Nubank Didn't Win Credit Market Share by Cutting Rates. It Trained a Model.

Nubank's Q2 2026 earnings — $1.1 billion in net income, 49% growth, a record 33% return on equity — landed exactly as most expected. Analysts tracked the profit. Almost nobody wrote about the 50 basis points.

In Q4 2025, Nubank gained half a percentage point of credit card purchase volume in Brazil. That sounds modest. In context, it was the single largest quarterly gain by any institution in that market in ten years. The cause was nuFormer.

nuFormer is a transformer neural network — the same class of architecture that powers large language models — trained specifically to assess credit risk using Nubank's proprietary behavioral transaction records. Traditional credit scoring works from your past: bureau history, declared income, time-since-default. nuFormer reads your present. Every transaction sequence, every repayment rhythm, every spending pattern across Nubank's 139-million-customer network becomes signal. The result: 70% reduction in credit risk for an equivalent borrower population. Three times the improvement of a typical single-model generation.

The structural implication is what Q2 commentary glossed over. nuFormer is already deployed across credit cards in Brazil and Mexico, and personal loans in Brazil. SME credit and Colombian cards are in testing. Each deployment adds more behavioral data, which refines the model, which improves approval accuracy, which enables credit to reach borrowers traditional lenders have consistently underpriced or rejected outright.

Brazil's credit market has always had a structural problem: sparse bureau coverage for lower-income borrowers, concentrated incumbents with pricing power, rates that reflected uncertainty rather than actual risk. Open Finance — Brazil's data-sharing framework that lets consumers authorize lenders to access their transaction history held at other institutions — partially solved data accessibility. But Open Finance levels the field on access, not on model quality. A model trained on years of Nubank-specific behavioral data from 139 million accounts reflects patterns no consent-sharing arrangement can replicate for latecomers.

That is the asymmetry the market is undervaluing. Every major Brazilian bank now runs an AI team. Itaú's IAI assistant reached 300,000 users last month. Banco do Brasil is piloting agentic relationship management. The industry calls itself AI-native. But there is a difference between deploying AI and accumulating training advantage. Nubank started extending digital credit to thin-file borrowers when traditional banks were still debating smartphone apps. That head start has compounded into the training dataset nuFormer now draws on.

The interesting question is not whether competitors will respond. They will. The question is what a model looks like after two more years of training on 150 million customers — at a time when Brazil's Open Finance infrastructure keeps generating new behavioral signal with every Pix transaction, credit card swipe, and investment consent.

Credit moats used to be built on branch density and rate floors. This one grows with every approval.