David Vélez on Nubank's AI Strategy and Banking's Future

Imagine a store where nine out of ten customers still drive across town to rent a physical video, years after a mail-order service proved people would rather have it delivered. That is David Vélez's picture of global banking today. Vélez, founder and CEO of the Brazilian digital bank Nubank, tells Patrick O'Shaughnessy that despite a wave of fintech startups over the past decade, incumbent banks still process about ninety percent of the ten trillion dollar global consumer financial services industry (02:38). "It's still almost as if ninety percent of the world was going to Blockbuster to get VHS and DVDs, and Netflix-like services still had something like ten percent of that industry," he says (03:03). The analogy sounds like a founder's humble brag. It is actually the spine of the whole conversation: a argument that the transition from physical to digital finance, the kind of shift that already remade media, retail, and transportation, has barely started in banking.

Nubank is Vélez's evidence. Since the two last spoke five years ago, the company has grown from thirty-five million customers to one hundred forty million (01:39), and in Brazil it now reaches roughly sixty-five percent of the adult population (03:31). More telling than scale is behavior: sixty percent of Nubank's customers use it as their primary bank account, not a secondary wallet (03:49). The company has not just added users. It has replaced the incumbent.

The cost structure nobody sees

The mechanism behind that replacement is unglamorous: real estate. Nubank serves its customers with ten thousand employees, while competitors need roughly one hundred thousand employees and six thousand branches to serve a smaller base of one hundred million (05:32). The result, Vélez says, is that Nubank's cost to serve a single customer runs at about three percent of a traditional bank's cost (05:32). That gap is the entire business model: it buys enough margin to drop fees, pay better interest, and still turn a profit.

Vélez organizes the opportunity that follows into three buckets: access, cost, and intelligence (06:45). On access, he points out that three billion people worldwide still have no credit product at all (06:26), a gap he attributes to the slow digitization of financial data rather than a lack of demand. On cost, he offers a number designed to unsettle: U.S. banks hold about fifteen trillion dollars in retail deposits but pay an average yield of only forty basis points, even with interest rates near five percent (08:19). Do the arithmetic and that gap is worth roughly six hundred billion dollars a year, money that in a digital banking system would stay in depositors' accounts instead of funding branch networks.

"This is a massive wealth tax, the fact that banks don't pay any yield. If you're high income, you do. The lower your income, the less you get paid." — David Vélez [08:43]

The third bucket, intelligence, is where Vélez gets most provocative. Financial literacy campaigns, he argues, have failed for a mundane reason: people do not show up to the class. His proposed fix is not more education but an AI system that manages the decisions people would otherwise need expertise for, the way a private banker already does for wealthy clients (09:52). "You're not gonna solve this problem of intelligence through education. You're gonna solve it through AI," he says (10:07). Nubank already runs a version of this for fifty million Brazilian customers, flagging when a bill at another bank is about to trigger a steep penalty rate and offering cheaper credit in its place (12:06).

Why data, not speed, is the real moat

Asked how he would attack his own company if he were a new founder with nothing to lose, Vélez gives a candid answer: speed. An unregulated AI-native startup could move faster than Nubank, which now holds banking licenses across several countries and has to clear regulatory review for new products (13:55). But he argues that advantage runs out quickly in banking, because licenses take years to obtain and no AI model can shortcut that. The deeper moat, he says, is data: Nubank holds twenty to thirty thousand data points per customer across thirteen years for one hundred forty million customers (15:02), the raw material for its own proprietary foundation model, a large AI system trained specifically on Nubank's own financial data rather than general internet text. Nubank built this model after acquiring a Palo Alto startup called Hyperplane, whose technology had struggled to sell into slow-moving incumbent banks (18:14). Vélez says Nubank was likely the first financial services company to build its own foundation model for credit decisions and has since extended it to fraud detection in Mexico and Colombia (18:34).

That data advantage is also why Vélez pushes back on a simple reading of his own Blockbuster analogy. Blockbuster had enormous amounts of customer data too, he notes, but no system or culture built to use it, which is why Netflix could out-build it with far less history (19:50). Scale alone does not create an edge. Scale paired with a culture that treats every decision as a modeling problem does.

Flattening the company to see the truth

The most concrete operational claim in the conversation concerns Nubank's own management structure. As the company scaled from one million customers to one hundred forty million, it organically added layers of management to handle the growing pains, eventually reaching ten hierarchical levels (48:17). Vélez says AI tools, paired with a deliberate choice, let the company cut that down to five layers, while he personally went from one direct report to twenty (47:31). The logic draws on an idea he calls the dictator's dilemma: the more layers a message travels through, the more distorted it becomes by the time it reaches the person at the top (49:02).

"Simply by removing a lot of those layers, you are much closer to the ground truth, and you reduce a lot of those layers by having information systems that have access to everything." — David Vélez [49:02]

Mapping that shift took work: Nubank identified one hundred twenty internal workflows, from coding to compliance to product design, and rebuilt each one natively with AI tools, building evaluation infrastructure to measure the change at every step (46:10). Vélez frames this less as an AI story than as a forcing function: companies always have latent opportunities to redesign how they work, and the current technology shift simply makes the case for doing it now undeniable.

The conversation closes on a distinction Vélez clearly thinks matters more than any metric: the difference between a founder and a chief executive. A founder, in his telling, is impatient, wants to move fast, and resists the slow work of coaching a team. A CEO has to build systems and be patient (59:49). Most people assume the CEO version is the better manager, he says, but the company's long-run advantage comes from staying closer to founder mode, especially during a technology shift like the current one in AI, when the instincts that built the company from nothing become more valuable than institutional polish (60:27). It is a fitting place to end a conversation that began with a store nobody visits anymore. The argument throughout is not that incumbents are incompetent. It is that an industry built around physical branches, inherited trust, and static yield never had to ask what a bank looks like once branches, trust, and yield can all be rebuilt from data instead of real estate.