Margin notes on Spiegel.

Evan Spiegel sat down with Lenny on a podcast called How to win when software is not a moat, and I watched it on a Sunday evening with the bKash post I'd published the week before still open in another tab. Half of why that tab was open was procrastination on a separate draft. The other half was that Spiegel kept arriving at sentences I had already pencilled in a different accent. The talk is worth the fifty minutes; the embed is below.

What follows is five places where his sentence in Santa Monica met a paragraph I had already pencilled in Dhaka, plus one place where I want to push past where he stops. I'll cop to the obvious caveat upfront, which is that I'm a Dhaka product owner reading a Snapchat CEO's interview and pretending the lessons port across, but the unfair advantage of working in this market is that the moat question is never abstract here, because the infrastructure under the app is visible from the office window.

That sentence is the one Lenny puts at the top of the interview, and it's the one a generation of founders in this country quietly pay for every quarter without ever naming. The pattern is so common in the Banani decks I've reviewed over the last three years that I now look for the distribution slide before the product slide, and the few decks that have a real answer for it tend to belong to teams who built the distribution before they wrote the app. The clearest local example, the one I wrote about last week, is bKash, where the moat is not the USSD menu or the polished app but the roughly three hundred thousand shopkeepers across the country who keep a poster laminated pink on the wall and a float account topped up twice a week. The app, in any honest accounting, is the receipt for what the shopkeeper already did.

The Spiegel framing lands harder when you sit with the corollary, which is the part most founders skip. The question on the slide deck is not what will we build, but who will hand the thing we built to the next user in a setting that user already trusts. Hamilton Helmer makes this point at book length in 7 Powers, where the durable advantages he catalogues (scale economies, cornered resources, branding, switching costs) all sit downstream of distribution in a way the strategy slides in this town do not usually price in. Spiegel is essentially saying the AI wave has reset the field back to the same fight Helmer's framework already named. A lot of teams who thought they had a moat built from process power will discover by next Eid that the moat was actually rented.

I half-laughed at this one because the sentence is so obvious in hindsight that it sounds glib, and then I caught myself, because half my own essays are also half-arguing this same point under different headers and I have not always been generous about saying so out loud. The cleanest local proof is Pathao on the segment for two wheelers. Uber's ride share business is still operating in Bangladesh on the car side, but the dominant share of two wheeler trips runs through Pathao. The model Uber brought in from elsewhere assumed banked drivers with smartphones and credit cards, while the median bike rider in Dhaka is on a 125cc Honda, has never had a bank account, and gets paid in cash at the end of each day. Pathao captured that segment by building an onboarding flow for unbanked riders, an upfront cash payout, and a rider network you could pull a bike from in eleven minutes on Mohakhali flyover at 6pm. The Pathao app and the Uber app render almost identically on a phone screen, and the rider network underneath one of them has fifteen years of relationships with the upcountry families whose sons are now driving the bikes, while the other one had a Salesforce instance in Singapore. The same shape played out in food delivery the other direction, where Uber Eats exited Bangladesh while Pathao Food and Foodpanda kept the courier network they had spent years assembling.

This is also where I have to be honest about my own work, because SupplyLine, the thing I've been building on the side for the last two and a half years, started with the lending UX as the thesis and arrived at the distribution thesis only after the first eight months of trying to acquire merchants through a loop of paid app placement went predictably nowhere. The moat for SupplyLine today is the roughly four thousand retailers on the platform and the ten tons of FMCG product moving through the network on an average weekday, settled through a layer of credit underwriting that exists because the retailer relationship was already there. If you removed the lending product tomorrow, the network would still ship the ten tons. If you removed the network, the lending product would have nothing to underwrite. I had to learn this twice, the second time more expensively than the first, and Spiegel saying it in a Santa Monica office in two clean sentences is the kind of compression that makes a working PM in Dhaka feel both validated and mildly embarrassed.

The Spiegel reading of TikTok and Threads is worth dwelling on, because it cuts against the romantic version of the story that gets told at startup events in Gulshan, where TikTok is presented as a triumph of content recommendation and Threads as a clever product pivot. Spiegel is unromantic about it in a way I appreciate. TikTok figured out distribution by spending billions of dollars subsidising both sides of a video marketplace, paying creators to make videos and paying for users to come watch them, until the network bootstrapped past the point where the subsidy mattered. Threads figured out distribution by inheriting Meta's existing graph and shoving a notification down the throat of every Instagram user in a week. Neither one of those is a product story; both of them are stories about capital and channel, not about product.

The closest local analogue I can name is Daraz, which is now part of Alibaba and which has, for the last six or seven years, been quietly building the only real e-commerce logistics network in Bangladesh, with delivery routes that handle the part nobody else can. The country does not have reliable formal addresses, so the moat is not the website but the rider who knows that the house with the green gate two doors down from the masjid in sector 11 is the one you actually want, and the call you place from the corner when you cannot find it. Aarong does a softer version of the same thing in handicrafts, with well over a hundred retail outlets and an artisan supply network that took thirty years to assemble, and which no competitor built around e-commerce first has come close to reproducing despite three serious attempts since 2019. The TikTok and Threads cases that Spiegel cites are the same lesson in San Francisco vocabulary, which is that the distribution layer ends up being the product the company sells and the screen on the phone ends up being the receipt the customer happens to read it through.

This is the sentence I had to read twice, because it answers a question I have been turning over for about six months without quite landing it. The AI tools we ship now, including the ones I write into the SupplyLine codebase three times a week, are extraordinary at compressing the middle of product work. The spec writes itself, the integration scaffolding writes itself, the email to the merchant writes itself in cleaner English than I'd manage on a Sunday evening. What none of that compresses is the part where you have to put a human in front of another human in a setting that other human already trusts. The agent network is not a software problem. Patrick McKenzie has been making a quieter version of this argument for years over at Bits About Money, where the recurring point is that the moat in unit economics for financial services lives in the part of the business that touches the regulator, the cash, and the customer's actual paperwork. That is exactly the part the model cannot do for you no matter how big the context window gets.

The teams I see in this country who will survive the next two years are the ones who have already figured out that the AI wave makes the distribution moat more valuable, not less. If everyone has access to a roughly equivalent stack of models and roughly equivalent ability to ship a feature on a Tuesday afternoon, then the only thing left to compete on is the part the model cannot touch. That part is the four thousand retailers on the SupplyLine roster, or the three hundred thousand bKash agents, or the Pathao rider on Mohakhali flyover at 6pm. The work nobody can buy out of a Claude subscription is the work that compounds.

This is the bit of the interview where Spiegel sounds least like a Snapchat CEO and most like someone who has lived through one full cycle of consumer technology and is bracing for the next one. The Bangladeshi version of his observation is more visible than the American one, because adoption here is mediated through layers of family, neighbourhood, mosque, and para that the Silicon Valley assumption of frictionless rollout doesn't model at all. The reason bKash took as long as it did to reach the rickshaw wala on Kemal Ataturk Avenue was not a UX problem; it was a trust handshake that needed roughly seven years and three hundred thousand shopkeepers to harden. The reason a chunk of the AI product layer landing in this market over the next eighteen months will stall is going to be the same shape, where the technical answer ships in a sprint and the social answer takes the half a decade it takes.

The honest reading of Spiegel's last point is that the people who win the buildout of AI in Bangladesh will be the ones who treat the social adoption layer with the same seriousness Snap treated camera and lens design with for a decade. Andy Grove called this kind of moment a strategic inflection point in Only the Paranoid Survive, and the part of his framing that has aged the best is the bit about people inside the firm usually being the last to notice that the fundamentals have shifted. The fundamentals here shifted somewhere between November 2022 and the present, and the firms that are still pretending the moat is the cleverness of their feature ladder are about to discover what it felt like to be Nokia in 2009. The ones who already had a thousand boots on the ground will keep their cohorts.

One place I'd push past Spiegel is on the framing that splits consumer from enterprise, because the interview stays mostly on the consumer side and the harder version of the argument lands in enterprise. The B2B SaaS pitch in 2026 is increasingly indistinguishable from the next B2B SaaS pitch, because the models and the design systems and the integration patterns have converged. Distribution in enterprise is a sales motion staffed by humans who have run that motion for a decade, not a subsidy in the TikTok style or a notification push in the Threads style. The firms in this country who will sell into Singapore and Dubai and London over the next three years are the ones who have already hired those humans, not the ones whose founder led sales loop still bottlenecks on a single bhai with an Apollo.io subscription. The compounding here is even slower than the consumer kind, which is part of why it is more defensible when you have it.

The rickshaw wala who will pedal me home tonight knows my face and my flat and which gate to stop at without me telling him, which is the handshake, seven years in the making, that gets repeated, with different actors, across roughly three hundred thousand bKash agents and four thousand SupplyLine retailers and the Pathao rider waiting on a fare at the Mohakhali flyover. That handshake is the only moat the AI wave does not reset, because the feature ladder any of us are climbing this quarter is rented from the same model provider as everyone else's, and the distribution layer underneath was paid for in a currency Anthropic does not sell on a price list.