A 50bp jump in credit card share, the largest by any Brazilian competitor in a decade, paired with stable 2.8% write-offs as AI model nuFormer cut projected risk 70%.
Nubank (NYSE: NU) executives used a late-2025 videocast to dissect the digital bank’s artificial-intelligence underwriting engine and its measurable impact on market share and asset quality. The headline numbers: a 50-basis-point expansion in Nubank’s share of credit card spending volume in Brazil during the fourth quarter, the largest single-quarter gain by any competitor in a decade, and a write-off rate that stayed flat at 2.8 to 2.9 percent. Chief Financial Officer Guilherme Lago, Credit Card Foundations General Manager Jeremy Selesner, and Vice President of Credit Risk Tyler Horn traced both outcomes to a model pipeline that makes lending decisions sharper without lowering acceptance standards.
The simple read is that Nubank is using AI to lend more aggressively. The better market read is that the company has built a structural edge in customer selection, one that allows it to grow share while keeping credit costs pinned. Traditional Brazilian banks, weighed down by legacy cost structures and slower-refreshing risk models, cannot easily replicate this dynamic.
The centerpiece of the technical update is nuFormer, a transformer-based AI model that delivered a 70 percent reduction in projected risk for comparable customer segments relative to the older modeling suite. Successive upgrades across the pipeline produced roughly triple the performance improvement of a standard model refresh. Nubank channels these gains into precision customer selection, not into broader approvals.
The model is the product of a deliberate, early commitment. Nubank launched with credit cards in Brazil rather than easing in through deposit or payment products. “We had to manage credit and manage economic cycles,” Selesner said. Delaying that credit-led entry would have been far harder, he added. The result is seventeen iterations of credit-limit adjustment models, ten iterations for customer acquisition, and an enormous data repository exceeding 100 terabytes of customer-behavior information. A dedicated risk team scrutinizes more than 1,000 monitoring indicators each week.
That scale of data ingestion and model refresh frequency creates a feedback loop that legacy banks, which commonly lean on bureau scores and slower update cycles, cannot match. The 70 percent risk reduction is not a one-off step change. It reflects compounding gains from better data, better architecture, and a corporate culture that treats credit risk as a real-time optimization problem.
The jump in credit card spending share during the fourth quarter of 2025 is the tangible output of that modeling edge. No other competitor in Brazil recorded a larger single-quarter gain over the past ten years. The advance did not arrive through easier underwriting. Horn stressed that his team “continue[s] to expect the future to be worse than the past to maintain that bar of resilience.”
That deliberate pessimism is a design choice. By calibrating models to a harsher environment than the one currently visible, Nubank avoids the pro-cyclical trap of overlending when conditions look benign. The flat write-off rate of 2.8 to 2.9 percent in the fourth quarter validates the discipline. A credit portfolio that is growing about 40 percent year-over-year, with a stable delinquency ratio, signals that the marginal borrower being added is not materially riskier than the existing book. The mechanism is a combination of better selection, dynamic limit management, and early-intervention triggers enabled by the real-time monitoring framework.
Revenue generated per active customer averaged $15 in late 2025, well below the roughly $40 posted by traditional Brazilian banks. The gap does not signal weakness. It reflects a customer base still early in its credit journey, many users holding their first card or first unsecured loan. That $15 base gives Nubank a long runway to deepen wallet share through higher limits, new products, and increasing engagement.
Customers already approved for credit hold about US$11 billion in untapped lines. That unused capacity is a latent asset. As the bank’s models gain confidence in individual repayment behavior, Nubank can activate those lines without acquiring new customers, driving revenue per user higher at minimal incremental acquisition cost. The 40 percent portfolio growth, set against $11 billion in headroom, points to a multi-year trajectory that does not require heroic assumptions about new-customer acquisition.
The videocast coincided with Nubank’s confirmed role in the federal government’s Novo Desenrola Brasil debt-relief program. The bank also rolled out an in-house initiative for borrowers who fall outside official eligibility. Both efforts run entirely through the Nubank app and rely on individualized assessments of repayment ability.
Built-in safeguards cap future debt accumulation. Users receive ongoing guidance on budgeting and financial habits. The program does more than rehabilitate borrowers who might otherwise remain excluded from the formal credit system. It generates a stream of behavioral data–every repayment and default becomes a training signal–that feeds back into the risk models. For a bank that treats credit as a data problem, the program accelerates the learning loop.
Horn’s insistence on expecting the future to be worse than the past is embedded in model calibration, not rhetoric. The 1,000-plus weekly monitoring indicators act as an early-warning system, flagging shifts in portfolio behavior before they show up in aggregate delinquency numbers. That stands in contrast to the typical credit-cycle playbook, where banks expand aggressively during good times and tighten abruptly when losses rise. Nubank’s approach aims to smooth the cycle by maintaining a constant margin of safety.
The risk is that a severe macroeconomic shock could still expose vulnerabilities in a relatively young loan book. The model’s built-in pessimism is designed to absorb exactly that kind of scenario. For investors sizing up the stock after the update, the variable that matters most is whether the market share gains continue without a corresponding uptick in credit costs. The next set of data–first-quarter 2026 delinquency figures–will show whether the 50-basis-point share grab brought any delayed deterioration. If write-offs stay inside the 2.8 to 2.9 percent range while revenue per customer ticks upward, the thesis that Nubank can compound its data advantage into sustained structural share gains becomes harder to dispute.
Drafted by the AlphaScala research model and grounded in primary market data – live prices, fundamentals, SEC filings, hedge-fund holdings, and insider activity. Each story is checked against AlphaScala publishing rules before release. Educational coverage, not personalized advice.