Nu Holdings Q3’25: Building the Moat While Breaking Records
How Nu is converting scale and data into a self-reinforcing AI advantage — proving the playbook in Mexico and revealing the intelligence engine behind its 31% ROE.
TL;DR:
Proof in Mexico: Deposit loyalty held firm despite yield cuts; ARPAC surged to $12.5, validating Nu’s international playbook and signaling long runway for monetization.
The AI Engine: Launch of nuFormer, a 330M-parameter foundational model trained on proprietary data, delivered 3× performance gains in credit decisions—marking the shift from UX to decision-quality dominance.
Investing from Strength: Efficiency ratio at 27.7% reflects deliberate investment in AI and market expansion, not cost slippage—Nu achieved record $783M net income and 31% ROE while widening its moat.
Nu Holdings reported another record quarter. Revenue hit $4.2 billion, net income reached $783 million, and Return on Equity achieved a stunning 31%. The company’s customer base swelled to 127 million, with engagement at 83%. By every headline measure, it was a resounding success.
Yet the market’s reaction was muted. One number gave investors pause: Nu’s efficiency ratio stood at 27.7%. For a company whose thesis is built on structural cost advantage, any efficiency question feels threatening. This created a paradox: how can a company become more profitable while costs haven’t fallen proportionally?
The answer is that the market is misinterpreting the signal. This quarter was not about extracting every last basis point of margin. It was about the deliberate cost of building an unassailable, AI-powered moat. To understand this, look at Nu not as a bank that uses technology, but as a three-layer system where each component reinforces the others.
The Three-Layer System
Nu’s advantage operates across three layers. The Trust & Distribution Layer is the customer-facing ecosystem serving 127 million users—banking, NuTravel, NuCel, NuMarketplace. The Data Layer captures every transaction, login, and interaction into a proprietary, closed-loop system unavailable to competitors. The Intelligence Layer is the “brain” that converts this data into predictive power—better credit decisions, fraud detection, and personalization at massive scale. Last quarter’s costs were deliberate investments widening the moat around all three layers simultaneously.
The Mexico Proof
The most significant investment was proving the Trust Layer’s durability outside Brazil. In my previous analysis, I framed Nu’s decision to cut Mexico deposit yields from 15% to 11.5% as “The Mexico Deposit Test”—would customer loyalty survive a rate cut while competitors offered 13% or higher?
Q3 provided a definitive answer: loyalty is earned, not bought.
Despite the rate cut, Nu’s consolidated deposits grew 6% quarter-over-quarter and cost of funding improved to 89% of interbank rates. In Mexico specifically, ARPAC surged to $12.5—a monetization level that took Brazil years longer to achieve. Cost-to-serve fell below $1, and the loan-to-deposit ratio remains at just 15%, implying years of deployment optionality.
The customers didn’t just stay—they became more profitable.
Nu has proven a repeatable playbook: lead with superior products to acquire customers, earn trust to build low-cost funding, then monetize efficiently across multiple products. This is fundamentally more sustainable than competitors burning cash for deposits. Mexico’s success comprehensively de-risks Act 2 international expansion.
What makes Mexico particularly compelling is a structural advantage: an 80% revolver rate compared to Brazil’s 10-15%. As Nu’s product stack fills out with lending and payroll services, the ARPAC ceiling is structurally higher. Today’s $12.5 is an early milestone on a path likely reaching $25-30. Mexico could contribute 20-25% of consolidated earnings by 2026.
The AI Engine Revealed
The quarter’s biggest revelation was the first detailed look at Nu’s Intelligence Layer: “nuFormer,” a proprietary foundational AI model. This is not a chatbot project—it’s a foundational investment in core technology that could redefine Nu’s competitive position.
The specifications are significant: a 330 million-parameter model trained on over 600 billion data tokens. These aren’t scraped from the internet—they’re proprietary outputs from Nu’s Data Layer. The model learns from signals only Nu can access: credit card patterns combined with NuCel mobile usage, NuTravel booking behavior, and NuMarketplace shopping activity.
Management stated nuFormer has delivered a 3x performance uplift over traditional machine learning—not 10% or 20% better, but three times more effective. Its first production application: optimizing credit card limits in Brazil. The model grants meaningful credit line increases—driving higher purchase volumes and interchange—without increasing overall risk appetite.
This is already visible in the numbers. Lower provisions and improved recovery rates are direct results of better decision-making. But credit limits are just the beginning. The model is deploying across fraud detection, collections, personalization, and into Mexico for new customer underwriting.
Here’s the moat: while traditional banks ask “What did this customer do?”, Nu predicts “What will this customer do next?” The competitive gap isn’t about having AI versus not having it—it’s about prediction quality, determined by data quality and breadth. Competitors can copy features and hire engineers. They cannot replicate trillions of proprietary data points spanning banking, telecommunications, travel, and commerce.
The announcement timing is revealing. Nu recognizes AI-powered decisioning will become the 2026-2027 competitive battleground. By establishing their lead now—while generating $783 million quarterly to fund R&D internally—they create both technical and perception moats.
The nature of competition is shifting. The old game was UX and cost structure—Nu won decisively. The new game is decision quality at scale. Traditional banks are blocked by culture (slow iteration), structure (legacy tech), and strategy (profit pool protection). Even well-funded fintechs lack the scale and breadth to train equivalent models.
The Financial Resolution
Through this lens, the margin paradox resolves completely. The most important metric—Risk-Adjusted Net Interest Margin—expanded to 9.9%. This measures profit after credit losses, the truest sustainability indicator. Even as Nu shifted toward safer, lower-yielding products, better risk pricing through the Intelligence Layer meant profit after losses increased. This is earnings quality improvement disguised as margin compression.
Headline NIM declined to 17.3% for three intentional reasons: strategic mix shift toward secured products, higher Brazil funding costs to defend primary banking relationships before incumbents respond, and Mexico’s 15% LDR creating mix pressure. None reflects competitive weakness—all reflect strategic choice.
Elevated operating costs are largely engineering and R&D for nuFormer, plus Mexico acquisition spending while unit economics are exceptional. The remarkable signal: Nu absorbed these investments while delivering 31% ROE. This is the power of the underlying model.
Consider the capital allocation: 46% LDR generating 31% ROE isn’t inefficiency—it’s optionality. Nu maintains dry powder for rapid deployment without external capital. Recent examples illustrate the pattern: pausing Pix financing despite profitability, offsetting FGTS headwinds with public payroll, keeping Mexico deliberately unprofitable to maximize long-term position. These are decisions optimizing for decades, not quarters.
Looking Forward
Probability weights from my Q2 analysis have shifted. The bull case moved from 40% to 45%, with the base case from 45% to 40%. Mexico exceeded expectations, and nuFormer provides concrete evidence of sustainable advantage.
What would change my mind? If risk-adjusted NIM compresses, if Mexico ARPAC stalls below $15, or if efficiency stays above 28% through 2026 without corresponding growth acceleration. None of these signals are present.
Nu’s edge is shifting toward something more durable: decision quality at scale. The company makes better, faster, cheaper decisions for 127 million customers simultaneously, improving automatically with every new customer and data point.
Mexico proved the playbook is repeatable. The AI reveal showed the engine. Financial results demonstrated durability while absorbing significant investments. The 27.7% efficiency ratio isn’t a warning—it’s evidence of investing from strength.
The market prices Nu as a bank at 18x forward earnings. Q3 revealed it’s becoming a platform where banking is merely the first application. The question is no longer whether Nu can replicate success beyond Brazil, but how fast it will scale its proven model globally.
The validation quarter is complete. The acceleration quarter has begun.
Probability weights from my Q2 analysis have shifted.
(Read the Nu Analysis Q2 2025)
$NU
Subscribe to Kristal Lens today!
Disclaimer:
The content does not constitute any kind of investment or financial advice. Kindly reach out to your advisor for any investment-related advice. Please refer to the tab “Legal | Disclaimer” to read the complete disclaimer.