Happy Friday, Fintech Takers!

I really appreciate the flood of San Diego recommendations I received after publishing Wednesday’s newsletter. Rest assured, Legoland is at the very top of our to-do list!

If you have additional recommendations, please send them my way!

- Alex

Was this email forwarded to you?

Subscribe Here

Fintech Threads: Stablecoins and Model Distillation

One of my favorite things about writing this newsletter is the feedback loop that has formed around it — I share my takes and then, almost immediately, the email replies start to roll in.

The replies span the gamut. Some agree with my takes. Some (politely) disagree. Many ask follow-up questions that help to push my thinking on the topic(s) forward. A few turn into substantive back-and-forths over weeks or even months.

In all cases, I’m extremely grateful because these replies (which I always read and almost always reply to) help me get smarter.

In Monday’s newsletter, I wrote about Zelle’s expansion into international remittances and the upcoming launch of its ZLUSD stablecoin, as well as Klarna being sued by its former partner Pagaya. Both of these write-ups prompted some thoughtful email replies from subscribers and exposed a few loose threads on those news stories that I would like to pull on a bit more on.

So, that’s what we’re going to do!

USD-backed Stablecoins & Payments Sovereignty

The headline here is that Early Warning Services (EWS), a financial infrastructure company owned by a consortium of large banks, is planning to expand its Zelle P2P payments network into other countries, starting with India. It is also planning to launch a Zelle stablecoin (ZLUSD), which will, eventually, be one of the underlying payment rails that Zelle runs on.

My take on this news in Monday’s newsletter was that EWS’s openness to launching a stablecoin is a signal that the big U.S. banks (which own EWS) are less worried about crypto deposit displacement for consumer use cases like remittances, compared to commercial use cases like treasury management for multinational corporates, which their other jointly-owned payments company (The Clearing House) is building a tokenized deposit network to support.

I still believe in this take, but there’s a different thread in this story that I want to pull on: The threat of USD-backed stablecoins to other countries’ national sovereignty.

When we talk about stablecoins eating into payments, we tend to frame it as a fintech disruption story. The less sophisticated version of this story is that stablecoins will disrupt the card networks (“that 2-3% interchange fee is a scam!”) and lower costs for merchants. The more sophisticated version of this story is that stablecoins will be absorbed into existing global payment networks as a superior settlement mechanism for cross-border transactions.

However, the part of the stablecoin payment story that we’re not thinking about enough is the threat that USD-backed stablecoins pose to foreign central banks. That might sound overly grandiose (“are central bankers in other countries really worried about ZLUSD?”) but it’s important to remember that payment systems are one of the best places to exercise soft power, because every sanction, every AML regime, every "you can't do business with that person/company/country" rule ultimately has to be enforced somewhere in the flow of funds.

If you own the corridor, you own the off-switch.

The U.S. has obviously understood this for a long time. Dollar clearing and the correspondent banking system have been very potent foreign policy instruments for decades. They represent, functionally, the ability to export your values (and your enemies list) to the rest of the world.

When we talk about U.S. dollar-backed stablecoins in this context, the discussion usually focuses on countries like Argentina, where currency instability and hyperinflation make the U.S. dollar a highly desired store of value for consumers and business owners. This stablecoin use case is highly frustrating to the central bankers in those countries, but it’s also highly logical from the perspective of the end users, and thus very difficult to prevent. Consumers and business owners in these countries are going to find a way to hold U.S. dollars, one way or the other.

The more pernicious national sovereignty threat posed by USD-backed stablecoins is in countries that have stable currencies and manageable levels of inflation. Consumers and businesses in these countries are not demanding U.S. dollars as a store of value. However, they may still end up transacting in (and even holding) U.S. dollars if the payment tools that they utilize run on top of USD-backed stablecoin rails.

Payments looks neutral, but smart central bankers know that it’s not. It's a Trojan horse, and the soldiers inside are denominated in dollars.

Payments solutions like ZLUSD or PYUSD (PayPal’s stablecoin) don’t have to require users to hold their money in USD-backed stablecoins. They can simply make it the most convenient option (by driving last-mile merchant adoption so that funds never need to be converted into the local currency) or the most rewarding option (by layering on yield or other financial incentives), and suddenly holding U.S. dollars becomes a default behavior.

Central bankers in countries with unstable currencies won’t be able to stop people from wanting to store their money in dollars, but central bankers everywhere else absolutely can fight to keep dollar-based stablecoins from becoming the payment rails that everyday transactions run on. National payments infrastructure is a defensible front line, which is exactly why the countries with the most to lose are already digging trenches.

India and Brazil offer two instructive examples.

India's UPI now processes north of 20 billion transactions a month and has gone international, with live or in-progress links across nine-plus countries (the UAE, Singapore, Nepal, Mauritius, and counting) and a stated ambition to reach a dozen-plus. More importantly, India's central bank joined Project Nexus, which is the Bank for International Settlements' effort to interlink domestic real-time payment systems across borders into a single network. The entire point of Nexus is to let money move between countries instantly without having to route through dollar infrastructure.

Brazil's Pix instant payments system is run by the central bank of Brazil. Pix moved nearly $7 trillion last year and continues to grow quickly. Brazilians love it. And in March 2026, the central bank switched on Pix's first international corridor in Argentina, with Portugal, the U.S., and other Latin American countries reportedly on deck.

That may look like a payments product expanding, but I think a more accurate read is that it’s a sovereign payment rail going on offense.

And the U.S. government has taken notice. In July of last year, the Office of the U.S. Trade Representative (USTR) opened an inquiry into Pix, alleging that it imposes unfair competition on U.S. credit card operators because it offers an alternative to transaction fees. Following the investigation, the USTR determined that certain Brazilian practices are unreasonable and proposed retaliatory tariffs of up to 25% on various Brazilian imports.

On the surface, this looks like the Trump Administration going to bat for Visa and Mastercard, which have been losing volume to Pix and are (I’m guessing) terrified of its expansion into other countries. There’s definitely some truth to that, but I also think this fight is about more than interchange revenue. It's about who gets to control the rails, and, downstream of that, whether dollar-based infrastructure can become an important part of one of the largest economies in the Western Hemisphere.

So when Early Warning Services takes Zelle to India with a dollar-backed stablecoin in tow, the remittance use case is the part everyone will talk about. The more interesting point is that every new corridor for a USD-backed rail is a small expansion of the dollar's footprint, and the central banks on the receiving end know it. 

Model Distillation

On Monday, I wrote about Pagaya's lawsuit against Klarna, which boils down to an accusation that Klarna partnered with Pagaya to support the growth of its U.S. POS lending business, and then spent four years watching Pagaya's underwriting model make decisions — the approvals, the declines, the pricing, the repayment performance — and used those observations to backwards-engineer a competing model of its own. I was skeptical that Pagaya can win (and I still am), but I ended that write-up with a question I haven't been able to stop thinking about: Where's the line between illegally using your partner's trade secret and just getting smarter by working next to them for four years?

As far as I can tell, this is the first time a model-distillation-as-IP-theft claim has surfaced in fintech or consumer finance. (If you know of an earlier one, hit reply … I'd love to be corrected.) However, a broader and much more consequential fight over distillation has been happening for a while now in the most consequential industry in the world: AI.

When OpenAI trains a new version of GPT, what is it actually doing?

Essentially, it’s studying an enormous volume of human output (everything we've ever written down) and distilling that into a model that can reproduce the patterns contained within. The collected works of any great author are, in a sense, that author's proprietary model. And the AI labs have distilled it and reproduced their own version of it, without the author’s permission.

[Channeling my colleague Kiah Haslett]: IS THAT A CRIME?!?

So far, the courts are saying, “no.”

Authors and publishers have sued over AI training, and courts have mostly leaned toward the view that training a model on copyrighted material is transformative enough to qualify as fair use. (The New York Times' case against OpenAI is still grinding through discovery and hasn't been decided on the merits, but the broader "training is fair use" position has picked up real wins in adjacent cases.)

Distilling a human's life's work into a model? “Largely fine,” says the law.

However, what might happen when someone tries to do the same thing to OpenAI? How would OpenAI respond?

Well, we already know the answer because it has (allegedly) happened!

OpenAI has accused DeepSeek of "distilling" ChatGPT — querying its API at scale to train a cheaper competing model in violation of OpenAI's terms of service — and the White House's former AI czar called it outright IP theft. Anthropic has leveled similar accusations at DeepSeek, Moonshot, and MiniMax for distilling Claude.

It’s basically the same mechanism — observe the outputs, learn from them, build something new — but suddenly it's a crime.

There is a rich irony here. OpenAI distilling humanity’s collective intelligence is fair use, but someone distilling the resulting OpenAI model is theft. (Anthropic, to its credit, at least acknowledged that distillation is a normal industry practice even while objecting to it being done to Claude.)

But the irony isn't the interesting part. The interesting part is the legal machinery underneath it.

Because notice what changed. When it's a human (or a model) learning from copyrighted human work, we reach for copyright law, and copyright law has spent centuries building pressure release valves for exactly this kind of knowledge transfer. Fair use. Transformative use. The right to make derivative works. The whole doctrine is designed around the premise that learning from, remixing, and building on top of existing work is something we want to permit or even encourage.

But the moment it's model-to-model, the doctrine suddenly changes from copyright to trade secrets. And trade secret law has none of those pressure release valves. It's binary: either the thing is a protected secret that was improperly acquired, or it isn't. There's no "your competing model is transformative enough" defense. There's no fair-use carve-out for getting smarter by watching.

So we've built two completely different legal regimes for what is, mechanically, the same act. Copy a human's work into a model: copyright law applies and it’s probably fair use. Copy a model's behavior into another model: trade secret law applies, and it’s probably misappropriation.

The only thing that changed is whether the thing that is doing the learning is carbon-based or silicon-based.

And here's the part that should worry anyone who thinks this is a containable problem.

In Klarna/Pagaya, the alleged distillation required privileged access. Klarna only got to observe Pagaya's outputs in that kind of granular detail because the two companies had a four-year commercial relationship. That access is the bottleneck. It’s also the thing that gives Pagaya something to point at in court. The secret was protected by a relationship, and the relationship was (allegedly) abused.

However, that bottleneck is the result of current technology constraints. It’s not a fundamental law of nature. Increasingly, you don't need the privileged relationship in order to learn.

A sufficiently capable AI system wouldn't need to sit next to Pagaya for four years. It could start with the public record — the securitization disclosures, ABS performance data, and investor reports that Pagaya itself has to publish in order to sell its loans — and distill a "close enough" model from that. Or it could go further and probe the system directly: deploy agents to actually apply for loans, at scale, mapping out exactly where the approve/decline line sits and how pricing moves. Mystery shopping, except automated, relentless, and running thousands of experiments a day.

To be clear, this isn’t happening today (I don’t think!), but it's not science fiction either. It's a straightforward extension of agentic AI applied to a problem that, until now, required humans.

And it breaks trade secret law completely. That body of law assumes knowledge transfer happens at human scale, through identifiable people who had access that they shouldn't have had or used their legitimate access in ways that they shouldn’t have. It assumes you can protect a secret by controlling who's in the room and assigning contractual obligations that they are required to adhere to.

But if a model can reconstruct, from public outputs and active probing, what used to require privileged human access, then the "secret" was never really secret. It was just expensive to reverse-engineer. And once reverse-engineering gets cheap, the legal concept of trade secrets starts to dissolve.

It also scrambles the question of who is even responsible. When a human distills a competitor, there's a paper trail and a bad actor to point at. When a self-improving model does it autonomously, the knowledge transfer becomes both undetectable and unpreventable. “Who misappropriated the trade secret?" stops having a clean answer. The company? The model? Nobody?

So, yes, the Klarna/Pagaya case isn’t, by itself, a huge deal (and I'd still bet against Pagaya winning in court unless discovery turns up a bombshell), but it's a preview. Very soon we are going to have to decide, collectively, what counts as legitimate distillation and what counts as theft. And we're going to have to do it in a world where the distiller increasingly isn't a person, doesn't need permission, and can't be caught in the act.

Sponsored by T-Mobile for Business

The financial services sector rarely benefits from one-size-fits-all solutions.

Legacy infrastructure, regulatory complexity, and the uneven pace of digitization across institutions all require something different.

That’s why many of the world’s largest financial institutions choose T-Mobile for Business and their 5G Advanced Network Systems.

With expertise in regulatory requirements, they provide a comprehensive tech ecosystem for businesses in financial services.

Whether through public, private, or hybrid mobile networks, T-Mobile configures deployments so institutions invest only in what they truly need.

See what that looks like for yours

MORE QUESTIONS TO PONDER TOGETHER

Big news for the endlessly curious (yes, you): I’m collecting your fintech questions on a rolling basis. 

What’s keeping you up at night? What great mysteries in financial services beg to be unraveled? Think of it this way, if a stranger is a friend you just haven't met yet, your question is a Fintech Takes conversation waiting to happen. 

One that could headline a Friday newsletter or be answered in an upcoming Fintech Office Hours event.

Drop your question here, whenever inspiration strikes!

WHERE I'LL BE

There are some great virtual events coming up soon, and planning for the fall fintech season is well underway. Here's where I'm planning to be this summer (virtually) and in September (in person).

💻 Fintech Office Hours | June 25 | Virtual

Fintech Office Hours are now open to everyone! Come hang out with us, hear about what's new in fintech (and what I'm cooking up for the newsletter), and ask questions!

💻 Stablecoins as Payments Infrastructure (A Conversation for Skeptics) | June 30 | Virtual

Western Union is leaning all the way in on stablecoins, and the more you learn about what its doing, the more it makes sense. We'll be talking about it in this virtual event, sponsored by Rain.

✈️ FinovateFall | September 9-11 | New York City

My can't miss fall conference! September in New York is glorious and the fintech conversations will be too.

✈️ FDATA Global Open Finance Summit | September 17 | Toronto

This will be my first time at an FDATA event and my first time back to Toronto in a long time. If you work in open banking in Canada and want to yell at me for my bad takes in the past, this is your chance!

✈️ AI-Native Banking & Fintech Conference | September 29 | Salt Lake City

The name of this event is a mouthful, but the content and networking are both A+.

✈️ *Bonus: Cash Flow Intelligence Summit 2026 | September 10 | New York City

I've been watching this event grow for three years, and the brainpower it draws is impressive. Cash flow intelligence is now embedded across fraud, line management, and retention; and how AI is changing the underlying models is still very much an open question. Full day on where cash flow intelligence goes from here, invitation-only. Apply to attend.

*This rec is brought to you by one of our fantastic brand partners.

Thanks for the read! Let me know what you thought by replying back to this email.

— Alex