Latest / The CTO Podcast with Fexingo: Technical Leadership, Architecture, and Engineering Org / How Nubank Cut Cloud Costs While Scaling to 80 Million Customers
Transcript
- Lucas: So when Nubank hit eighty million customers earlier this year, they also quietly published something that most fast-growing fintechs would rather bury: their cost per active customer. And it's roughly half of what a traditional bank spends per customer on technology. Luna: Half. That's a huge number when you're moving money for tens of millions of people. How did they get there without starving the engineering org? Lucas: A big part of it is that they treat cost optimization like a product feature, not a finance exercise. They have a dedicated team, but the real lever is cultural. Every engineer at Nubank sees the cost of every API call they make. There's a dashboard, it's part of the pull request review. Luna: That sounds like a recipe for death by a thousand cuts. How do you keep engineers from spending all day shaving dollars off a Lambda function? Lucas: That's the tension, right? And Nubank's answer is that you don't make it the engineer's problem alone. You give them the tools and the data, but you also set a budget at the team level. Each squad has a cloud budget, and they're expected to stay within it. It's not punitive, but it does force trade-offs into the open. Lucas: And they pair that with a 'build vs. buy' framework that's aggressively biased toward building. They built their own core banking system from scratch, their own card processing, their own data platform. They argue that in the long run, owning your stack gives you more control over cost and performance. Luna: But building everything is expensive upfront. Didn't that almost kill them in the early days? Lucas: It did, and that's the real story. In 2018, they were burning through cash and their cloud bill was growing faster than revenue. They had to make a choice: keep building and hope investors kept funding them, or start optimizing aggressively. They chose both, but the optimization had to be smart. Lucas: So they started measuring everything. They introduced a metric called 'cost per transaction' and tied it to engineering OKRs. They also did something that sounds counterintuitive: they moved to a multi-cloud strategy, which usually increases complexity. But for them, it created negotiating leverage and let them pick the cheapest option for each workload. Luna: Multi-cloud as a cost lever. That's a contrarian take. Most folks I talk to say multi-cloud is a myth, that nobody actually runs it in production. Lucas: It's rare, but Nubank does it for specific services. Their core banking runs on a private cloud they built on top of open source. Their analytics and machine learning run on one public cloud, and they've kept some legacy workloads on another. They're not shy about the fact that it's a mess to manage, but they've built internal tooling to abstract it away. Luna: So the tooling is the secret sauce. That's a lot of engineering investment just to save on cloud bills. Where's the return? Lucas: They published that their cost per active customer dropped twenty-five percent year over year for three straight years. That's not a one-time optimization. It's a structural advantage. And they've been able to pass that on to customers with lower fees, which keeps them competitive against the incumbents. Lucas: One concrete example: they moved their real-time messaging infrastructure from a managed service to a self-hosted Kafka cluster. They claim it cut the cost of that system by seventy percent. But they also had to hire three engineers just to run it. Luna: So it's a bet that the engineering salaries are cheaper than the managed service. That's a classic build vs. buy calculation, but it only works if you have the talent and the scale to justify it. Lucas: Exactly. And Nubank has that scale, eighty million customers. They're processing billions of transactions a year, so a system that needs three engineers to run still pays for itself. But for a startup at a million customers, that would be a terrible trade. Luna: That's the kind of nuance that gets lost when people talk about 'cloud cost optimization' as a generic initiative. Lucas: Right. And it's also why they've been able to grow so fast without blowing up their expense ratio. Their operating costs as a percentage of revenue have actually gone down over the past two years, which is almost unheard of for a fintech in hyper-growth mode. Luna: So what's the one thing a CTO at a smaller fintech should steal from Nubank? Not the multi-cloud mess, probably not building your own core banking system. Lucas: I'd say the cost per unit metric. Pick the unit that matters for your business — it could be cost per active user, cost per transaction, cost per API call — and measure it relentlessly. Then make it visible to every engineer. You can't optimize what you can't see. Luna: And you have to be willing to have those uncomfortable conversations about trade-offs, like 'do we really need to log everything?' Lucas: Yeah, and that's where the culture piece comes in. Nubank engineers are known to push back on new features if the cost is too high. There's a story of a product manager who wanted to add a real-time notification for every transaction, and the engineering lead said, 'that would double our message queue costs, let's find a cheaper way.' They ended up batching the notifications and still got the feature out. Luna: That's a great example of cost awareness leading to better product decisions. And it's a far cry from the 'move fast and break things' mentality that still dominates a lot of tech. Lucas: It is, and I think that's the bigger lesson. Nubank has managed to maintain speed while being cost-disciplined, and that's rare. They've got a strong engineering culture that rewards efficiency, not just feature velocity. Lucas: And you know, this is a topic that's close to us here at Fexingo. Running a podcast network with three hundred shows, we're constantly looking at our own cost per episode, our cost per listener. It's not a bank, but the same principle applies: measure, optimize, and don't be afraid to make trade-offs. Luna: And that's a good segue to something we've been meaning to mention. These episodes take a ton of time to research and produce, and they're only possible because of listeners like you who chip in a few dollars a month. Lucas: It genuinely makes a difference. Buy me a coffee dot com slash fexingo — that's where you can support the network. Even a couple of dollars a month helps keep this show ad-free and lets us keep digging into stories like this one. Luna: Yeah, and we're not going to guilt-trip you. If you've gotten value from these episodes, consider it, and if not, that's fine too. Lucas: Alright, back to Nubank. So they've also been open about their failures, like the time they tried to migrate their database and hit a major outage in 2021. That was a costly mistake, but they learned from it and redesigned their migration approach. Luna: What did they change? Because that's a classic 'how we shot ourselves in the foot' story. Lucas: They moved from a big-bang migration to a strangler pattern, gradually shifting workloads over months instead of a weekend. They also invested heavily in automated rollbacks and chaos engineering. It's a reminder that even the best cost optimizers can get bitten by operational risk. Lucas: And that's maybe the final piece of their playbook: they accept that they'll make mistakes, but they design for fast recovery. That resilience is what lets them keep innovating without fear. Luna: So for a CTO listening, what's the single most actionable takeaway from Nubank's story? Is it the metric, the culture, or the multi-cloud strategy? Lucas: I'd say it's the coupling of cost awareness with engineering autonomy. You give engineers the data and the freedom to make cost trade-offs, and you hold them accountable with budgets. That combination is powerful, and it's what lets Nubank scale efficiently. Luna: And it doesn't hurt to have a CFO who understands technology. Nubank's CFO came from a tech background, which probably helps align incentives. Lucas: Absolutely. When the finance team and the engineering team speak the same language, you can have productive conversations about unit economics instead of just slashing budgets across the board. Lucas: So as we wrap up, I think the takeaway is this: Nubank proves that cost optimization and growth aren't mutually exclusive. It takes a deliberate strategy, a culture that values efficiency, and the willingness to make hard trade-offs. Luna: And it's a model that could work for any company, not just fintech. The principles are universal. Lucas: Right. Measure what matters, make it visible, and empower your teams to act on it. That's the Nubank way.