Itaú Unibanco has agreed to pay R$2.925 billion, about $589 million, for all of NeoSpace, the startup whose foundation models it already runs for credit, CRM and fraud prevention. Latin America's largest bank decided its predictive model layer was worth owning outright, and any other bank, insurer or telco using NeoSpace now buys from a company its competitor owns.
The deal was disclosed on October 26, 2026 in a material-fact filing to Brazil's securities regulator, the CVM. Reuters reported that part of the price is due at closing and the remainder is deferred over five years, and converted the figure at 4.9654 reais to the dollar. Closing needs approval from both CADE, Brazil's antitrust authority, and the Banco Central do Brasil, per the filing summary. Founder Bruno Pierobon stays as CEO.
If you run credit, propensity or fraud models at a bank, the question this deal raises for you is direct: do you rent the model layer that decides who gets an offer, or do you own it?
What Itaú Is Actually Buying
Itaú is buying a team and a model family it already depends on. NeoSpace has worked for the bank for almost two years, building models trained on Itaú's own data that are already in use for credit, CRM and fraud prevention, Money Times reported from the bank's announcement.
A tabular foundation model is a large model pretrained on rows and columns (transactions, account events, bureau fields) that can be fine-tuned for a specific prediction such as default risk or the likelihood that a customer takes an installment loan. It plays the role a gradient-boosted tree has played in bank analytics for a decade, with one pretrained base shared across many tasks.
NeoSpace's platform is called NeoData. NVIDIA's customer case study describes it turning tabular, event and unstructured records into real-time predictions for banks, telecom operators and insurers, with NeoSpace deploying GB200 NVL72 systems in production through Oracle Cloud Infrastructure. The same page says NeoData runs for an unnamed Latin American private bank with more than 60 million customers, and reports a 19% lift over NeoSpace's own baseline model after fine-tuning, without tying that figure to a named deployment. Those are vendor-reported figures in a supplier's marketing story, and nobody outside the two companies has checked them.
The stated rationale is control. Itaú says the purchase expands its capacity to build its own models by combining NeoSpace's technical work with the bank's data and scale, Exame reported, and that the models will also feed ia.i, its generative AI assistant. CEO Milton Maluhy Filho said he wants Itaú to be among the institutions that define the next chapter of the global financial industry.
How Fast the Price Climbed
On reported figures, the price rose roughly sixfold in under two years. In January 2025 Itaú took a 15% stake in NeoSpace through an $18 million round it led, InfoMoney reported at the time, alongside a commercial agreement for Itaú to use NeoSpace's base models and co-develop exclusive products for its digital retail platforms. Itaú did not disclose what it paid at the time; Exame now reports $15 million for the 15%, citing Bloomberg Línea.
If those reported figures hold, the January 2025 stake valued NeoSpace at about $100 million. This week's price for the whole company is about $589 million. That is our arithmetic on reported numbers, and the deferred portion over five years "subject to specific conditions" means the cash actually paid may land below the headline.
Steel-man the price, because it is defensible. Itaú is paying for models already fitted to its own customer data and already in production, a team that knows its systems, and the right to stop sharing any of it. Pierobon also founded Zup, which Itaú agreed to buy in 2019, so the bank has bought this founder once before and kept doing business with him.
Who Else Was Using NeoSpace
The public record names only one NeoSpace customer, Itaú, and every other customer should be asking what happens now. Itaú's filing, as summarised, mentions no other clients and no exclusivity terms.
But NeoSpace did sell beyond Itaú. The NVIDIA case study describes a second deployment at "another major Latin American financial institution": a propensity model for installment credit that reached 6 million customers in production and was trained on more than 10 TB of data covering over 50 million customers, with a 10% increase in installment credit offers and a 10 to 20 point KS improvement. NVIDIA does not name that institution. If it is a separate company from the 60-million-customer bank in the same story, its supplier is about to become a subsidiary of the country's largest private bank.
That is the uncomfortable position. A rival bank's AI team will own the company that holds your fine-tuning pipeline, has seen your feature schema, and decides which model improvements ship first. The filing says NeoSpace will keep autonomous management and its current team. Autonomy is a governance promise from the acquirer; it says nothing about whether NeoSpace will keep serving outside customers, at what price, or on what roadmap.
What Happened the Last Time a Bank Bought Its Model Vendor
The closest precedent points one way: the buyer absorbs the vendor and the outside customers stop being the story. In June 2024 Nubank acquired Hyperplane, a Silicon Valley startup whose platform let banks train, evaluate and deploy deep-learning models on their own first-party data for risk, collections and marketing. Nubank said Hyperplane would power its core machine learning and help build the foundation models its product teams use.
Nubank's announcement did not say what happened to Hyperplane's existing bank clients, and contemporary coverage said only that the team would continue its work at Nubank. When a bank buys a tool built for banks, the external roadmap is the first thing nobody mentions.
The Itaú case differs in one respect that helps NeoSpace's other customers: Itaú explicitly says NeoSpace keeps operating independently, with its CEO, team and R&D intact. Treat that as the strongest card you hold in a renewal conversation, and get it written into your contract rather than relying on a press summary.
Rent, Own or Use an Open Model
For most banks the realistic choice is between three routes, and Itaú's deal removes one supplier from the rental shelf. You can rent a vendor's pretrained tabular model and fine-tune it; you can own the layer by building or buying a team, as Itaú and Nubank did; or you can run an open tabular foundation model on your own infrastructure.
The open route is more viable than it was two years ago. Prior Labs, the company behind TabPFN, raised a €9 million pre-seed round in February 2025 and claimed its model beat state-of-the-art methods on small tabular datasets in over 96% of cases. We have compared the current open options in Mitra vs TabPFN vs XGBoost and NVIDIA's Kumo Tabular against TabPFN's licence, and where foundation models beat tuned XGBoost. The short version: the public benchmarks where they win are mostly small datasets. A bank scoring tens of millions of customers is outside that range, so the open models have to be tested on your own data before anyone assumes they match a vendor fine-tuned on it.
Owning is what Itaú chose, and it costs a figure most banks cannot spend on a model layer. A platform such as DataRobot sits in between: you own the models, the vendor owns the tooling. The supplier risk is smaller, because a change of control at the platform vendor does not move your trained weights.
Questions for Regulated Buyers
A model vendor's change of control is a model risk event as well as a procurement event. US supervisors have been asking banks how they govern third-party AI (see their recent supervisory questions), and Singapore's MAS now expects an inventory of AI embedded in third-party tools. A credit model whose developer is now owned by a competitor is exactly the dependency those inventories exist to surface.
Three questions matter. Who owns the fine-tuned weights trained on your data, you or the vendor? Can you export them, with the feature pipeline, in a form your own team can run? And does the vendor's new parent get any access, even aggregated, to what your data taught the base model? We made the same argument when Visa bought BioCatch, a fraud vendor whose bank customers also had to ask for neutrality terms in writing.
What to Do About It
This Week:
- Check whether NeoSpace, NeoData or a reseller of either appears in your vendor inventory, including pilots run by a single analytics team.
- If it does, pull the contract and find the change-of-control, assignment and termination-for-convenience clauses. Note the notice period.
- Ask your NeoSpace account team, in writing, whether service to non-Itaú customers continues after closing and on what terms.
This Month:
- Get written answers on weight ownership, export format and whether any model trained on your data, or improvements derived from it, can flow to Itaú.
- Have model risk re-tier any NeoSpace-built model in production and document a fallback, typically the incumbent gradient-boosted model, with a tested switch-over.
- Benchmark one live use case against an open tabular model on your own data, so you know the cost of leaving before you need to.
Before Closing:
The deal still needs CADE and central bank approval, which gives you leverage you lose once it closes. Put data separation, export rights and a service-continuation commitment into an amendment while NeoSpace is still negotiating as an independent company.
The Bottom Line
This is the second time in just over two years that a large Brazilian bank has bought a startup that builds foundation models for banks, after Nubank and Hyperplane in 2024. Both buyers can afford a dedicated model team; most banks cannot, and will keep renting from a shrinking set of independent suppliers. Itaú has decided the model that chooses who gets credit is too close to its business to rent, and it paid about $589 million to stop renting it.
If you rent yours, find out this month who owns the weights trained on your customers.
Continue Reading
- Mitra vs TabPFN vs XGBoost Puts Mitra-v2 Ahead of Tuned XGBoost
- Foundation Models Beat Tuned XGBoost Below 50,000 Rows
- NVIDIA's Kumo Tabular Undercuts TabPFN's License Below 60,000 Rows
- Visa Bought BioCatch. Get Neutrality in Writing.
- US Regulators Ask Banks to Explain Their AI Agent Safeguards
- Best MLOps for Regulated AI: Domino, Then a Sign-Off Layer