Itaú’s $589 Million AI Deal Tests Banks’ Case for Ownership

The Brazilian lender is paying nearly six times NeoSpace’s earlier implied valuation, even as rivals develop financial AI without acquiring suppliers.

Itaú Unibanco agreed Monday to buy Brazilian AI developer NeoSpace for R$2.925 billion ($589 million), giving the bank control of technology designed to improve financial decisions. The agreement raises a question for banks investing in artificial intelligence: must they own the developers of their most valuable predictive models, or can they capture comparable benefits from externally supplied technology?

The price, equivalent to $589 million at the R$4.9654-per-dollar exchange rate used by Reuters on October 26, is nearly six times NeoSpace's implied valuation when Itaú first invested in January 2025.

Itaú invested $15 million for a 15% stake during an $18 million funding round, implying a roughly $100 million post-money valuation. The comparison is indicative rather than a like-for-like investment return, because the latest deal concerns the whole company and includes deferred consideration.

Under the agreement, part of the purchase price is payable at closing, with the remainder spread over five years subject to conditions. NeoSpace will retain autonomous management and its existing team. Completion requires approval from Brazil's Central Bank and competition authority CADE. Itaú disclosed no completion date.

The companies have worked together for nearly two years, developing models using Itaú's proprietary data for credit, customer relationships and fraud prevention. Itaú says the collaboration has generated significant results but has not quantified the financial gains.

Some NeoSpace founders previously helped establish Zup Innovation, which Itaú took control of in 2020. The established relationships, continued management autonomy and deferred payments suggest retaining technical personnel may be important to the deal, although Itaú has not disclosed any retention or performance-based payment terms.

Sources: Itaú's SEC filing · Reuters

The value of better predictions

NeoSpace develops tabular foundation models that identify patterns in structured financial and operational data. Its NeoData platform is designed to support multiple predictive applications without requiring banks to develop entirely separate machine-learning systems for each task.

An NVIDIA case study describes a deployment at an unnamed Latin American financial institution using customer characteristics, eligibility information, contracting history and transactions to predict who was most likely to accept installment-credit offers.

The model affected 6 million customers in production. Against the bank's previous system, it reportedly improved the Kolmogorov-Smirnov statistic—a measure of predictive discrimination—by 10–20 percentage points. Installment-credit offers increased approximately 10%.

The benefit was improved marketing efficiency: fewer customers could be contacted while producing better results. The study did not demonstrate lower credit defaults or higher risk-adjusted lending profits.

NVIDIA also reports 30–50% higher accuracy than traditional machine-learning approaches across NeoSpace applications. These supplier-published claims have not been independently validated publicly. A separate enhancement combining tabular models with transaction sequences remains experimental.

Other banks, other routes

Brazilian rival Nubank acquired specialist developer Hyperplane in 2024 and subsequently deployed transaction-based predictive models.

In research revised August 10, Nubank reported a 1.25% relative improvement in credit-default prediction AUC against a LightGBM baseline, using the same underlying data. AUC measures how well a model distinguishes borrowers who subsequently default from those who do not.

Nubank says that improvement is three times the typical gain from a successful model upgrade with material business impact. The system is deployed in production, serving credit decisions for more than 100 million users. Researchers also reported a 32% improvement in an undisclosed long-term surrogate performance metric. Neither finding establishes a corresponding reduction in actual credit losses or increase in profitability.

Revolut developed its PRAGMA foundation-model family in collaboration with NVIDIA. In an NVIDIA case study, Revolut reported 64.7% higher fraud recall and 16.7% higher precision against its previous production model.

Recall measures the proportion of actual fraud detected; precision measures how many flagged transactions are fraudulent. Improvements in both could reduce losses and unnecessary customer interruptions. The figures are Revolut's internal measurements, published by a technology supplier with a commercial interest in the results.

Santander offers a more direct challenge to Itaú's acquisition strategy. On October 19, the Spanish bank announced that it was developing specialized financial intelligence using NVIDIA's Kumo open-weight foundation models for structured data.

Santander is developing proprietary applications without purchasing the underlying model developer. In separate experiments, largely involving conversational AI, it reported reducing serving costs from approximately €5 to €0.03 per interaction. Those savings cannot be directly compared with predictive-model improvements.

Banks are pursuing predictive AI through different ownership arrangements

Sources: company disclosures and technical research, 2024–2026. Results concern different applications and metrics and are not directly comparable.

What ownership buys

The contrasting strategies suggest that a bank's competitive advantage may reside less in ownership of the original model developer than in combining specialized models with proprietary customer data, operational systems and decisions that competitors cannot easily reproduce.

Santander's use of open-weight technology illustrates one alternative to acquisition. Itaú's purchase illustrates what a bank may be willing to pay for greater control of technology and talent.

There are reasons to question whether that premium is justified. Banks have used predictive machine learning for decades, and conventional methods remain formidable competitors. A 2023 NeurIPS study of 19 algorithms across 176 datasets found that tuning established gradient-boosted models could matter more than replacing them with neural networks.

Proprietary development also entails research, computing, integration, validation and maintenance costs. Independent suppliers may have an advantage by spreading those expenses across multiple customers.

Itaú says the acquisition will also support ia.i, its generative-AI assistant, suggesting that the bank intends to combine specialized predictive capabilities with customer-facing language-model applications.

Itaú has put R$2.925 billion behind the proposition that greater control is worth the investment. The test will be whether NeoSpace delivers enough additional revenue, lower losses and operating savings to justify that price when competing banks can pursue similar intelligence without buying the supplier.