Build Bridges, Not Moats. (Cuz you can’t.)

“There’s a big difference between mostly dead and all dead. Mostly dead is slightly alive”.
”All dead, well, with all dead there’s usually only one thing you can do. Go through his clothes and look for loose change.” — Miracle Max

“Moat”, “Unfair Advantage”, “Right To Win”

This is the language both of early stage companies looking for product market fit, and companies looking to protect or optimize value. But I have news. Moats are dead. Or maybe, if we’re being charitable, they’re just “mostly dead”.

Dead. Unless you have regulatory shelter, you’re leveraging a monopolistic and/or anticompetitive position, or… Or… you’re doing something that is inherently really hard.

Let’s sharpen and define that edge a little more. If what you’re doing is software, of any kind, “AI” buzzword compliant or not, you have no moat. There’s nothing you’re doing that someone else can’t replicate. And they will.

The Dead Things List™

Pure AI/Software plays (dead on arrival)
Network effects without capital barriers (Instagram Threads vs Twitter… $0 to 100M users)
Brand as primary advantage (Kodak invented this faceplant)
First-mover advantage (fast followers win 70% of the time now)
IP protection (meaningful only in pharma… for now)
Access to capital (everyone has it or can get it)

The Mostly Dead List™

Switching costs in B2B (hanging by a thread)
Distribution advantages (Amazon can still hurt you)
Trade secrets (if you’re Coca-Cola, maybe)

So if defense/moat is dead, what’s left? Maybe just “movement”. The ability to build bridges to value; temporary advantages that get you somewhere valuable before they dissolve. You can’t keep others out. You’re trying to get somewhere where there is still value, extract it, and move on before the bridge collapses or someone builds a better one.

So, what sort of problems still have legit moats? The shitty problems. The impossible problems. The problems that require long chains of exponential compound learning to attack. The problems that require unique “unavailable” data that captures rare, low-incidence real-world phenomena, from which you can build models. The problems that are really problems about atoms.

Thus the only sustainable competitive advantage is: the ability to repeatedly do really hard things, very quickly, over and over. It’s like the Innovator’s Dilemma times a million.

Consider the AI gold rush, in whose tender embrace we presently rest. OpenAI had everything, right? The “best researchers”, first-mover advantage, Microsoft’s checkbook, and a brand that is pretty much synonymous with AI itself. 100 million users faster than any product in history. Moat, moat, moat!

Wrong. Meta (hello metaverse faceplant) gives Llama away for free. Google hail marys Gemini. Anthropic raised $7 billion and built Claude. Oops then… Qwen! Alibaba suddenly beats GPT-4 on reasoning benchmarks. Truly outta right field, DeepSeek (Chinese hedge fund nobody heard of) matches GPT-4 performance at 1/10th the training cost. Eighteen months from “OpenAI will dominate everything” to “which model should I use today?” That entire moat evaporated in less time than it takes to earn a degree.

If your secret recipe is just mathematics and enough capital to buy GPUs, everyone else can bake the same cake.

Software…? Consider poor Slack. Moment of silence for Slack. Remember when Slack was unstoppable? They invented a category! Pwned the narrative! Exponential growth! Microsoft Teams didn’t even exist. But then “suddenly”, Teams is a throw-in with Office, looks “about the same”, except, palatable for the enterprise. Goes from zero to 250 million users while Slack sold itself to Salesforce for less than their private valuation implied they were worth. But let’s be clear: Slack didn’t fail. They built a $27 billion bridge to Salesforce and crossed it before Microsoft finished their Teams highway.

Or we can put Zoom under the same microscope. Pandemic darling! Verb status! $159 billion market cap! Oops, here comes Google Meet, and oops Teams again. Sorry, Zoom’s stock is down 90%. The pattern is pathological: every successful software company is just teaching Microsoft, Google, or Apple what features to add next quarter. Your innovative software product is their roadmap.

Consider network effects. “Our network can’t be replicated!” More users means more value means more users. The flywheel of dominance. Then TikTok showed up. Instagram had a billion users and network effects, unstoppable, right? So they launched Reels, the “perfect” TikTok clone. Whomp whommmp. Didn’t matter. TikTok kept growing. Why? Because ByteDance’s recommendation algorithm was better, and they had infinite capital to pay creators. Oh wait let’s not forget BeReal. Had killer network effects too, right? Your friends were on it, that was the whole point. But then TikTok and Instagram added “dual camera” features. BeReal went from 20 million daily users to statistical noise in just about 100 days. BeReal thought they built a moat. They built a footbridge. The moment Instagram and TikTok noticed people crossing it, they built six-lane highways right next to it. Lesson: your “network effect” is a customer list waiting for someone with deeper pockets to harvest.

Consider brand. “Nobody gets fired for buying IBM” used to be the most famous expression of brand moat. Kodak was photography. Nokia was phones. BlackBerry was business communication. Guys, at their time, these weren’t just products. They were part of our culture. Institutions. Kodak invented the digital camera and still died. Nokia had 50% market share in phones, and it was gone in three years. AND this lifecycle keeps accelerating! Peloton went from luxury fitness brand to meme stock in 18 months. Even fresh brands die fast. Allbirds was (briefly) “the” shoe of Silicon Valley. Everyone was wearing ‘em. Sustainable! Comfortable! Obama wears them! Fast forward: stock down 97%. Brand value is just nostalgic residue. Ask anyone under 25 what IBM even is. “Ok, boomer.”

Consider… capital. Quibi raised $1.75 billion. Not million. Billion with a “b”. Jeffrey Katzenberg and Meg Whitman at the helm. Every Hollywood studio chips in $$. Content, oh the grade-A prime content, from Spielberg, Jennifer Lopez, LeBron James. Dead in six months. TikTok murdered it, starting from literally nothing. Wait, let’s not forget Better.com. $900 million raised to “revolutionize mortgages”, valued at $7 billion. Then rates rose and they fired 900 people on Zoom. Twice. My point? Even infinite capital can’t help when you’re solving the wrong problem or solving it the wrong way. Capital is no moat. It may still be “fuel”, but there are more places than ever to “fill up”.

The most defensible businesses are those attacking problems with compound learning curves where:

Each iteration requires significant time and money
Iterations are gates, so you can’t get to N+1 until you finish N
Failures are catastrophic, expensive, and/or time consuming
The total iterations needed exceeds what any new entrant can afford

Let’s be careful, though. This learning curve should have been Tesla’s fortress. They checked every box. Each car was expensive to build, you couldn’t skip straight to autonomous driving, failures meant recalls and deaths, and they had a 5-million vehicle head start. Perfect compound learning moat, right? BYD has entered the chat… with their own batteries, chips, and motors

BYD collapsed Tesla’s entire learning curve with one move: they made their own batteries. While Tesla was optimizing software to squeeze 5% more range from Panasonic cells, BYD changed the game entirely. Blade batteries that cost 40% less and don’t catch fire. Five million iterations became irrelevant overnight.

The lesson isn’t that compound learning doesn’t matter by itself. You have to compound the right thing. Tesla was learning how to optimize around constraints. BYD just eliminated them.

Tesla’s mistake was they thought they were building a moat around EVs. They were actually building a bridge to electric transportation. BYD thanked them for proving the destination was valuable, then built their own bridge with better materials.

So if you’re betting on compound learning as your moat, ask yourself: Am I learning how to do something hard, or just learning how to cope with someone else’s bottleneck? Because if it’s the latter, someone will eventually remove that bottleneck and make your entire learning curve worthless.

Are any advantages really left? If they remain at all, they might be where the moat is literally made of nightmares:

Being somewhere others can’t be (Mars, the ocean floor, inside a human body)
Doing anything that’s highly regulated
Planet-scale networks of sensors gathering data (Tesla, iPhone, Ring doorbells)
Solving problems where the downside terrifies capital, where IRR calculations make VCs physically ill

I’ll offer a couple of supporting examples.

SpaceX doesn’t just iterate. They blow shit up. Every Starship explosion generates data no competitor can buy, steal, or copy. Get your own ~~tots~~ rockets. At $100 million each. Then kaboom them in public. Most investors would rather fund another “disruptive customer service chatbot”.

Another one: ASML makes the only machines that can etch 3-nanometer chips. One machine: $380 million, 180 tons, 100,000 parts from 800 suppliers, and years to build. Good luck disrupting that with a React app. ASML has got the compound learning right. Each generation of lithography requires solving problems from the previous generation’s solutions. You can’t “skip to 3nm” without solving 5nm. That’s true compound learning, when new knowledge isn’t just accumulated but architecturally required.

Intelligence has been democratized. It left the Valley and now you can find it anywhere with a cell signal or a clear view of a Starlink bird. It’s not “just” a level playing field. It’s a moat-less field. They were all just temporary information asymmetries anyhow. Now that information is free, only physics remains. And atoms are hard.

Implications: If you’re raising money for a software or AI play in 2025, you’re already dead. You just don’t know it yet. If they’re smart, the VCs funding you are betting you’ll get acquired before everyone else figures it out.

Is there any good news? Maybe there is. If you’re willing to tackle the genuinely hard problems, I think the moats might still be found. Ones involving atoms (or electrons, or proteins and pathways), time, and the possibility of catastrophic failure.

The so-so news? If moats are dead, we can stop wasting time trying to dig them. It’s not about permanent defense but strategic movement. Build a bridge to something valuable, like a new market, an acquisition, a customer base. Cross it quickly. Extract value. Then build the next bridge.

Your competitive advantage isn’t the bridge itself, but knowing which shores to connect, when to cross, and recognizing when it’s time pack up camp and go build the next one.