What this covers

Patrick talks with David Vélez, founder and CEO of Nubank, about how a bank with 140 million customers and 10,000 employees is rebuilding itself around AI. The most concrete material is internal. After a code red over missing the moment AI began to work, Nubank mapped 120 workflows, from know-your-customer to internal audit, and redid each with AI tools and eval infrastructure. It also cut management layers from ten to five, raised the CEO's direct reports from one to 20, and moved from fully remote to hybrid. Vélez also discusses founder versus CEO temperaments and AI-driven employee feedback.

The banking economics come next. Vélez calls the gap between US deposit rates and market rates a roughly $600 billion annual "tax" that falls hardest on lower-income savers. He pictures AI giving everyone the private banker only the wealthy have now, and open banking aggregators making fees visible and switching easier. He argues that data, with 20,000 to 30,000 data points per customer, supports a proprietary foundation model for credit, and that licenses slow AI-native entrants.

The more contested turns are about money. Vélez forecasts that today's roughly 190 currencies could shrink to about ten, with billions banking offshore in stablecoins and Bitcoin. He also expects the dollar to share reserve status. On credit, he describes Nubank's $20 starting lines and argues that digitizing cash transactions, like a taco stand's, creates underwritable cash-flow history.

Sharpest takeaway

David Vélez argues that incumbent banks are in a Blockbuster-era position: branch-based cost structures let digitally native banks like Nubank undercut fees and pay better yields, and AI widens that gap, so market share and profit pools will migrate to them globally.

  • Nubank's cost to serve is about 3% of an incumbent's, so it can pass efficiency back to customers as no fees and better yields
  • AI multiplies this efficiency and removes banking profit pools, while data, licenses, and culture become moats
  • Open banking, real-time payments, and stablecoins reduce customer inertia and bring the US and other markets into play

The argument · threads14 threads · 69 claims
0.72

Lower cost to serve lets digital banks undercut incumbent branch-based economics.

5 pointscentrality 5/5
  • Nubank serves 140 million customers with 10,000 employees, while competitors serve about 100 million customers with 100,000 employees and 6,000 branches, so Nubank's cost to serve a customer is about 3% of a traditional incumbent's, leaving margin to return to customers as lower fees and better rates.

    “our cost to serve a customer is about 3% of what a traditional incumbent bank's is.”

  • Nubank has surpassed the largest Brazilian incumbent in scale, with close to 120 million Brazilian customers, over 65% of the adult population, and about 60% of its customers use it as their primary bank account rather than a secondary wallet.

    “About 60% of our customers use us as their primary bank account. We replace the incumbent bank. We're not a secondary wallet.”

  • Big incumbent banks in Mexico, Brazil, and similar markets pay zero yield on about 70% of their deposits and pay some yield on the remaining 30%, mostly to the highest-income customers and corporates, so the cheapest funding is used to subsidize branch infrastructure.

    “The big incumbent banks, about 70% of their deposits, they pay zero, and 30%, they pay some yield, generally to the highest income population or the corporates.”

  • AI will further multiply banking efficiency so that banks could serve a billion customers with 5,000 employees, and digital banks can either harvest that efficiency as margin or reinvest it to expand ambition into 100 countries.

    “You could imagine banks that have a billion customers with 5,000 employees. That's within sight.”

  • The end of cash, driven by real-time payment systems like Brazil's Pix and India's UPI, removes the need for branches and increases the speed of money movement and the amount of data created in an economy, which can then be used to extend credit.

    “Cash goes away, disappears. Enter a number of real-time payment systems, as you've seen in Brazil with Pix or India with UPI.”

0.57

Founder's personal care and consumer obsession drive Nubank's leadership, founding, and success.

9 pointscentrality 3/5
  • Vélez is motivated by impact and sees himself as a wartime general rather than a peacetime general, so disruptions like AI energize him to execute, and he concluded after one to two years of reflection that no other pursuit could have more impact than scaling Nubank to 100 countries and a billion customers.

    “I find that I'm more of a wartime general than a peacetime general.”

  • Vélez initially hired experienced external managers believing the company needed them, but concluded the organization actually needed more of his own clear leadership, motivation, and big-thinking, and that he should keep pushing hard beyond what the company may feel ready for.

    “What the organization needed was more of what I had already brought into the organization, which was this clear leadership, the motivation, the willingness to think really big.”

  • The most important lesson of entrepreneurship is that the founder who created a company cares about it more than anyone else, a visceral level of care that makes them the right leader for a long time, regardless of their expertise, which helps overcome imposter syndrome.

    “I don't think anybody in the world cares as much for this company as I do, to a level that is almost visceral.”

  • Building an organic brand in the US, where consumers rank brand as a key attribute in choosing banks, is likely the hardest challenge, and Nubank wants to do it more capital-efficiently than heavy marketing spend.

    “Building a very strong brand will be one of the hardest things. US consumers mention brand as one of the key attributes to choose banks, and it's not cheap.”

  • The hardest part of leading for a long period is imposter syndrome as the business grows, which Vélez says must be resolved through reflection, experience, and feedback, and he resolved it by concluding he is probably the right CEO while still hiring complementary skills.

    “The imposter syndrome. It's fine to be a founder at the beginning, but then suddenly when you see yourself running a very large organization”

  • Nubank has studied Amazon for consumer obsession and Disney for creating magical experiences, and applies this to banking by designing even mundane product touchpoints, such as the US welcome kit for first cards, as magical moments.

    “we study Disney very carefully because this is a company that wants to create magic experiences.”

  • Scaling a company will be far more painful than anyone expects, so founders need a mission they are truly excited about, one they cannot stop thinking about, because conviction is what carries them through.

    “Scaling a company won't be easy, will be much more painful than anybody anticipates.”

  • Nubank began as a reaction to the poor consumer banking experience in Brazil that Vélez personally felt, which he took as a proxy for the broader population's dissatisfaction, and this made consumer obsession the company's founding principle.

    “We began the company as a reaction of a consumer pain”

  • A CEO should go directly to whoever has the best information or does the work, ignoring the chain of command, which Vélez contrasts with traditional organizations that require strict channel respect.

    “I'll go to whoever has access to the best information if I have a question”

0.55

Incumbent banks' legacy systems and culture leave them unable to match digital entrants.

4 pointscentrality 4/5
  • Traditional banks operate with around 15 systems from acquisitions and are still converting COBOL mainframe systems to the cloud after 10 to 15 years, so AI applications for them mostly involve transforming legacy infrastructure rather than building new products.

    “traditional banks... still operating with 15 different systems as a result of a huge amount of acquisitions”

  • Blockbuster had customer data but lacked the systems and data culture to use it, whereas Netflix built a better recommendation algorithm without initial data, so the binding constraint is culture and systems, not scale of data alone.

    “Scale and data is important, but equally important is your systems and your culture.”

  • The hard part of consumer obsession is executing it when inconvenient, since companies claim customer focus but invent new fees when they see short-term revenue, and in the AI era most attention goes to efficiency rather than the consumer experience, where humans still beat AI.

    “Most companies will be consumer obsessed, but then when they are suddenly seeing themselves making more money because they invented a new fee.”

  • Some US incumbent banks now have higher return on equity than emerging-market banks, which Vélez reads as opportunity: since incumbents earn very high margins despite heavy costs, digital entrants have room to undercut them on price.

    “your margin is my opportunity. The typical Amazon view or Bezos view of that incredible ROE that even incumbent banks have”

0.55

Nubank rebuilt its workflows natively around AI tools and its own foundation model.

4 pointscentrality 4/5
  • After the code red, Nubank mapped 120 workflows across the organization, from product design and know-your-customer to internal audit and regulatory capital management, reinvented them natively with AI tools, and built eval infrastructure to measure each step.

    “we mapped out 120 different workflows across the entire organization... 120 workflows were reinvented natively using AI tools.”

  • With its own foundation model, Nubank cut the cost of adding and testing a feature from about six months to a much shorter cycle, so that nearly every organizational decision can be automated and made by a model rather than a human.

    “Pre-AI, you had to spend six months adding a new feature... every decision in the organization can now be automatized and be done by a model versus a human.”

  • Nubank requires senior managers to build actual software products every quarter so they understand AI capability firsthand, on the principle that everyone should use the tools actively rather than just read about them.

    “We ask all of our senior managers to build products every quarter, actual software and products, so that we are integrated and very aware of what's going on.”

  • Over five years Nubank moved from fully remote to hybrid with significantly more office time, because AI can do the individual tasks that remote workers did while physical presence remains necessary for collaboration and coordination, so AI increases rather than decreases the value of in-office work.

    “AI increases the argument to spend time in person more, because increasingly the work that AI can do is more or less the work that people at home individually could do.”

0.51

Scaled neobanks' data, trust, and licenses give them AI advantages startups cannot replicate.

7 pointscentrality 3/5
  • When cash-based transactions like a Mexico City taco stand become digital ledger records, those small sellers build a recorded cash flow history that banks can underwrite, enabling credit and expansion; with about 70% of Brazilian employment in small businesses, this could mean trillions of dollars of credit for billions of people.

    “The moment that that transaction goes into a digital ledger... that taco seller can take that information to a bank and says, hey, look at my cash flow, gimme access to credit.”

  • AI-native startups have an advantage in speed and in launching products quickly, especially if unregulated, but in banking regulation caps that speed, and no model can obtain a banking license or shorten regulatory approval from two to three years to six months, so licenses and data become incumbent advantages.

    “you still cannot really get ChatGPT to get you a banking license. And you still cannot really get them to have regulators give you licensing in six months instead of two to three years.”

  • Nubank holds roughly 20,000 to 30,000 data points per customer across 13 years for about 140 million customers, enough data to build a proprietary financial services foundation model for credit and collections decisions that a new startup would struggle to replicate.

    “For each customer, we have something like 20,000 to 30,000 data points across 13 different years.”

  • About 3 billion people globally have no access to credit, and in Mexico and Colombia credit card penetration is only about 13%, a gap Vélez argues will close as data, underwriting sophistication, and digital rails mature, potentially making 100% of populations and small businesses banked and credit-served within 10 to 20 years.

    “you still have 3 billion people in the world that don't have access to a credit product.”

  • Scaled neobanks are in a sweet spot to benefit most from AI because they already possess scale, data, brand, trust, and licenses that startups lack, while still retaining the technology infrastructure to exploit AI faster than before.

    “We are doubling down on the larger neobanks or neo-fintechs that are now right in the sweet spot”

  • Nubank began using machine learning in 2013 to price credit risk for Brazilian consumers paying 500% to 600% annual rates, and chose to take credit risk on its own balance sheet, making it capital intensive by default and building a model-driven organization early.

    “We started using machine learning since we started the organization, since 2013, and we looked at Brazilian interest rates for consumers paying 500% and 600% a year.”

  • Early-stage AI-native startups will have speed but will struggle to reach the scale of customer data needed for credit decisions, so data becomes a moat for established digital players, while entrants may need to compete through applications layered on top of AI models.

    “The early-stage startup entrepreneur, while they would have the speed, would have a much harder time getting to that level of data.”

0.51

Nubank wins by passing efficiency and financial intelligence to consumers, taking incumbent share.

7 pointscentrality 3/5
  • The bull case is that Nubank captures incumbents' share over decades using technology tailwinds like real-time payments, open finance, and AI, then passes efficiency gains to customers across 100-plus countries, eventually addressing the $10 trillion profit pool.

    “if we have a cost structure that is 50x theirs, and this is pre-AI”

  • Once consumers realize they earn nothing on balances, the roughly $600 billion now kept on incumbent balance sheets will be redirected to depositors, with compounding making the effect larger over time.

    “The first obvious rearrangement is those $600 billion that banks are keeping in their balance sheet... get forwarded to all the banking consumers”

  • An AI copilot can tell customers which spending, saving, and investing moves earn credit points and raise creditworthiness, acting as a coach that helps people improve their access to credit over time.

    “you enable an AI copilot for that customer that tells them all the different moves”

  • Nubank proactively contacts about 15 million Brazilian customers to flag credit card bills due at other banks, offering money to avoid 15% daily interest or refinancing loans from 10% to 5%, which removes profit pools on the consumer's behalf.

    “about 15 million of our customers every day — we're telling them, hey, your credit card in that other bank is due tomorrow.”

  • Financial education programs have never solved the intelligence gap because people do not sit down for classes, so the problem can only be solved by delivering personalized AI guidance inside the product.

    “You're not gonna solve this problem of intelligence through education. People don't sit down to do classes.”

  • Banks concluded people did not want credit cards because they feared taxes and would not use the product, but the real gap was a product matching their needs, not a lack of desire for credit.

    “what they were really lacking was a product that really matched their needs.”

  • Banking is ultimately a commodity, so the best product with no fees should win all the market, but it doesn't because of customer inertia, and open banking tools like Plaid can break that inertia by connecting 120 million accounts for intelligence layers to optimize fees.

    “banking is a commodity product ultimately, so you should be able to get 100% of the market share. That doesn't happen because there is a lot of inertia in the system.”

0.41

Open banking and AI will compete away banks' fee-based profit pools.

4 pointscentrality 3/5
  • US banks hold about $15 trillion in retail deposits but pay on average about 40 basis points, so at 4.5% to 5% market rates roughly $600 billion a year is kept by banks to fund branches, a hidden wealth tax that falls hardest on lower-income savers.

    “there's about $15 trillion of retail deposits in the US banking system, but banks pay on average about 40 basis points of yield a year.”

  • Combining open banking, which connects over 120 million accounts through aggregators like Plaid, with AI that shows each customer where fees come from, lets banks compete away their profit pools, as Nubank does for about 15 million Brazilian customers by offering refinancing and credit card payoff help.

    “you put intelligence on top of that, you're effectively allowing each customer to know how to optimize their financial life and where the fees are coming from.”

  • Remittance fees of 5% to 10% and monthly account fees of $5 to $15 in the US and similar fees in Europe and Japan are a cost that will be competed away, because moving money is just moving bits and trends toward zero in a competitive environment.

    “Moving money is moving bits. That will trend towards zero in a more competitive environment.”

  • Nubank charges no fees and so has less revenue than incumbents, yet its return on equity is two to three times theirs, which Vélez argues shows a cost structure advantage and a technology-driven moat.

    “we end up with a return on equity that is 2 to 3x their return on equity. That basically means that you found a cost structure advantage.”

0.38

Open banking and the end of cash will fragment the entrenched banking oligopoly.

3 pointscentrality 3/5
  • Global consumer financial services is roughly a $10 trillion industry, about 90% of which still runs through bank branches and incumbent banks, so neobanks and fintechs together hold less than 10% of the market, which is early like Netflix's share of video rental in the Blockbuster era.

    “If you look at global consumer financial services, it's about a $10 trillion industry, and still something like 90% of that industry goes through banking branches.”

  • Open banking and open finance let consumers own and share their data across institutions, which reduces the inertia that incumbents have profited from in banking, telecom, and utilities, and creates competition for consumer data and less concentrated banking.

    “The moment that you can connect all your history from all your services and pass it to somebody else, you're reducing that inertia and you're enabling more competition for your data.”

  • Banking worldwide remains an oligopoly, with five or six US banks holding 60% to 70% of all assets, and similar concentration in Europe, Asia, and Latin America, which open banking should fragment.

    “You have five, six banks in the US that own 60, 70% of all the assets. Same in Europe, same in Asia, same in Latin America.”

0.34

Nubank's scaling across markets shows digital banks can win underserved customers.

6 pointscentrality 2/5
  • Nubank became the largest private financial institution in the entire Latin American continent and passed incumbent banks in Brazil after a decade of being told it could not compete, which Vélez describes as the most emotional milestone.

    “Becoming the largest private bank in the entire country was something incredible.”

  • Playbooks from Brazil transferred to Mexico more than expected, with the main difference being cash preference in less-banked countries, where Nubank as a digital shark struggled without branches or ATMs for a couple of years before becoming Mexico's second largest bank and largest digital bank.

    “I always say that we are a shark in digital waters. When we are forced to compete in the offline world with banking branches, we're much less competitive.”

  • Banks believed people did not want credit, but Nubank became the first credit card for 30 million Latin Americans by offering a $20 starting credit line that grew to $100 and $1,000 over time, building credit history that AI copilots can then coach customers to improve.

    “We were the first credit card for 30 million people in Latin America... people really appreciated that, that $20 was better than zero.”

  • The key metrics for evaluating Nubank are customer count growth, net promoter score, and cost to acquire customers measured as LTV to CAC, because growth can be bought with marketing spend, so sustainable growth must be assessed against acquisition economics.

    “What is your LTV to CAC, and ultimately what is your return on equity?”

  • Nubank announced it is launching a US market entry and a stablecoin-backed account in up to 35 countries as part of its thesis that global retail banking will be transformed by digitally native companies.

    “we also announced launching to upwards of 35 different countries a stablecoin-backed account”

  • The bear case for Nubank is that it has saturated its three Latin American markets, holds about 55% of Brazil, cannot internationalize successfully, and will run out of Brazilians to acquire given low natality, leaving it unable to exploit AI opportunities.

    “We're going to run out of Brazilians. We did not succeed internationalizing.”

0.34

AI is transforming how companies build products, serve customers, and gather feedback.

6 pointscentrality 2/5
  • Nubank's early AI tests for Clojure programming were poor through 2025, until a December 2025 model release (Opus 4.6 and newer GPT models) made AI work well for Clojure, and Nubank did not catch the switch for about 60 days, prompting a code red across the company.

    “And so we were testing — and Opus 4.6 came in in December of '25, and that was sort of the moment that a lot of people suddenly realized, no, this actually works.”

  • Nubank acquired the Palo Alto company Hyperplane, which had struggle selling foundation models for financial services to incumbents with two-year sales cycles, and used it to build what Vélez claims was the first financial services foundation model trained on proprietary data for credit decisions, now used for fraud in Mexico and Colombia.

    “We think we were the first financial services firm in the world to build our own foundation model with its own proprietary data to do credit decisions specifically.”

  • Annual or semiannual feedback is distorted by recency bias and is not actionable, whereas AI copilots can give every employee daily specific feedback, such as reviewing engineers' pull requests, and give managers real-time information that makes one-on-ones low-information meetings unnecessary.

    “Olympic athletes have somebody that literally every day is coaching them and providing feedback. That is the opportunity that ultimately AI is going to create for every single person in an organization.”

  • AI agents become a customer segment as important as consumers, requiring products, marketing, and SEO aimed at agents, though in banking trust and branding still require marketing to humans, so both must be pursued at once.

    “I think agents become a customer segment as important as consumers as a segment, and you have to provide a product for the agent and you have to market to the agent and you have to do SEO for the agents.”

  • AI lowers the barrier to starting companies, so founders who pursue easy problems are displaced, and the best founders are now going after impossible areas such as hardware and space, where a solver can believe only they can succeed.

    “I think that's why some of the best founders that we're seeing right now are going for the impossible, truly. There is this shift now into hardware and space”

  • Investor Micky Malka told Vélez he is doubling down on large neobanks that combine AI transformability with scale, data, brand, and licenses, while early-stage entrants will need to find less obvious entry paths like applications on top of AI models, since building a consumer brand with hundreds of millions of customers is now hard.

    “We are doubling down on the larger neobanks or neo-fintechs that are now right in the sweet spot”

0.27

AI-driven visibility lets leaders flatten management layers and widen spans of control.

4 pointscentrality 2/5
  • Founders are impatient, urgent, and typically better as individual contributors than managers, while CEOs must be patient, systematic, and systems-oriented; the best leaders combine both, and founder-led companies outperform in technology shifts because founder advantages are most valuable then.

    “I think the key for me, though, has been to realize that the CEO title itself and the founder title, you can redesign in different ways too.”

  • Each management layer distorts messages, so by the time information passes through ten layers it is badly altered; cutting layers and deploying an AI-based company brain that is proactive and reactive gets the CEO closer to ground truth, though culture must also allow direct contact with those doing the work.

    “Every part of that chain, there is a possibility for a message to get distorted, and you get the distorted vision by the time it navigates ten different layers.”

  • Great leadership combines a differentiated vision that is sold to inspire people with active listening and humility, and the hardest part is navigating between revolutionary rejection of consensus and being consumed by it, since the status quo often persists for good reasons and leaders must identify which elements to keep and which to change.

    “And I think leadership is about knowing what doesn't have to change and pushing to change the things that have to change.”

  • Nubank cut management layers from ten to five and raised the CEO's direct reports from one to 20, arguing that AI-driven visibility allows wider spans of control and fewer managers, while scaling had caused an organic buildup of layers as managers were added to solve fires.

    “We went from ten layers to five, so the organization went very horizontal.”

0.27

Nubank's Latin American model proves digital banking can travel globally as products homogenize.

4 pointscentrality 2/5
  • Nubank's 2013 seed deck thesis was not specific to Brazil or Latin America but that digitally native companies would transform global retail banking by solving cost, access, and intelligence problems, and 13 years in Latin America were used to prove the model before going global.

    “The thesis really is that global retail banking will be transformed by digitally native companies.”

  • Certain high-touch products like mortgages, large investments, and some small business needs still value a physical human interaction, but that applies to perhaps 5% of cases, while savings accounts, credit cards, and loans no longer require branches.

    “But that might be 5% of the cases today, versus if you have a savings account, a credit card, or a loan, you just don't need that anymore.”

  • As countries digitize, financial products become more homogeneous globally, as with social media, Netflix, Spotify, Uber, and Amazon, so cultural particularities matter less than in the past.

    “The future looks much more homogeneous than the past, even for financial services.”

  • Nubank says it is now the largest private financial institution in Latin America, with the highest net promoter score and the highest return on equity in Brazil, which it treats as evidence the model works and justifies going global.

    “We're now the largest private financial institution in the entire continent.”

0.24

Stablecoins and AI advisors will reshape money and democratize access to financial advice.

3 pointscentrality 2/5
  • Of the roughly 190 currencies in the world, Vélez expects ten to 20 years from now there could be only about ten, because fiscal conditions in many countries are so dire that much fiat money is not real money, and billions of people will bank in an offshore system of stablecoins and Bitcoin.

    “Today, there's about 190 currencies around the world. My view is that ten, 20 years from now, those 190 will be maybe ten”

  • US dollar stablecoins may help preserve the dollar's reserve currency role by democratizing access to dollars, because US fiscal problems are less severe than those of the 100-plus other countries, though Vélez expects the dollar to share reserve status with five or six fiat currencies and some crypto tokens.

    “You can make the case US stablecoins ultimately will be the big savior or one of the biggest saviors of the US dollar because it democratizes access to the US dollar.”

  • Wealthy people get a private banker who optimizes deposits, tax structure, and loan pricing, while the bottom 99% get nothing; AI can provide every person a superintelligent financial advisor, which competes away today's banking profit pools, because financial education has historically failed to deliver this.

    “You're not gonna solve this problem of intelligence through education. You're gonna solve it through AI.”

0.07

Bitcoin and stablecoins will form parallel offshore banking systems beside distrusted domestic currencies.

2 pointscentrality 1/5
  • Vélez has long been a Bitcoin supporter and believes it plays a role in a diversified investment portfolio.

    “I've always been a Bitcoin fan and it plays a role in a diversified investment portfolio.”

  • People increasingly bank in a combination of their domestic fiat system and a global offshore system including stablecoins and Bitcoin because they distrust their own currencies and banks, which will create two parallel interrelating systems.

    “you operate in a combination of the fiat system in your country with a global offshore system.”