Nubank Bets on GPUs to Scale Its Proprietary AI Models

Nubank is investing in GPUs despite warning that AI spending will raise near-term operating costs. The Google News headline captures the hardware decision, but not the larger conflict. Nubank believes its proprietary AI models can improve credit decisions enough to justify greater control over scarce and expensive computing capacity.

That position challenges a simpler enterprise strategy: rent models and computing power from established cloud providers as needed. Nubank is instead combining its transaction data, internally developed foundation models, engineering research, and additional GPU capacity. The approach demands more capital and specialist talent before its long-term returns become visible.

The bank has already placed NuFormer, its proprietary family of foundation models, into production for credit cards in Brazil and Mexico. It also uses the models for unsecured lending in Brazil. Yet Nubank has not publicly disclosed the number, supplier, location, or ownership structure of the new GPUs referenced by management.

That gap matters. An investment in GPUs can mean purchasing servers, leasing dedicated capacity, reserving cloud instances, or financing infrastructure through another operator. Each arrangement creates different costs, security responsibilities, and exposure to hardware depreciation.

The important story is therefore not that a digital bank wants more chips. It is that Nubank is trying to turn its private transaction history into a durable underwriting advantage. GPUs are becoming part of that strategy, but the financial case still depends on measurable lending and operating results.

What the Google News Report Actually Signals

Nubank is treating AI compute as operating infrastructure, not a temporary experiment.

The original report followed management comments about investing in research, talent, and GPUs. During Nubank’s February earnings call, executives said those expenses would create short-term costs. They argued that medium-term benefits should more than offset the increase.

Management did not provide a precise GPU budget. It also grouped AI spending with two other initiatives that could pressure operating expenses: office expansion and international growth. That makes it difficult to isolate the financial effect of the company’s AI infrastructure program.

The timing is still revealing. Nubank’s efficiency ratio, which compares operating expenses with revenue, ended the fourth quarter of 2025 at 19.9%. It then improved to 17.6% in the first quarter of 2026, according to the company’s quarterly results.

A lower ratio generally indicates that a financial institution is generating more revenue for each unit of operating expense. Nubank warned investors not to treat the first-quarter result as its new steady state. It expects office, international, and AI investments to raise costs as the year progresses.

That warning creates the central tension. Nubank is spending more on the infrastructure behind a technology that it also presents as an efficiency engine. The company needs AI-driven gains in revenue, credit performance, or internal productivity to grow faster than the associated expenses.

The public evidence shows that the strategy extends beyond general-purpose chatbots. NuFormer processes sequences of financial transactions and learns reusable representations of customer behavior. Product teams can apply those representations to credit, cross-selling, and other predictive tasks.

This is different from asking a public language model to summarize a support ticket. Credit decisions must operate within strict latency, governance, privacy, and risk constraints. Small improvements can influence millions of individual limits, approvals, and collection decisions.

Nubank says its AI Private Banker features already serve more than 15 million monthly active users. It has also reported that Pix with AI, a conversational interface for Brazil’s instant-payment system, serves more than 10 million monthly active users.

Those products create ongoing inference demand. Inference is the computing process that applies a trained model to new information. It happens whenever a system generates a response, scores an application, or produces a personalized recommendation.

Training demand is more concentrated but can be much heavier. Nubank must repeatedly retrain, test, and adapt models as customer behavior and economic conditions change. More models, markets, and use cases can turn GPU availability into a persistent operational constraint.

The Google News item therefore points to a broader transition. Nubank is moving from buying access to AI capabilities toward maintaining a substantial internal AI production system. However, the available disclosure does not prove that every GPU involved will sit on Nubank’s balance sheet.

NuFormer Makes Compute a Banking Input

The economic value of Nubank’s GPU investment depends on NuFormer improving decisions that directly affect revenue and losses.

NuFormer is not a single consumer-facing language model. Nubank describes it as a proprietary set of foundation models trained on financial transaction data. A foundation model learns general patterns from a large dataset before teams adapt it to narrower tasks.

For Nubank, those narrower tasks include credit-card underwriting in Brazil and Mexico and unsecured lending in Brazil. The company says the models are already in production, which distinguishes this program from a research demonstration without customer exposure.

The models learn from sequences rather than isolated account snapshots. A customer’s deposits, purchases, transfers, bill payments, and repayment behavior form a changing timeline. Modeling that timeline can reveal relationships that manually designed variables might overlook.

Nubank says its real-time underwriting system evaluates every personal-loan request individually. The system estimates the predicted net present value, meaning the expected future value of a loan after accounting for risk and economics. Management says this evaluation occurs in under one second.

That speed matters because consumer lending combines risk management with product experience. A slow system can interrupt an application, while an overly conservative system can reject profitable customers. A poorly calibrated system can increase approvals while quietly creating future losses.

Nubank claims its newer models have supported credit-limit expansion. That remains a company attribution, not independent proof that AI alone caused the portfolio’s performance. Changes in customer mix, interest rates, policy rules, and economic conditions can also affect lending results.

The scale of the platform increases both the opportunity and the risk. Nubank reported more than 135 million customers across Brazil, Mexico, and Colombia in March 2026. It added roughly four million customers during the first quarter.

Its revenue exceeded $5 billion in that quarter, while net income reached $871 million. Return on equity was 29%. These figures give the company room to invest, but they also raise the consequences of a model error affecting a widely deployed decision system.

Credit models do not need to resemble the largest public generative models to consume substantial computing resources. Financial models must be retrained across markets, validated against several outcomes, monitored for drift, and tested under different policy settings.

A single production model can generate many related workloads. Engineers need experimental runs, benchmark models, validation jobs, stress tests, and controlled deployment candidates. Regulatory or internal governance reviews may require teams to reproduce earlier results.

Nubank’s engineering research illustrates the cost. In one published experiment, a training workflow using five million rows required about 12 hours on eight Nvidia A100 GPUs. Expanding the workflow to 40 million rows required approximately 95 hours on the same hardware.

That comparison does not disclose Nubank’s total fleet or annual GPU spending. It does show how quickly training demand can grow as researchers use more data. Multiplying the experiment across products and countries strengthens the case for predictable compute access.

Owning or reserving capacity can reduce the risk that a critical project waits for suitable cloud instances. It can also allow engineers to optimize a stable environment. However, dedicated capacity only saves money when utilization stays high enough to justify the commitment.

Model Efficiency Changes the GPU Calculation

Nubank is investing in hardware while simultaneously trying to make every training run consume less of it.

That combination can appear contradictory. If a new training method reduces compute requirements, why buy or reserve more GPUs? The answer lies in how organizations typically use efficiency gains.

Lower compute per experiment rarely causes an ambitious AI program to stop at its previous workload. Teams can instead run more experiments, train on larger datasets, test additional products, or refresh models more frequently. Efficiency increases the number of economically viable projects.

Nubank has focused part of its research on Muon, an optimizer that updates a neural network during training. An optimizer determines how model parameters change after the system measures its errors. Nubank compares Muon with AdamW, a widely used optimizer.

The company’s Muon experiments used 330-million-parameter models trained on 20 million samples. Nubank reported that Muon converged faster and reached lower validation losses than AdamW across the tested learning rates.

Validation loss measures how well a model performs on data excluded from its direct training process. Lower loss can indicate better generalization, although it does not automatically guarantee better lending or customer outcomes.

Nubank also cited broader scaling experiments suggesting that Muon can reach comparable model quality with approximately 52% of the training operations required by AdamW. That figure describes a research result, not a guaranteed reduction across every NuFormer workload.

The bank later contributed research on an 8-bit version of Muon. Quantization reduces the number of bits used to represent certain values. The researchers reported a 62% reduction in optimizer-state memory while maintaining performance in their tests.

Memory reduction can let researchers fit larger models or batches onto the same hardware. It does not always produce an equivalent increase in training speed. Additional conversion work or different hardware bottlenecks can limit the wall-clock benefit.

Even so, memory efficiency can improve infrastructure economics. A job that previously required more devices may fit on fewer GPUs. Another job might use the freed capacity. The result depends on scheduling, model architecture, networking, and actual utilization.

Nubank’s research also indicates that it wants technical control below the application layer. The bank is not only connecting a commercial model to its products. It is studying how its models train, how gradients are updated, and how memory consumption changes.

That depth supports the argument for dedicated computing capacity. Researchers can configure environments around their workloads and maintain more consistent performance. Sensitive financial data may also remain within a controlled architecture, depending on where and how the equipment operates.

However, better optimizers do not settle the purchasing decision. GPU generations improve quickly, and a device acquired today can lose economic value before its physical life ends. Cloud capacity transfers part of that obsolescence risk to the provider.

Dedicated infrastructure also requires networking, storage, orchestration, security controls, monitoring, and specialist operations. A GPU sitting idle still carries a cost. A heavily utilized GPU can create additional cooling, power, and maintenance demands.

The most credible interpretation is not that Nubank has solved AI efficiency. It is that the company sees sufficient model demand to pursue efficiency and capacity at the same time. The models use less compute per unit of work, while the total amount of useful work expands.

Nubank Is Choosing Control Over Pure Cloud Flexibility

The primary contest is between controlled compute capacity and the flexibility of renting everything on demand.

Cloud platforms make it relatively easy to start an AI project. A company can provision accelerators without building a data center or committing to a full hardware lifecycle. Teams can also select different devices for training and inference.

That flexibility carries tradeoffs. High-demand accelerators may be unavailable in a desired region. Long-running workloads can become expensive. Moving large datasets between services can introduce latency, operational complexity, and additional charges.

Banks face another constraint: data governance. Customer transactions are highly sensitive, and financial institutions must control access, retention, lineage, and auditability. Cloud systems can satisfy regulated workloads, but the bank remains responsible for configuring and supervising them properly.

Nubank already operates a proprietary, cloud-native banking stack. The phrase cloud-native means its systems were designed around distributed cloud services rather than moved from traditional banking servers without major redesign. That history makes a hybrid approach plausible.

A hybrid architecture might combine reserved cloud GPUs, leased clusters, and directly controlled machines. It could route training jobs according to cost and availability while keeping certain datasets or model stages within stricter boundaries.

Public comments do not yet establish which configuration Nubank selected. The company has discussed investment in GPUs, but it has not identified a vendor or disclosed a purchase quantity. Describing the strategy as ownership therefore goes beyond the most detailed public evidence.

The distinction is more than wording. Buying hardware creates capital, depreciation, and utilization questions. A long-term capacity contract creates a different expense profile. Renting on demand preserves flexibility but can expose the company to volatile availability and pricing.

The selected model also affects bargaining power. A company with predictable workloads can negotiate long-term capacity or distribute jobs among providers. A company tied to one vendor’s proprietary services may find switching more difficult.

Nubank’s acquisition of Hyperplane in 2024 provides useful context. Hyperplane developed self-supervised foundation models using financial institutions’ proprietary data. Nubank said the acquisition would help product and engineering teams apply shared models across customer journeys.

The Hyperplane acquisition brought model-building talent and technology inside Nubank. Its systems were designed to train, evaluate, and deploy numerous customized models for financial use cases.

That move shifted Nubank toward internal model development before management emphasized GPU investment. The hardware decision follows the software and talent decision. Once a company maintains proprietary models, compute becomes a recurring production dependency.

Traditional banks are under pressure from the same trend, although they begin from different technology structures. Many have large proprietary datasets and substantial infrastructure budgets. They may also carry older core systems that make unified model deployment slower.

Other digital financial companies are developing transaction foundation models too. Stripe has published research on a payments foundation model, while Visa, Mastercard, Revolut, and Plaid have explored related architectures. The contest is not simply Nubank against one named bank.

The competitive question is who can connect data, models, compute, and product decisions most effectively. Large datasets alone do not create an advantage if teams cannot clean them, train models, validate results, or deploy changes safely.

Nubank argues that its unified data and internally built systems shorten the path from experimentation to production. That claim has strategic weight, but investors still need comparable evidence. Deployment speed, model lift, and risk-adjusted returns matter more than the number of AI projects announced.

The Efficiency Claim Still Needs a Harder Test

A lower expense ratio does not prove that Nubank’s AI infrastructure has paid for itself.

Nubank’s first-quarter efficiency ratio improved from 19.9% to 17.6%. Management attributed the result to stronger revenue, portfolio growth, operating improvements, and some timing effects. The company also excluded selected strategic investments when discussing a lower core ratio.

That creates an important measurement issue. If AI infrastructure spending is excluded from a core efficiency measure, the adjusted figure cannot answer whether the complete AI program is economically efficient. It can still show how the underlying platform behaves before selected investments.

Nubank expects its reported efficiency ratio for 2026 to finish near the level recorded at the end of 2025. That guidance implies that current investment will offset at least some underlying operating leverage during the year.

Management has said roughly one-third of the first-quarter cost outperformance reflected structural gains. It cited AI improvements in operations and collections, software consolidation, and hiring discipline. The remaining portion involved expenses expected to arrive later.

These categories are not sufficiently detailed to calculate a return on AI spending. Investors do not know the separate costs of models, GPUs, AI hiring, external services, or supporting infrastructure. They also lack a precise revenue attribution for NuFormer.

The credit portfolio offers a more consequential test. Nubank reported a 5.0% ratio of loans between 15 and 90 days overdue in the first quarter. That was 89 basis points higher than the previous quarter, which the company largely attributed to normal seasonality.

Loans more than 90 days overdue declined by 10 basis points to 6.5%. Nubank also said improved risk models gave it confidence to extend credit into higher-risk segments. That strategy can increase profitable lending, but it also raises the importance of monitoring later delinquencies.

A model can approve more borrowers and appear successful before losses fully emerge. Consumer credit performance develops over months. The relevant comparison includes net revenue, funding costs, expected losses, collections, and capital usage over the loan’s life.

NuFormer therefore needs to demonstrate more than approval growth. It must produce attractive risk-adjusted returns across changing economic conditions. Performance in Brazil may not transfer unchanged to Mexico or Colombia because customer behavior and credit systems differ.

Model governance creates another uncertainty. Foundation models can capture patterns that are difficult to express as traditional underwriting rules. That flexibility can improve predictions, but it makes documentation, monitoring, and explanation more important.

Bias is a related concern. A model can learn correlations associated with protected or vulnerable groups even without receiving an explicit sensitive attribute. Financial institutions need testing that examines who receives approvals, limits, pricing, and adverse decisions.

Nubank’s public materials describe its model strategy more thoroughly than its independent validation process. That does not show that controls are absent. It means readers cannot use the available disclosure to compare model governance with the claimed performance gains.

The Google News framing also risks merging two separate claims. Nubank has published evidence that Muon improves training efficiency in controlled experiments. Management has separately said its proprietary models produce significant lift in credit and cross-selling.

Neither claim automatically proves that buying or reserving GPUs is cheaper than using cloud capacity. That requires workload-level accounting, including utilization, energy, support, networking, depreciation, and the opportunity cost of committed capital.

The strongest evidence will arrive through business outcomes. If credit expands while later delinquencies remain controlled, NuFormer’s underwriting case becomes stronger. If productivity improves while total expenses remain disciplined, the infrastructure case becomes easier to defend.

What to Watch After the Initial Google News Cycle

Three signals will show whether Nubank’s GPU investment is creating an advantage or merely adding an expensive infrastructure layer.

The first signal is credit performance across model-backed portfolios. Watch growth in card limits and unsecured lending alongside early and late delinquency ratios. Revenue expansion matters, but it cannot be separated from the cost of defaults.

Nubank should ideally provide cohort data showing how loans approved under newer models perform over time. A cohort groups accounts opened during a similar period. It helps distinguish recent underwriting quality from the behavior of older portfolios.

If model-backed cohorts maintain better risk-adjusted returns while serving more customers, the proprietary-model thesis gains support. If losses rise faster than expected, approval growth will look less persuasive.

The second signal is the relationship between AI spending and operating efficiency. Nubank has already warned that its investments will create near-term pressure. Future results should clarify whether underlying cost savings and revenue gains begin to outpace that pressure.

The headline efficiency ratio alone will not be enough. Readers should look for disclosures about collections productivity, customer-service workload, engineering throughput, credit decision speed, and cross-selling. These operating measures connect model use to identifiable business processes.

Nubank previously reported more than 50% year-over-year growth in engineering throughput and a tenfold increase in weekly token consumption since January. Those are adoption indicators, not standalone financial returns. More generated tokens can represent valuable usage or simply higher compute consumption.

The company’s investor materials should eventually reveal whether adoption translates into durable operating leverage. A stable reported ratio during a heavy investment year would support management’s argument, provided credit quality remains controlled.

The third signal is infrastructure disclosure. Nubank does not need to reveal sensitive system architecture, but the market needs enough information to understand the commitment. Useful details would include whether capacity is purchased, leased, or reserved.

Hardware generation and utilization would also help. The economics of a small research cluster differ from those of a production fleet serving real-time inference. Training capacity can be scheduled in batches, while customer-facing systems require dependable availability.

Supplier concentration is another factor. Reliance on one GPU vendor or cloud provider can create pricing and availability risk. A multi-provider strategy can improve resilience, although it also increases engineering complexity.

Nubank’s technical publications offer a partial window into its choices. They show an organization working on optimizer behavior, memory use, training efficiency, and transaction-specific models. Continued peer-reviewed research would strengthen confidence in its engineering depth.

Yet research quality and commercial success remain separate tests. A useful optimizer can reduce training costs without making a credit product profitable. A successful credit model can create value even if it runs on rented hardware.

The ultimate judgment should connect all four layers: model quality, infrastructure economics, product adoption, and financial outcomes. Leaving out any layer produces an incomplete picture.

Nubank has made the direction clear. It wants AI to become part of underwriting, customer interaction, operations, and product development. Its more recent results describe NuFormer as a production system rather than a future ambition.

What remains unclear is how much computing capacity Nubank will control, what that control costs, and how quickly the investment generates measurable returns. The company has provided enough evidence to take the strategy seriously, but not enough to declare the GPU bet successful.

The next earnings cycles should narrow that gap. Watch the loan cohorts, the full efficiency ratio, and any disclosure about compute ownership or contracts. Those signals will determine whether the Google News headline marked an infrastructure advantage or only the start of a costly test.