What Rails Can’t Solve
From the Fault Line No.2
Francesco worked on neobankbeat to watch neobanks silently for a year then published The Rise and Quiet Death of the Neobank. H/t for the valuable resource.
This my reply to his “Three predictions I’m willing to be wrong about.”
The first AI-underwriting blow-up happens within two credit cycles. Most of those 67 production models have never been through a real downturn in their current form. Some are about to learn what their training data didn’t contain.
I expected to find Catena Labs and similar bleeding edge AI x fintech startups in neobankbeat. Startups who want to become banks in a single bound are not neobanks. Noted.
For now, AI for neobanks is predominantly about underwriting.
the AI leaders mostly aren’t the famous names. They’re emerging-market lenders in Nigeria, the Philippines, Mexico, Bangladesh where credit bureaus are useless and a model underwriting thin-file borrowers isn’t a feature, it’s the entire reason the business can exist. The West talks about AI banking. The Global South ships it, because it has to.
It’s one thing for individuals in your company to be tokenmaxxing, another when no usable credit bureau exists and you have no choice but to train a model on thin, short-term, unique customer data. The thinness is why the business exists and why it breaks—something more proprietary data doesn’t fix.
The Thinking Machines x Bridgewater PoC, where the two companies finetuned Alibaba’s open weight Qwen3 on Bridgewater’s proprietary financial data and expert labeled examples, points at the shape of a resilient model: a customized model plus the credit officer who’s seen a dozen downturns.
The obvious reason to finetune an open weight model compared to a frontier model is cost, roughly 13.8x less per inference task. The nonobvious one: it’s cheap enough to do repeatedly.
You don’t have to ship-and-pray, but retrain as conditions shift, reinjecting human judgment each time.
The credit officer’s job is not to label the loans. She invents the credit SNAFUs you haven’t seen yet. What the Great Recession would do to this book, when it is not in the data.
Judgment scales badly. Moats usually do.
The license gap closes from both ends. The strong unlicensed players buy or earn charters; the weak ones become the 2027 deletions. The middle disappears.
Kulipa, a B2B neobank infra platform that powered card programs, shut down and deactivated all crypto debit cards tied to its partner neobanks and wallets. Textbook prediction 1.
Charters as a moat are overrated. Because regulators ration approvals, they’re hard to get. But once the strong players get their own charters and the weak ones are gone, charters become a pricey commodity among the survivors. A defensive necessity, not a weapon.
Every neobank in that middle was somebody’s customer.
The battle moves from who holds the licenses to who holds the rails.
Meanwhile, stablecoins are the frontline.
Nobody’s figured out who eats the loss when a stablecoin breaks. An uninsured instrument now competes with an insured one, at volume, with no resolution mechanism. And nobody knows how anyone clears and nets at scale.
We’re using them, yet these systems can break at any moment.
The customer of the next wave isn’t human. Banking rails for AI agents wallets agents operate, cards agents issue, machine-to-machine payments is seven companies today. It looks exactly like web3-native looked in 2021: tiny, weird, and structural.
Agentic banking can be relatively boring or very weird. An agent acting for a verified human, within limits that humans set, sounds too ordinary while AI agents as principals push into sci-fi. The interesting problem in agentic payments, like stablecoins, is liability, not rails.
I’m still forming my views on agentic banking. I’ll come back to this when I’m ready.
If you’re building solutions for loss allocation or netting that make stablecoin systems more resilient, let’s talk.
From LAVA
Ryan, The Protean in his finest VC stance giving out VC bread-and-butter—
Ryan on the 3 questions he grills stablecoin neobanks with and how not to answer.
cryptowanderer concluded his 3 part series:
how local market makers can build deep, liquid emerging-market stablecoins and our open source software as a starting point.
our experience and conclusion of providing liquidity
for cNGN across networks
See Eun on how banks could prefer CIPS and RMB between Africa <> China
No Comment
Parimutuel Market Maker Litepaper for permissionless prediction markets. Pitch to an AI-twin VC. Bottom-up research on AI adoption across African startups. Exits for Africa. Busha’s one step towards an everything exchange. Know what’s been built onchain. Merchant banking stands out from banking. Hide yo books, AI is coming. Invisibility as an edge. Guo chronicles. Trade-as-an-African-curse. ELI5 $FWA.
What’s Next
I last published From the Fault Line 3 months ago. Going forward, instead of writing commentary on a handful of links shared during the week à la Matt Levine, I will write one commentary focused around the juiciest link that week.
I leave you with a reminder for work humans must own regardless of technological progress.
