ChatGPT can rewrite your product, but it can't get a banking license
Nubank's founder argues this is precisely why scaled players capture most AI gains—they've accumulated a decade-plus of data and hold licenses; AI amplifies their edge rather than leveling the playing field.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Ten thousand staff scale to what takes legacy competitors one hundred thousand and six thousand branches
Nubank operates with roughly ten thousand employees serving 140 million customers, while traditional competitors deploy one hundred thousand employees and six thousand branches to serve 100 million at roughly three percent the service cost of legacy banking. Vélez attributes this to the first phase of efficiency gains—mobile internet freed finance from physical branch networks, a result that predates AI. AI becomes the next efficiency multiplier on this foundation, enabling banks to serve more customers with fewer people or redirect these gains into aggressive global expansion.
— David VélezThe interest-rate spread is a hidden wealth tax on depositors
Vélez ran the numbers: the US banking system holds roughly $15 trillion in retail deposits, banks pay an average of 40 basis points in interest, yet market rates sat between 4.5 and 5 percent, leaving a spread of around $600 billion annually—money that should stay in consumer pockets instead maintains bank branch networks. He calls this a hidden wealth tax, with lower-income customers losing a higher share because high-income depositors at least capture some interest income.
— David VélezSpeed has a ceiling: ChatGPT doesn't get you a banking license
If someone were to build a company from zero to disrupt Nubank, Vélez sees one advantage: speed. Unconstrained by regulation, they could iterate rapidly. But that advantage has a ceiling. Finance ultimately requires a license, and ChatGPT cannot secure a banking charter for you any more than it can compress regulatory approval from two or three years into six months. Nubank already holds licenses across multiple jurisdictions, has accumulated fifteen years of data on 140 million customers, acquired Palo Alto-based Hyperplane, and built proprietary foundation models for financial services—occupying the sweet spot where scale, data, licenses, and AI capabilities converge.
— David VélezBanks have decades of data and still can't build algorithms with it
Vélez counters the "banks hold massive data so they're safe" argument with Blockbuster: it recorded vast customer rental patterns daily but never built better recommendation algorithms because it lacked systems to collect data and—more fundamentally—a culture of data-driven decisions. Netflix, with no historical data, beat it through algorithms. Traditional banks today still run dozens of systems left by acquisitions, migrating mainframe workloads to the cloud, and this organizational and cultural lag is harder to overcome than data volume itself.
— David VélezNinety percent of today's currencies will lose relevance in ten to twenty years
Vélez predicts that as more of life straddles home-currency and offshore USD stablecoin systems, today's roughly 190 fiat currencies will shrink to perhaps ten with any meaning within ten to twenty years—many nations' fiscal positions are so weak their currencies have essentially become crypto assets. The twist: this actually extends the dollar's life as reserve currency. Stablecoins make it easier globally to hold dollars, but dollars won't be the only option—five or six national currencies plus several crypto assets will share the role of store of value.
— David VélezModel capability flipped overnight; the company noticed sixty days later
Nubank extensively uses Clojure, and throughout 2025 no LLM generated Clojure well—even the company's most senior engineers thought it impossible. Then a new model generation arrived (Vélez mentioned Opus 4.6 in December 2025) and the capability inverted overnight. Yet Nubank took a full sixty days to notice the inflection—a few partners learned about it in conversation with early-stage founders in California and realized something fundamental had shifted. Back in Brazil, the company executed a cold restart across the entire organization, inventoried 120 workflows, and rebuilt each one AI-native.
— David VélezDirect reports grew twentyfold as management layers collapsed from ten to five
After this AI-driven organizational overhaul, Vélez's own direct reports expanded from one to twenty, and management layers compressed from ten to five. The rationale extends beyond efficiency: more hierarchy means more distortion as information travels upward, while AI systems let him see ground truth directly without relay chains. The company built an internal "corporate brain" that proactively pushes information about AI transitions, product launches, and KPI progress rather than waiting for quarterly reviews.
— David VélezFounder and CEO are opposite temperaments; in a pivot, choose founder mode
Though founder and CEO are often the same person's two titles, they're opposite temperaments: founders lack patience and crave speed, willing to charge forward recklessly; CEOs demand patience and systematic decision-making, excelling at building management systems. As companies scale, people default toward CEO mode, but Vélez's conclusion is to choose founder mode when forced to pick one—especially during a technological pivot like AI, where founder instinct and urgency beat CEO process and patience.
— David VélezIn their own words · checked verbatim
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 and is owned by incumbent banks.
David Vélez2:03
we have 140 million customers with 10,000 employees or competitors have 100 million customers with 100,000 employees in 6,000 branches.
David Vélez5:04
there's about $15 trillion of retail deposits in the U.S. banking system. But banks pay on average about 40 basis points on yield a year.
David Vélez8:19
you still cannot really get ChatGPT to get your banking license.
David Vélez14:37
My view is that 10, 20 years from now, those 190 will be maybe 10 because the fiscal situation in so many countries around the world is so dire that a lot of that fiat money is not real money.
David Vélez25:54
holy shit, this is actually working out. It can do Clojure very, very well.
David Vélez45:28
The founder is impatient, huge amount of sense of urgency. The founder will want to go run very, very fast, will move fast, extremely fast. The CEO will need to be more patient.
David Vélez58:58
I don't think anybody in the world cares as much for this company as I do, right? To a level that is almost miserable.
David Vélez1:08:09
Figures
| Current customer base | 140 million (35 million five years ago) | 1:03 |
| Brazil market penetration | 120 million Brazilian customers, roughly 65% of adult population | 3:03 |
| Global population without access to credit | 3 billion people | 6:11 |
| US retail deposits and interest | $15 trillion in deposits, average annual rate ~40 basis points | 8:19 |
| Annual interest-spread loss to depositors | ~$600 billion | 8:19 |
| Predicted surviving fiat currencies in 10–20 years | From ~190 to ~10 | 25:54 |
| Organizational restructuring: direct reports and management layers | Direct reports increased from 1 to 20; management layers compressed from 10 to 5 | 46:33 |
Glossary
- foundation model
- A large-scale pretrained general-purpose model; here, Nubank's proprietary model specialized for credit risk decisions.
- open banking
- Regulation allowing consumers to export their financial data across institutions and authorize third parties to use it.
- stablecoin
- A cryptoasset pegged to fiat currency like USD; here, a product form enabling cross-border users to hold dollars easily.
How to listen
Investors and founders tracking fintech regulatory moats, enterprise AI reorganization, and founder-led management philosophy.
Sponsor ads at ~0:00–1:03 and ~53:48 can be skipped.