Latin America's largest bank just made its most consequential AI bet: Itaú Unibanco disclosed Monday it has signed a Share Purchase Agreement to acquire all outstanding shares of NeoSpace A.I. Technologies Ltda. for BRL 2,925,000,000 — approximately $588 million USD — in a Material Fact filing submitted to the U.S. Securities and Exchange Commission. The deal crystallizes a thesis that Itaú has been stress-testing for nearly two years: that a bank's own transaction records, trained into a custom AI foundation layer, will outperform any generic model a cloud provider can sell it — and that the moment to buy the company building that layer is before a competitor does.

The acquisition is not a generic "bank buys AI startup" event. NeoSpace has spent close to two years training foundation models exclusively on Itaú's data — 60 million customers' worth of transaction histories, credit events, and behavioral signals. The company that built those models now becomes a wholly owned asset. Itaú is not acquiring a capability it is trying out; it is internalizing one it already depends on and permanently closing the door to rivals who might otherwise have made the same purchase.

The deal, structured with one installment payable at closing and the remainder spread over five years, requires prior approval from Brazilian regulators — specifically Brazil's antitrust authority CADE and the Central Bank of Brazil — before it can close.

From Zup to NeoSpace: Itaú Acquires the Same Engineering Team Twice

NeoSpace was founded in 2023 by Bruno Pierobon (CEO), Felipe Almeida (CRO), Gustavo Debs (COO), and Thiago Teixeira (CTO) — the same team that built Zup, a software engineering firm Itaú acquired in 2019 to accelerate its digital transformation. In other words, Itaú has now purchased this founding team twice, each time to bring a critical technology capability fully in-house.

The second acquisition follows a deliberate incubation pattern. In January 2025, Itaú led an $18 million pre-seed round in NeoSpace and took a 15% minority stake — becoming the startup's anchor client and its primary training-data source simultaneously. That structure gave Itaú a front-row view of what NeoSpace was building and how fast it was improving. By October 2026, rather than remain a customer and minority investor, Itaú moved to own the company outright.

The SEC filing states that NeoSpace's management and business operations will remain autonomous from Itaú following closing and that the current team will be retained in full — a structural choice that signals Itaú wants to preserve the startup's engineering culture, not absorb and dilute it.

What Large Data Models Actually Do — and Why They Are Not Chatbots

NeoSpace does not build chatbots. Its core technology is what the company calls Large Data Models (LDMs) — an architectural paradigm distinct from the Large Language Models (LLMs) that power products like ChatGPT and Claude. Where LLMs are trained on massive volumes of text scraped from the public internet to understand and generate language, NeoSpace's LDMs are trained on billions of structured and unstructured corporate records — financial transactions, customer behavior histories, credit events, fraud signals — to produce real-time predictions and business decisions.

IBM's research on the LDM paradigm describes the core mechanism: each database column value becomes a token, and the model perceives each database record as an unordered sequence in which each token maintains equal relationship with the others, deducing significant values from surrounding column context both within and across rows. The practical result is a model that can read two years of a customer's transaction history and instantly produce a credit limit recommendation, a fraud risk score, or a churn propensity signal — without requiring a human analyst to design a feature pipeline.

According to materials published alongside an NVIDIA case study of NeoSpace's work, Pierobon has described enterprise foundation models as "the missing layer between data and decisions." That framing is precise: NeoSpace is not competing with ChatGPT. It is competing with the fragmented machine learning pipelines that most banks still use — pipelines that require separate models for credit scoring, fraud detection, and CRM, trained separately, maintained separately, and frequently producing contradictory signals.

NeoSpace's NeoData platform, running on NVIDIA Blackwell via Oracle Cloud, replaces those fragmented pipelines with a single predictive layer. In benchmark comparisons documented in the NVIDIA case study, NeoData achieved at least 30–50% higher accuracy than traditional machine learning models for one of Latin America's largest private banks. The company featured on the main stage at NVIDIA GTC 2026 in March as one of the first companies in Latin America to deploy the NVIDIA GB200 NVL72 system in production.

Why the Data, Not the Model, Is the Moat

The deeper strategic logic behind the acquisition has been articulated most clearly not by Itaú, but by the research firms analyzing where AI competitive advantages actually sit.

Forrester's analysis of AI strategy in financial services identifies what banks are building as distinct from what any cloud provider can sell them: the true competitive asset is not the model weights themselves, which can quickly depreciate as new base models emerge, but behavioral data no vendor can replicate — the proprietary record linking customer actions to outcomes over time. An OpenAI model fine-tuned on publicly available financial data does not have access to two decades of a specific bank's loan repayment patterns, spending signals, and fraud decisions. Itaú does. And NeoSpace has spent two years learning to read them.

A June 2026 paper by Stanford Law School on defensible moats for vertical AI companies ranks what it calls the "Brain" layer — a proprietary data pipeline combining private customer records, real-time data, and operational insights — as one of the highest-durability competitive advantages available to industry-specific AI companies. The paper's core finding for banks specifically: each customer's accumulated behavioral context compounds over time, raising switching costs in a way that model architecture alone cannot replicate. A bank that trains a foundation model on its own data for two years, and then switches to an external provider, does not merely pay a financial cost — it loses the two years of learning that lived in the model.

That compounding informational value is precisely what Itaú is locking in by acquiring NeoSpace now rather than continuing to license its capabilities. The five-year deal structure for deferred payments suggests Itaú has priced in regulatory and integration uncertainty; the decision to buy outright, rather than negotiate exclusivity or expand the licensing arrangement, suggests it has concluded that partial ownership is not sufficient protection in a market where a competing bank could make the same acquisition.

The SEC filing's language is direct about the strategic rationale: the combination of NeoSpace's capabilities with Itaú's data, expertise, and scale "will create the conditions required to develop increasingly specialized models and expand their application across Itaú's various products and services."

How NeoSpace's Technology Already Powers Itaú's AI Stack

The acquisition formalizes a relationship that already runs through Itaú's core operations. The SEC filing notes that NeoSpace has delivered "significant and transformative results in areas such as credit, CRM, and fraud prevention" for Itaú — three domains where marginal improvements in prediction accuracy translate directly into basis points on loan portfolios and billions in risk-adjusted revenue.

The deal is also described in the filing as "directly aligned" with the development of ia.i — Itaú's flagship generative AI assistant, which launched publicly in July 2026 to 300,000 initial users. ia.i, which stands for Inteligência Artificial Itaú and doubles as a play on the Brazilian greeting "e aí?" ("what's up?"), gives customers a conversational interface — accessible by text or voice — to check balances, interpret spending patterns, seek financial guidance, and initiate transactions including Pix transfers.

Behind that conversational surface, the models NeoSpace has trained on Itaú's behavioral data are what produce the predictions ia.i surfaces to customers. The generative interface (text, voice) and the decision layer (credit, fraud, propensity modeling) are distinct stacks; NeoSpace builds the latter.

Itaú describes its current AI footprint as more than 1,300 AI models in production across its businesses. The bank has reported a 65% reduction in generative AI costs since beginning its AI investment program in 2022. The Evident AI Index for Banks in Latin America, published in July 2026, ranks Itaú as Nubank's closest rival for overall AI maturity in the region — holding the largest absolute headcount of AI professionals among incumbent banks evaluated.

Itaú joined Stanford University's Human-Centered AI Institute in February 2025 as the first Latin American company to do so.

Deal Mechanics and Valuation: A ~6x Jump in Under Two Years

The BRL 2,925,000,000 ($588 million USD) price tag represents a dramatic revaluation of NeoSpace from its entry point. When Itaú led the $18 million pre-seed round in January 2025 and took a 15% stake, the implied company value was in the range of $100–$120 million. The full acquisition at approximately $588 million represents an implied valuation increase of roughly 6x in under two years — a trajectory that reflects both NeoSpace's revenue growth and its deepening operational integration with Itaú.

By mid-2026, NeoSpace was projecting revenue above BRL 300 million (approximately $60 million USD) for the year. As recently as September 2026, the startup had been in negotiations for a new funding round of up to $250 million — a round that this acquisition has now superseded.

The $18 million pre-seed included angel participation from prominent fintech investors including Martin Escobari of General Atlantic, Micky Malka of Ribbit Capital, Nigel Morris of QED, and Hans Morris of NYCA Partners.

Payment at closing will be followed by deferred installments spread over five years, subject to certain conditions — a structure that protects Itaú's cash flow and likely ties portions of the deferred consideration to NeoSpace's post-close performance.

Regulatory Gauntlet: What CADE's AI Scrutiny Means for This Deal

The transaction faces a two-regulator approval process that is standard for significant acquisitions in Brazil's banking sector, but carries particular salience given the current moment in Brazilian antitrust enforcement.

CADE — the Conselho Administrativo de Defesa Econômica, Brazil's principal competition authority — has been closely scrutinizing AI deals throughout 2026. It ordered Microsoft to submit its Inflection AI investment for formal merger review in May 2026, investigated the Google-CharacterAI arrangement, and convened a Tribunal review of Amazon's investment in Anthropic to determine whether it triggered mandatory merger notification requirements. CADE President Diogo Thomson has been explicit that low revenue doesn't eliminate competitive relevance — in the authority's view, AI companies without significant revenue can still be competitively significant.

NeoSpace is not a zero-revenue startup — it is projecting BRL 300 million ($60 million USD) in 2026 revenue — so mandatory notification thresholds are clearly satisfied. But CADE's stated interest in AI market concentration and its active development of the D.E.I.A. (Defesa Econômica com Inteligência Artificial) program — an internal AI-powered system for reviewing AI acquisitions — means that the substantive review is likely to be more thorough than a standard merger of this size.

The Central Bank of Brazil must also approve the transaction, as it does for any acquisition involving a regulated financial institution or its affiliates. The combination of the two approval processes means the closing timeline is inherently uncertain. In recent complex fintech and AI deals, CADE and Central Bank deliberations have stretched from several months to more than a year.

The staggered five-year payment structure suggests Itaú has priced that uncertainty into the deal's architecture from the outset.

What Comes Next for NeoSpace, Itaú, and Latin American Banking AI

When regulatory approvals are granted and the deal closes, Itaú will hold full ownership of an AI infrastructure company whose models are already embedded in its credit scoring, customer relationship management, and fraud detection operations — and whose work underpins ia.i, the conversational interface Itaú is betting will define how tens of millions of Brazilians interact with their bank.

For competitors — Bradesco, Banco do Brasil, Santander Brazil, and Nubank — the acquisition raises a question that will define the next phase of AI competition in Latin American financial services: which bank controls the model layer, and which is left licensing capabilities from the outside? BBVA and Santander have pursued partnerships with external providers including OpenAI for banking AI; Itaú has now moved to own the alternative.

The deal's most significant precedent may be the one that is least visible in the headline: it establishes that two years of training an AI model on a bank's proprietary transaction data — not access to the most capable frontier models on the market — is the thing that is worth $588 million. For every financial institution now evaluating its AI strategy, that is the number to reckon with.

Currency conversions in this article are based on an exchange rate of approximately 1 BRL = 0.201 USD as of October 26, 2026; all USD figures are approximate.

Frequently Asked Questions

What are Large Data Models, and how are they different from the AI in ChatGPT?

Large Language Models like ChatGPT and Claude are trained on text from the public internet to understand and generate language. Large Data Models — the technology NeoSpace builds — are trained on structured corporate records: transaction histories, credit events, fraud signals, and behavioral data. They are designed to produce business decisions (credit limits, fraud alerts, churn predictions) rather than language. IBM and NVIDIA describe LDMs as a complement to LLMs — the missing layer between data and decisions for enterprises that need real-time inference over their own records, not general-purpose text generation. Both IBM's research on large database models and NVIDIA's case study of NeoSpace's work provide technical detail on this distinction.

Why would Itaú buy NeoSpace rather than just keep licensing its technology?

The core reason is that NeoSpace's models have been trained on Itaú's own data for nearly two years. That training history is not transferable — it is embedded in the model weights, and if another bank or a tech giant acquired NeoSpace, Itaú would lose both the capability and the institutional intelligence built into it. Forrester's analysis of banking AI strategy identifies this as the decisive factor: the true competitive asset is not the model architecture itself but the proprietary record linking customer behavior to outcomes — data no vendor can replicate. Buying NeoSpace locks in that moat permanently and prevents a competitor from making the same purchase.

What does CADE need to approve for this deal to actually close?

CADE, Brazil's antitrust authority, must determine that the acquisition does not substantially lessen competition in any relevant market — including the emerging market for AI infrastructure in financial services. Given that NeoSpace has documented revenue above BRL 300 million ($60 million USD) projected for 2026, the deal clearly meets mandatory notification thresholds. CADE has been actively scrutinizing AI-adjacent acquisitions throughout 2026, including Microsoft's Inflection AI deal and Amazon's investment in Anthropic, and has publicly stated that the absence of traditional revenue is not a barrier to competitive relevance in AI markets. The Central Bank of Brazil must also separately approve the transaction. Combined, both regulatory processes have stretched from several months to more than a year in comparable recent deals.

How does NeoSpace's technology connect to ia.i, Itaú's AI assistant?

ia.i is Itaú's conversational interface — text or voice — that customers use to check balances, understand spending, and initiate transactions. The conversational layer is the product customers see. The predictions ia.i surfaces — what your spending trend means, whether a credit offer is likely to perform, whether a transaction looks anomalous — are generated by the decision models that NeoSpace has trained on Itaú's transaction data. The SEC filing explicitly describes the acquisition as "directly aligned" with ia.i's development, and the ia.i launch announcement confirms that NeoSpace's models are the intelligence layer underneath the product.