Nubank and Revolut turn transaction data into AI moats

Proprietary foundation models trained on billions of customer events are reshaping how neobanks underwrite, sell and retain.

By Disrupts Editorial Team Thu, 10 September 2026

A computer monitor displaying system performance data, a keyboard, and a mouse are positioned on a white desk in a brightly lit data center, surrounded by rows of dark server racks featuring blinking blue and green LED lights.

Nubank and Revolut have each deployed proprietary foundation models built on years of customer transaction data, a move that signals a structural shift in how digital financial services companies convert data scale into competitive advantage. Where legacy banks built separate machine-learning models for credit, fraud and marketing, the two neobanks have trained single, large models on raw behavioural histories and are now routing most core functions through them. The result, both companies argue, is not incremental improvement but a genuinely different kind of infrastructure.

Nubank's model, nuFormer, was trained on more than a decade of transaction history across roughly 100 million customers in Brazil, Mexico and Colombia. On its second-quarter earnings call, founder and CEO David Velez said Brazil and Mexico now "run on the same technology stack and increasingly on the same brain." The model handles loan underwriting, credit-risk prediction, customer support and targeted marketing from a single pretrained base. AI agents already manage more than 60% of customer support conversations in Brazil, and nuFormer has underpinned more than 100 marketing campaigns. Velez describes the end point as an "AI private banker" for every customer.

Revolut's PRAGMA: one model, six functions

Revolut's equivalent, PRAGMA, was built in partnership with NVIDIA and published in April, with a dedicated research division, Revolut Research, announced last week. PRAGMA was pretrained on 24 billion events drawn from 26 million users across 111 countries, spanning transactions, app activity, trading and communications. From that single pretrained base, the model serves credit scoring, fraud detection, engagement, product recommendation, recurring-transaction prediction and customer lifetime-value estimation.

Revolut reports that in historical backtests, PRAGMA delivers 2.3 times the accuracy of the models it replaces on credit default prediction, catches 65% more fraud with 17% better alert precision, and generates 41% more relevant product recommendations. These are backtested figures rather than live production results, and should be treated accordingly, but the directional signal is consistent with what Nubank is reporting from deployment.

Pavel Nesterov, Revolut's head of AI, frames the strategy as "build, don't bolt on," a pointed contrast with incumbent banks retrofitting ageing core systems with off-the-shelf software. The underlying point is structural: legacy institutions never captured a unified behavioural signal because their data lived in siloed, function-specific systems. Nubank and Revolut built to capture everything from day one.

The convergence read-across

The strategic implications extend well beyond financial services. For cross-sector investors, the neobank foundation-model playbook is the clearest live demonstration yet of what proprietary data ownership is worth in an era of general-purpose AI. Both companies are effectively arguing that their data moat is now quantifiable, not merely rhetorical, and that it compounds: the same model that sets a credit limit can select the next product, sequence the next campaign and route the next support query.

This has direct implications for the broader AI infrastructure market. NVIDIA's role in building PRAGMA is a reminder that the sovereign-data strategy requires serious compute, and that partnerships between financial data holders and chip or cloud providers are becoming a structural feature of the landscape, not a one-off arrangement. It also raises the competitive bar for any institution considering a similar buildout: the training data accumulated over a decade cannot be replicated quickly, which is precisely why both companies are now talking about it loudly.

The parallel story developing in the US adds further context. Block has announced it will open its Cash App Score to outside lenders, offering a behavioural credit signal built on the money moving through tens of millions of accounts rather than traditional bureau data. More than 30 million American adults currently have no usable credit file, per the Consumer Financial Protection Bureau. The common thread across Nubank, Revolut and Block is the same: transaction data, accumulated at scale, is becoming the primary input for financial decision-making across underwriting, fraud and customer development. The institutions that own it are building with it; those that do not are buying it in.