# Ten Thousand People Serving 140 Million: David Vélez on Banking's Second Half

Source: Realpha Blog (blog.getrealpha.com)
Original article and charts: https://blog.getrealpha.com/en/blog/iltb-2026-10-13-david-v-lez-building-nubank-and-the-future-of-glob/

> Notes after listening to Invest Like the Best EP.495. David Vélez's numbers: global consumer finance is a $10T industry with ~90% still flowing through branches; Nubank serves 140M customers with 10,000 employees at roughly 3% of an incumbent's cost to serve. Educational discussion of his reasoning and where my own read was wrong — not investment advice, no stock recommendations or price targets.

Published: 2026-10-13
Locale: en
Tags: digital banking, financial inclusion, cost structure, AI and org design, podcast notes

TL;DR: He opened up the gap with headcount and branch counts, which convinced me digital banking is still early; I still haven't seen that advantage travel outside Latin America.


![An old bank hall at dusk, light pouring in from the far glass doors, empty queue barriers running into the depth of the room, a man standing midway looking down at his phone](/covers/iltb-2026-10-13-david-v-lez-building-nubank-and-the-future-of-glob-cover.png)

> The way to govern a state begins with making the people prosperous. Prosperous people are easy to govern; poor people are hard to govern.
> —— *Guanzi, "Governing the State"* (Warring States period, translation mine)

On 13 October 2026, Invest Like the Best EP.495 had Patrick O'Shaughnessy interviewing Nubank founder David Vélez. The framing numbers he gave: global consumer financial services is roughly a $10 trillion industry, about 90% of it still runs through bank branches, and every neobank and fintech combined holds under 10%. Nubank serves 140 million customers with 10,000 employees, which puts its cost to serve a customer at about 3% of a traditional bank's. All of those figures come from him on the show, with no third-party check, and the cost advantage still has to travel outside Latin America to count — his own bear case is a failed internationalization with three saturated countries.

## I Thought What Was Left Was Share-Grabbing

He was on the same show five years ago with 35 million customers. Now it's 140 million. My first reaction to that curve was that it should flatten: neobanks are a mature category by now, every market has a few, and from here it's subsidies, cross-selling, and poaching each other's customers. I filed it under redistribution of an existing pool.

I had also picked the wrong side on the AI layer. With the cost of writing software collapsing, my inference was that scale and data would get cheaper — a team of three wires up a model and rebuilds a decade-old product in months, so the entrant wins. On that logic, a company as large as Nubank becomes the next thing to get flipped.

## Ten Thousand Against a Hundred Thousand Plus 6,000 Branches

The way he takes this apart left me without an argument. He compares global banking to the years when Blockbuster was still open: 90% of people walking into a store for VHS tapes and DVDs, Netflix-type services holding about a tenth. I'd heard the analogy before; what was new was him putting the denominator on the table — every digital challenger added together is under 10%, so the migration is at its beginning.

![Two bars of equal length are stacked and aligned, each mostly grey with a short blue segment at the right end; the top bar is the video-rental market in the Blockbuster era and the bottom one is consumer finance today, and in both the blue takes up only about a tenth.](/figures/migration-still-ninety-percent-en.svg)

The cost section is the hardest part of the episode. His competitors serve 100 million customers with 100,000 employees and 6,000 branches; he serves 140 million with 10,000, which lands his cost to serve at roughly 3% of theirs — and that gap predates AI. Where does the saving go? His answer is zero fees and yield handed to the customer. That reasoning runs straight into his financials: he charges less than the incumbents and still earns two to three times their return on equity, which he says is what a cost-structure advantage looks like.

![Two groups face each other: on the left a very tall headcount column beside a row of branch blocks, with a shorter customer bar below; on the right the headcount column is a mere stub, the branches are empty outlines, and the customer bar is longer instead.](/figures/ten-thousand-versus-one-hundred-thousand-en.svg)

He sorts the problem into three buckets: access, cost, intelligence. Access is the 3 billion people worldwide with no credit product at all, and credit card penetration of 13% in countries like Mexico and Colombia. On cost he uses the US: about $15 trillion of retail deposits earning an average 40 basis points, which in a 4.5–5% rate world leaves roughly $600 billion a year sitting with banks to fund physical infrastructure, on top of 5–10% remittance fees and $5–15 monthly account fees. Intelligence is where he's sharpest: the top 1% of earners have a private banker optimizing accounts, structuring for tax, shopping a loan across several banks, while the other 99% get nothing — and solving that through financial education has never worked, because nobody sits down for the classes. What they do in Brazil today is tell 15 million customers daily: your card at that other bank is due tomorrow, there's no money in it, the overdue rate there is 15% a day, let us lend you the payment. Or, that loan of yours at 10%, let us refinance it at 5%.

So what about the entrants I'd bet on? He puts up two gates. The first is licensing: you can have a model rewrite an entire software stack, and you still can't get it to produce a banking license or compress a two-to-three-year approval into six months. The second is data: 140 million customers, 20,000–30,000 data points each, spanning 13 years. They used that to build their own foundation model for financial services, running credit and collections decisions on it, and published the research. A startup has speed and can't manufacture that depth.

![One hundred small squares form a ten-by-ten grid, with only the top-left square in blue and the other ninety-nine in grey, noted alongside as the one square whose person has someone handling things for them.](/figures/one-percent-has-a-banker-en.svg)

He doesn't stop at "we have more data," though. Asked how he beat banks with a century of history, his answer is that Blockbuster had mountains of data too — who walked in, which tape they stopped in front of, what they rented. None of it produced a better recommendation engine, because they weren't capturing it, had no systems to hold it, and had no culture of deciding from it, so Netflix started with zero data and built the better algorithm anyway. Scale counts once systems and culture sit under it, which is him putting an insurance policy on the moat he'd just described.

My favourite stretch of the episode is the taco stand on a Mexico City corner. He sells all day, takes cash, and none of those transactions land in any ledger, so he can't borrow. The moment the transaction goes digital, he can walk his own cash flow to a bank and ask for credit — and one stand becomes two, then five. The same logic is why they issued $20 credit lines. Banks thought it was absurd; to the customer, $20 beat zero, and $20 became $100, then $1,000. He says they were the first credit card for 30 million people in Latin America.

![Two rows of multiplication, each with three boxes standing for data, systems and culture as either full or empty, followed by an equals sign and a bar whose length is the result; the top row has data full but the other two empty and a result near zero, while the bottom row has only a sliver of data but the other two full and the longest result.](/figures/data-times-system-times-culture-en.svg)

The organizational part was the most concrete. They write a lot of Clojure, and through all of 2025 every model they tested was useless at it — their most experienced engineer concluded the path was closed. The turn was Opus 4.6 in December 2025, and they found out 60 days late: a few partners went to California, sat with early-stage founders, realized something had shifted, retested, and confirmed the models now wrote Clojure well. Those 60 days, in his view, were distance opening between them and best-in-class. Back in Brazil they went code red: 120 workflows mapped and rebuilt AI-natively, from product design and KYC through internal audit, compliance and regulatory capital management, with eval infrastructure built alongside so they could tell whether each step actually improved. Ten layers compressed to five; he went from one direct report to twenty. His reasoning is that a message crossing ten layers gets ten chances to be distorted.

## When the Shark Leaves the Water

My reading order has changed: first look for a physical gap an outsider can count, then listen to the narrative. Headcount, branch count, cost to serve, return on equity — those are countable from outside. "We have a more innovative culture" isn't. What convinced me here was 10,000 against 100,000, a checkable pair of numbers, rather than how well he tells it.

He marked his own ceiling. He calls Nubank a shark in digital waters, and says that when it's forced to compete offline against branches and ATMs it gets visibly weaker — which is what the first couple of years in Mexico looked like, in a country used to cash. That line explains the advantage and the limit at once: the US is still a branch-and-brand market, and he says building the brand there is the hardest thing they'll do, without carpet-bombing marketing to get it.

![A horizontal waterline splits the frame in two: below it on the left is a large shark-fin triangle, and above it on the right the same triangle appears shrunken and tilted on its side among a row of bank-branch and ATM blocks.](/figures/shark-out-of-water-en.svg)

Three things I didn't hear answered. How much of the data moat survives a border — can 13 years of Brazilian behaviour underwrite an American? His claim is they arrive in a new market with the foundation model on day one; the validation wasn't given. Second, the two-to-three-times return on equity carries emerging-market spread structure inside it and has to be recomputed for the US, which he half-concedes when he notes some US incumbents now out-earn emerging-market banks. Third, every number in the episode is self-reported.

If you've been stuck telling tailwind apart from skill in a company's growth, this episode hands you a usable cut: count the competitor's people and fixed assets, then look at how much more this company earns while charging less. A gap that lives in the cost structure stays in the filings; a gap that lives in the story is gone next quarter. His HR section is the same idea in another register — employees get feedback every six months, mostly loaded with impressions from the last 30 days, and he says coaching an Olympic athlete that way would be a disaster, because an athlete has someone correcting them daily. Things that can be checked should be checked often, and that rule reads the same inside an organization as it does in investing.

![Two bar groups sit side by side, fees charged to customers on the left and return on equity on the right; connecting each company's two bar tops draws one line rising and one line falling, and the two lines cross in the middle.](/figures/charge-less-earn-more-en.svg)

## One Thing to Take Away

That company was 60 days late to knowing the world had changed, with a room full of smart people in it, and their most senior engineer had signed off on "this path is closed." The difference was where they were standing — what dissolved the misread was a few people flying to California and sitting across from a group already using the new tools.

Here's something I've tried: take one thing you do daily — raising kids, running, bookkeeping, writing reports all count — find someone who does the same thing a different way, and ask them which step they've dropped lately, and why. Don't ask about tools; they'll volunteer those. The step they threw out is where they paid tuition. Twice now, asking that has shown me a habit I was guarding for no remaining reason.