Korva is a scroll of the internet’s best ideas.
Speed and licenses
Speed is a startup's biggest advantage in the AI era, especially one born unregulated. In banking that only carries it so far, because eventually the company needs regulation too. Nubank holds banking licenses in several countries, so it has to check with regulators and move at a set pace.
The data advantage
Data is the other moat. Nubank has 140 million customers, with something like 20 to 30,000 data points on each, across 13 years. That volume of information let it build its own foundation model for financial services, used for credit decisions, collection decisions, and now fraud decisions in Mexico and Colombia. A startup with speed would have a much harder time reaching that level of data.
The Hyperplane acquisition
The model began with an acquisition. Nubank bought Hyperplane, a Palo Alto company building foundation models for financial services, which was struggling to sell to big banks because of two-year sales cycles. Nubank believes it was the first financial services firm to build its own foundation model on proprietary data for credit decisions, and it published its research. As Nubank launches in the US and other markets, the model comes along from day one.
Why do banking licenses take years to get?
Regulators review a company's capital, the backgrounds of its owners and managers, its risk controls, and its anti-money-laundering systems before granting a charter. The review also covers whether the business plan is sound and whether deposits will be safe. Applicants often revise their plans during review in response to regulators' questions. In the United States, new bank charters are approved by federal and state authorities, and the process often runs well over a year.
How can a model learn from financial transactions?
Language models predict the next word in a sequence. A foundation model for finance can treat a customer's transactions as a similar sequence of events and learn patterns in how they unfold. A single model trained this way can then be adapted to credit, collections, or fraud tasks instead of building a separate model for each. Researchers at OpenAI described in 2020 how model performance improves predictably with more data and computing power.
Why do banks take so long to buy software?
Banks run security reviews, vendor due diligence, and compliance checks before they adopt a new tool. US regulators issued interagency guidance on third-party risk management in 2023, requiring banks to oversee the vendors they rely on. Several departments each need to sign off, so sales cycles at large banks can stretch to a year or two. A small startup can run out of money before the process ends.
More from Invest Like the Best with Patrick O'Shaughnessy
