The Data Moat Is Finally Real | The Public Ledger

Hello, Fintech Friends!

Two weeks ago I wrote that AI had given payments companies something better to do with their transaction data than lending. Toast, Block and Shopify are using it to build AI agents instead. I think I undersold what is happening.

On Nubank's second quarter call, founder and CEO David Velez said Brazil and Mexico "run on the same technology stack and increasingly on the same brain." The brain is nuFormer, the foundation model Nubank introduced about a year ago and trained on its own customers' transaction histories.

"One AI platform now powers underwriting, customer support, optimization and growth," Velez said. Nubank uses it to underwrite loans in Brazil and Mexico, and is testing it in Colombia. The model that predicts credit risk also predicts what a customer wants next, and Nubank has used it to target more than 100 marketing campaigns. AI agents also handle more than 60% of customer support conversations in Brazil.

Revolut published PRAGMA, its own foundation model built with NVIDIA, back in April, and last week set up a division called Revolut Research around it.

PRAGMA was pretrained on 24 billion events from 26 million users across 111 countries. The events include transactions, app activity, trading and communications. One pretrained model then serves credit scoring, fraud detection, engagement, product recommendation, recurring transaction prediction and lifetime value.

Against the models it is supposed to replace, Revolut reports 2.3 times the accuracy at identifying credit default risk, 65% more fraud caught with 17% better alert precision, and 41% more relevant product recommendations. Those results come from running PRAGMA on historical data rather than in production, but they are promising.

Banks have used machine learning models for decades, but each was built for a single job. The credit team trained a scorecard on repayment history, the fraud team trained a separate model on merchant patterns, and each started from scratch with variables its analysts defined by hand.

A foundation model is trained once on the raw transaction history, with no particular task in mind. It learns how a customer normally behaves, and the credit, fraud and marketing teams all build on that same model.

Everyone has been saying for years that data is the moat. Foundation models are what finally make that moat real.

Nubank and Revolut decided early that the data was the valuable part and built their systems from the beginning to capture all of it. Nubank calls more than a decade of transaction history across 100 million customers in three countries its main AI advantage. Pavel Nesterov, Revolut's head of AI, calls the strategy "build, don't bolt on," against legacy institutions retrofitting decades-old infrastructure with off-the-shelf software.

Fintechs have said for a decade that data is the moat. Until now, nobody could say exactly what it bought them.

Two weeks ago I wrote about Toast building agents to take over the marketing, payroll and bookkeeping its restaurants buy elsewhere. Nubank and Revolut can do the same with the investments, insurance and advice their customers buy elsewhere. The model that sets a customer's credit limit can also pick which product to put in front of them.

Velez calls the end point an AI private banker. I think this time he's underselling it. It's clearly the future of finance.

Jevgenijs Kazanins

p.s. Have feedback? Reach out on X

Block $XYZ ( ▲ 5.88% ) will open the Cash App Score to outside lenders, a plan it laid out at its investor day last November. The score reads paycheck deposits, spending, savings and repayment across Cash App rather than credit history, and already underwrites every Block lending product. In Cash App Borrow it approves 38% more customers at the same loss rate as traditional scores. "We built Cash App Score to see the financial activity of millions of people the traditional credit system misses," said Juan Hernandez, Block's head of credit and underwriting.

More than 30 million American adults have no credit record, or a file too thin or stale to score, per the CFPB. They earn and pay rent, and none of it reaches a bureau. FICO $FICO ( ▼ 0.38% ) is used in about 90% of US lending decisions, and its one serious challenger, VantageScore, is a cheaper read on the same bureau data. Cash App Score is built on the money moving in and out of tens of millions of accounts. Someone is finally building an alternative to FICO out of different data.

Read more: Block Will Open Its Cash App Score to Lenders

Affirm $AFRM ( ▲ 5.93% ) and Shopify $SHOP ( ▲ 1.47% ) launched Shop Pay Installments in Australia last week. Affirm bought PayBright to get into Canada, but came into the UK through Shopify in December, and now Australia the same way. It tried Australia once before, following Peloton in 2022 and winding the business down in early 2023 during cost cuts. Australia is the most saturated BNPL market in the world. Afterpay, now part of Block, and Zip both started there, and Klarna arrived in 2020. Can Shopify help Affirm succeed in Australia this time?

Affirm timed the launch to its fiscal fourth-quarter results. GMV grew 36% YoY to $14.1 billion, the eleventh straight quarter of 30% or higher growth. Revenue less transaction costs, Affirm's gross profit measure, grew 39% YoY to $589 million. Adjusted operating income grew 49% YoY to $353 million. Guidance for fiscal 2027 is more than $64 billion in GMV, about 27% growth, and Affirm says its expansion outside North America will not be a material contributor to it.

Read more: Affirm Q4 2026 Shareholder Letter

Advent and Stripe dropped their pursuit of PayPal $PYPL ( ▲ 4.33% ), Bloomberg reported last Friday. The two bid $60.50 a share in July, valuing PayPal at more than $53 billion, in what would have been one of the largest leveraged buyouts ever done. PayPal's board rejected that offer as too low, and the two sides spent August negotiating a higher price before the consortium walked away. Shares fell 12.7% on the news, but still trade above where they were before the bid.

Enrique Lores took over as the new CEO in March and has spent his first months on structure. He split PayPal into three businesses, each with one leader and its own revenue target, is removing three layers of management, and committed to at least $1.5 billion in cost cuts over two to three years. The bigger change is that he stopped treating branded checkout as the growth engine. It grew 2% YoY last quarter, and Lores expects credit, BNPL and Venmo to drive growth instead, starting in 2027.

Read more: Advent and Stripe Abandon $50 Billion Pursuit of PayPal

Goldman Sachs, Bank of America, Citi, Deutsche Bank and 17 other financial institutions are forming a company this year to issue a dollar stablecoin. They plan to launch it in the first half of 2027. The group started with 10 banks when it was announced in October 2025 and has doubled since. It plans to add stablecoins in other G7 currencies, with the euro as the priority.

A separate consortium of 37 institutions has already formed Qivalis to launch a euro stablecoin later this year, and some members belong to both. Circle $CRCL ( ▼ 0.94% ) leads the regulated stablecoin market today, and both consortia will compete with it, along with Open USD, which has more than 140 companies behind it, including Visa. The prize is clearly not what exists now, but the share of global payments that could move onto stablecoins. And apparently, banks see it as big enough to join forces.

Read more: Citi, Goldman, other global banks and asset managers team up on stablecoin venture