropped to near zero.
Mastercard is building the same thing at a larger scale. Their model, which they call a Large Tabular Model, is trained on billions of anonymized card transactions with a plan to expand to hundreds of billions, adding merchant location data, fraud signals, authorization patterns, and loyalty activity. Mastercard currently runs thousands of separate AI models across different markets and customer segments. A single foundation model would replace much of that, reducing maintenance costs and letting every product team build on top of one shared base.
Stripe's results are the most concrete. Their Payments Foundation Model, launched in May 2025, was trained on tens of billions of transactions. When deployed against card testing fraud, where criminals run small test transactions to check if stolen card details still work, detection rates jumped 64% almost overnight. Their previous approach had taken two years to reduce the same type of fraud by 80%. The new model did more in weeks.
The shared pattern across all three is worth naming plainly: transaction data, accumulated over years across millions of customers, turns out to be a more powerful competitive asset than anyone had fully appreciated. You cannot buy it, you cannot copy it, and you cannot replicate it quickly. The algorithm is not the moat. The data is.
This matters beyond the payments industry. Any business that sits on years of structured customer behavior data, insurers with claims histories, retailers with purchase records, logistics firms with shipment patterns, faces the same architecture question. The companies now racing ahead in financial services are not doing something exotic. They are doing something straightforward: training one large model on everything they know, instead of ten small ones each trained on a slice.
NVIDIA's toolkit, now available through Amazon Web Services and Nebius cloud, is an attempt to lower the barrier for institutions that have the data but not the in-house capability to build the infrastructure. The consulting partners signed up to deploy it, including EXL, Infosys, GFT, and Thoughtworks, suggest the pitch is aimed at mid-sized banks and insurers, not just the Revoluts and Mastercards of the world.
The honest caveat is regulatory. These models make consequential decisions: approving credit, blocking transactions, flagging accounts. Regulators in Europe and increasingly in other markets are asking financial institutions to explain those decisions. A model trained on 24 billion events is not easy to explain transaction by transaction. Revolut's own research found that the same model that improved credit scoring by 130% actually performed worse on anti-money laundering tasks, precisely because detecting money laundering requires looking across multiple customers and accounts simultaneously, not just at one customer's history in isolation. These models are powerful tools with real limits, and the limits matter in a regulated industry.
Institutions that start this work now are training a model that will compound in value every month as new transaction data flows in. Institutions that wait are not standing still. They are falling further behind a system that is getting smarter on its own.