The corpus of my “second brain” comes from the books I’ve read, the essays I’ve written, the meetings / conversations I’ve had, and the tweets I’ve consumed. Over the last few weeks, I’ve been using a personalized iteration of Karpathy’s LLM Knowledge Base to fully ingest all the notes I already have across Roam, Notion, Google Docs, etc.

I’m only ~1/50th of the way through that process and already I have 3M+ words of context in my second brain. But one of the immediate exercises I can do is asking it to look for connections of ideas I’ve tacitly made across book notes and essay research fodder that I may not have explicitly drawn myself. The first one that caught my eye was about moats.

---

Revisiting my notes on [[Brad Stone]]‘s [[The Everything Store]] that I read in 2021, there were two little notes I wrote to myself that, at the time, are direct companions that I didn’t ever connect before.

The first, scribbled around page 390, next to a passage about [[Amazon]]‘s cost structure: “Being willing to accept the slimmest margins and building an organization that can operate lean enough to not need margin can be a competitive differentiation.”

The second, a hundred pages later, next to a passage about how Amazon’s e-book market share fell from 90% in 2010 to less than 60% by 2012, just four words: “Moats don’t stay impenetrable forever.”

Over the last three years, I’ve written over and over and over again about competitive moats in large part because they represent a paradox to me; simultaneously held up as one of the most critical aspects of any business, while also acknowledging material evidence that they almost never really exist.

But seeing those notes together helped me appreciate that the paradox comes in acknowledging that moats come from a willingness to endure repetitive discomfort. It lasts as long as you’re willing to accept discomfort. Once you’re no longer willing to accept discomfort, its only a matter of time until your moat goes away.

Rather than being an impenetrable wall that remains stable in stasis, its more like a shelf-life; something can last a certain amount of time but it only persists if you’re willing to continue to dig it. The willingness to dig a new moat while the old one is still up.

Thin margins, it turns out, are the cleanest example of both halves of that idea.

Two Kinds of Retailers

The last time I wrote about [[Competitive Moats]], I argued that for most of a company’s journey, defensibility is for dummies. Founders fall in love with a fuzzy story about why they’re safe, right up until the moment they run out of cash, time, or talent and the moat becomes moot. I still believe that. The element of [[Jeff Bezos|Bezos]]’ moat formation around Amazon is unique from that idea of comfort in that it comes in the never-ending pursuit of moats that makes the difference.

The famous line is the one he repeated, in Stone’s words, “ad nauseam for years”:

“There are two kinds of retailers: there are those folks who work to figure how to charge more, and there are companies that work to figure how to charge less, and we are going to be the second, full-stop.”

Most people’s perspective on low margins is that they’re a weakness; a sign you have no pricing power, no brand, no defensibility. Every fresh would-be MBA thought-leader treats fat gross margin as the goal and thin margins as ritual humiliation.

But once again, Bezos takes the counter-position:

“Bezos believed his company had a natural advantage in its cost structure and ability to survive in the thin atmosphere of low-margin businesses. Companies like IBM, Microsoft, and Google, he suspected, would hesitate to get into such markets because it would depress their overall profit margins.”

And the cleaner statement of the thesis:

“Bezos believed that high margins justified rivals’ investments in research and development and attracted more competition, while low margins attracted customers and were more defensible.”

That’s a critical crux of the moat discussion. The thing that looks like the weakness is the wall. Fat margins are an open invitation that fund your competitors’ R&D budgets and signal to every adjacent giant that there’s room to come take a slice. Thin margins do the opposite. They’re the double black diamond warning sign that keeps all but the most disciplined operators away. IBM, Microsoft, and Google could have followed Amazon into the thin atmosphere. They structurally wouldn’t, because doing so would crater the blended margins their own shareholders were paying for.

This is, I’ll admit, the inverse of the usual moat story, and for a long time it looked insane from the outside. When Scott Devitt at Stifel Nicolaus upgraded Amazon to a buy in 2007, given their deeply in the red margin profile, here’s what happened:

“I was laughed out of portfolio managers’ offices,” Devitt says. “People were ripping apart every component of my investment thesis. At that point, they thought Amazon was some kind of nonprofit scam.”

A company that looks like a non-profit scam from the outside is exactly what a moat looks like when its being dug; just a giant hole in the ground.

A Wall vs. An Engine

So why does this kind of moat behave differently from the ones we usually obsess over? Because thin margins aren’t actually a thing you just have, they’re a thing you do.

The classic moats, be they patents, network effects, switching costs, regulatory capture, a brand, all get talked about as assets. Things you accumulate and then sit on. One of the other ideas that I touched on the last time I wrote about [[Competitive Moats]] was this idea that the real engine getting built is the actual advantage: the intangible muscles a company builds that “never appear on a balance sheet” but “are the single most valuable asset of a company.”

One cultural artifact that comes to mind is the “my body is a machine…” meme. I turn X into Y. That’s a system; an engine.

Thin margins build exactly the kind of muscle that can breed long-term success. Bezos didn’t just find a low-margin advantage lying around; he spent two decades building “an organization that can operate lean enough to not need margin” An engine that gets built piece by piece.

I’ve heard multiple people make an argument that [[Anthropic]]‘s product advantage has, in large part, come from forced constraints. From the beginning they were always more compute-constrained than [[OpenAI]] or any of the hyperscalers. As a result, they had to get better at the intelligence value chain of turning kWh into FLOP into output tokens, than anybody else. That’s an engine built over time.

What makes it a moat isn’t just that your competitor can’t copy it, but that they don’t even want to try. What’s more, accepting the consequences of running the engine aren’t just a matter of saying yes or no any more than me saying my Honda mini-van is now going to hit 160 MPH on the race track. You don’t just decide to have a muscle. You have to build it over time.

[[Elon Musk]]‘s perspective is that “moats are lame… what matters is the pace of innovation.” My argument in [[Momentum != Moat]] was that the whole point is building “high-quality, long-term sustainable economic engines,” not heat-seeking the next data point on your ARR chart. The even harder thing about moats as an engine vs. a wall is that a wall usually breaks down when other people try to pull them down; otherwise, they could just sit there; monopoly style. But an engine breaks down in the process of running. It requires significant maintenance.

Moats Are Never Forever

The key question around competitive defensibility relies on the rate of decay; how quickly does the engine break down once you’ve built it? Because its not just a question of “moats are moot,” and therefore don’t exist.

Even Amazon has run face-first into its own fair share of other people’s moats. One example? In the 90’s, Amazon threw a ton of talent and capital into online auctions and built what they felt was clearly a superior auctions product. But in the end they failed. [[eBay]] won. “Network effect mattered.”

But that was a moment in time. Because eBay’s engine was just left running, with no consideration for maintenance:

“At the same time that Amazon’s flywheel was accelerating, eBay’s was flying apart. The appeal of online auctions had faded; a customer wanted the convenience and certainty of a quickly completed purchase, not a seven-day waiting period to see if his aggressively low bid for a set of Cobra golf clubs had won the day… Amazon had battled and mastered chaos; eBay was engulfed by it”

The competitive landscape shifted like sand on the beach, and eBay’s moat fell apart because its engine never evolved. I’ve written before about the way [[Permanent Equity]] has described this dynamic:

“Moats are routinely discussed in a static sense, as if they exist in some form or fashion for most businesses. But the facts simply don’t bear this out… Averaged across industries, a business in its 25th year has roughly the same probability of dying as it did in its 10th year.”

Twenty-five years of “moat” and your odds of dying this year are the same as they were at year ten. So… like I said, that’s more like a shelf-life than a defensible wall.

Your Competitors Today vs. Tomorrow

So was Amazon’s ultimate victory assured just by low-margins? Or the willingness to invest in an engine? It’s actually the same. Going back to Bezos’ framework of two retailers; his willingness to be the kind of retailer looking to charge less forced them to trim the fat.

One clear playbook was Amazon’s productization of its cost structure. Since 2005, Amazon has taken every major aspect of their cost structure and turned it into a business line they can not only take advantage of but also monetize. In fact, the final line-item, G&A, you could argue that as Amazon continue to invest in AI and automate portions of its overhead, that could likely also become a product line for them.

But the natural consequence of that kind of responsiveness will expose you to a dramatically broader playing field of competitors. In the 90’s, Amazon was laser-focused on eBay and Barnes & Noble. Now, I can’t imagine either ever comes up as a concern to Amazon. Instead they’re focused on Wal-Mart in ecommerce, Google and Microsoft in Cloud Computing, Netflix and Disney in streaming, UPS and FedEx in logistics, Costco in grocery, and on and on.

When I think about OpenAI, I think that’s one of the reasons their empire expansion is failing. They tried to do everything very quickly because it was there, not because it was necessarily advantageous. Competing across language, image, and video models while also trying to buy Jony Ive to build consumer hardware and raise a trillion dollars for your own compute; it lacked cohesion.

Anthropic, on the other hand, I would argue leaned into a core competency and expanded from there. Their swings are much less egregious; launching a new cybersecurity or legal or design product is quite logical as an extension of their core competency around the intelligence value chain. Meanwhile, OpenAI’s expansion was never funded by its own existing engine; it was speed in pursuit of an engine.

What Elon focuses on with the “pace of innovation” and what I’ve been clumsily circling for three years is really just this: the only durable moat is the capacity to build the next moat faster than the current one decays. Otherwise, you just have an engine that is slowly breaking down. Or worse, a wall that’s just standing there defending nothing worth defending.