If I ever build something, this is exactly how I'd think about where to stand.
It started, as most useful thoughts do, with a question that felt almost too obvious to ask: why does NVIDIA keep winning no matter who's actually ahead in the AI race?
Think about what that race actually looks like from the outside. Anthropic, OpenAI, Google DeepMind, Meta AI, Mistral, DeepSeek - all of them extraordinary, all of them furiously competing, all of them eating each other's lunch every few months when someone drops a benchmark-destroying model. The leaderboard is a revolving door. Nobody stays on top. The narrative shifts every quarter. And through all of it, NVIDIA just keeps compounding.
The reason, once you actually sit with it, is almost embarrassingly clean: NVIDIA doesn't compete in the AI race. It profits from the existence of the race itself. Every lab - the ones winning, the ones losing, the ones burning cash to catch up - all of them run on NVIDIA hardware. NVIDIA isn't a player in the game. It built the stadium, and it charges rent to everyone who wants to play.
Once that framing clicked, I started seeing the same pattern everywhere I looked.
Take ASML. Most people outside semiconductor circles have never heard of it. ASML makes EUV lithography machines - essentially the only printers on earth capable of printing the most advanced chips. TSMC uses them. Samsung uses them. Intel uses them. Without ASML's machines, the chips that run your phone, your laptop, your car, and increasingly your entire world simply cannot be manufactured.
Now consider TSMC - one of the most impressive industrial operations ever built. Except TSMC has a gun permanently pointed at its own head in the form of geography. Taiwan's relationship with China is not stable. If that situation deteriorates seriously, the western world would scramble. New fabs would get funded in Arizona, Japan, Germany. The chip manufacturing map would redraw itself.
But through all of that chaos, ASML still wins. Because whoever replaces TSMC - whatever new fab rises in its place - will still need ASML's machines to print the chips. The printer-maker doesn't care who's doing the printing. China has spent hundreds of billions trying to replicate EUV technology and is still roughly a decade behind. That's not a moat. That's a canyon with no visible bottom.
The principle, once you see it: the further you are from the end consumer, and the more your product is a non-substitutable dependency for an entire competitive ecosystem, the more durable your business.
Not: who is my customer? But: what would have to be true for my product to become structurally unavoidable? What position in the value chain makes every competitor in the layer above me also my customer? What do I build such that the more intense the competition above me gets, the better it is for me?
Those are different questions. Most people never ask them.
I should say upfront: the dependency layer principle isn't the only path. There's a second, entirely different way to build something durable - and Apple is the proof of it.
I was an Android person for years. Genuinely believed Apple was overpriced, overhyped, and designed for people who didn't want to think. Then I used an iPhone for a week and something shifted. Not because any single feature was revolutionary. It was the feeling that everything just worked together - that the device, the software, the services, all of it was made by a single entity that had thought carefully about how they'd coexist. I couldn't name what was missing on Android until I felt its presence on iOS.
What I was feeling was the result of full-stack vertical integration executed with near-obsessive coherence. Apple designs the chips. Writes the OS. Builds the hardware. Designs the services. Owns the payment layer, the wearables, the laptops. Every dollar that would have left the ecosystem stays inside it. Every design decision is made holistically rather than as a compromise between vendors who've never been in the same room.
Android is a coalition by design. Qualcomm and MediaTek make chips. OEMs assemble devices. Google writes the OS but doesn't control how Samsung or Xiaomi ships it. No two things are truly under one roof. The result isn't necessarily worse in any individual dimension - but it never feels whole.
The reason Apple users don't switch isn't loyalty or aesthetics, though those play a role. It's that the switching cost compounds with every Apple product you add. iMessage, AirDrop, iCloud, Apple Watch, Handoff - none of these are individually irreplaceable. But collectively they form a web that becomes more expensive to leave every single year. Apple built lock-in not through contracts or restrictions, but through coherence. That's a significantly harder thing to fight.
If I'm honest with myself about what kind of thing I'd want to build - the Apple model is the more romantic one. The idea of owning every layer and making something that feels genuinely complete is deeply appealing. But the Apple model has a prerequisite that most people skip over: you have to be willing to serve one master, ruthlessly and exclusively. One customer. One experience. Everything else be damned.
Apple killed features. Abandoned technologies early. Refused to licence iOS at scale even when doing so would have dramatically expanded their market. The coherence isn't accidental - it's the product of a thousand decisions to say no to things that would have diluted the singular vision. Most people building things aren't willing to make that call consistently enough, for long enough, to get the outcome Apple got.
Everything up to this point has been NVIDIA, ASML, Apple. Trillion-dollar companies. Abstract principles applied to abstractions. Easy to nod along to and equally easy to file away as "interesting but irrelevant to my life."
Then I thought about Mold-Tek Packaging Limited.
Most people reading this have never heard of them. Mold-Tek is a Hyderabad-based company that makes rigid plastic packaging - the containers you don't think about when you buy a tin of Asian Paints, a litre of Castrol, a tub of Vadilal ice cream. As of late May 2026, it has a market cap of roughly ₹2,300 crore. Not a giant. Not a headline. Not something your relatives bring up at dinner as a hot tip.
But think about the position it occupies. Asian Paints fights Berger. Berger fights Indigo. Indigo fights Kansai. They compete on advertising, dealer margins, colour cards, brand ambassadors - the full consumer brand war. Now ask: who packages all of them? Who makes the container that every one of these companies needs before their product can leave the factory floor and reach a consumer?
Mold-Tek doesn't care who wins the paint war. It invoices every combatant. Same with lubricants - Castrol, Gulf, Servo, all fighting each other, all needing containers. The company nobody roots for is quietly billing all of them.
The in-house tooling is where it gets structurally interesting. Mold-Tek designs and manufactures its own molds, integrates IML labelling in-house, and maintains quality consistency that a client has spent months calibrating to. Switching to a competitor means rebuilding tooling, re-calibrating quality standards, absorbing operational disruption - for a cost saving that probably isn't worth it. That's a real switching cost, engineered through capability rather than contract.
Revenue of ₹886 crore for FY26, 13.5% year-on-year growth, net profit up 20%. These are not exciting numbers by the standards of a hot sector. No one is writing breathless posts about Mold-Tek's Square Packs division. And that's almost the point. The businesses that fit this framework tend to be exactly this quiet. They don't need to be loud. The dependency does the work.
I want to be honest about where the analogy strains, though. Mold-Tek is not ASML. Rigid plastic packaging with IML is an excellent capability but it is not technically irreplaceable. Other players exist. A large client with real motivation could develop an alternate supplier over 18-24 months - painful and expensive, but possible. Nobody replaces ASML in 18-24 months or 18-24 years. The moat is real but it's a wide ditch, not a canyon. The P/E of 32 on an ROE of ~11% means the market has already noticed this. You're not buying a hidden gem at current prices - you're paying a fair premium for a structurally sound business.
But the principle it illustrates is entirely sound - and it's the same principle operating at a scale that's actually accessible, in a market I can actually think about, made by a company in a city I've driven through. That matters to me.
I'm not naive enough to think building a business is as simple as picking the right layer of the value chain and waiting for the money to arrive. Execution is real. Timing is real. Capital is real. There are brilliant people with structurally sound ideas who still failed because they ran out of runway three months before product-market fit. None of what I've written above overrides any of that.
But if I ever start something - and I think about this more than I probably let on - the questions I'd ask myself first are not the ones most people start with.
Not: is this market large enough? Not: can I acquire customers cheaply? Not: is there a viral loop here?
The first question I'd ask is: am I building something that every competitor in my ecosystem eventually needs, or am I building something that competes with them?
Because if I'm competing - if my success requires me to beat someone else in a war of features, marketing, pricing, and distribution - then I'm in the AI lab position. I'm Berger trying to out-advertise Asian Paints. I'm fighting on ground where the outcome is genuinely uncertain, where a better-funded competitor can show up tomorrow and reset the board, where my moat is only as deep as my last product release.
If I can find a position where my success is a function of the entire ecosystem growing - where more competition above me means more revenue for me - that's a different game. A fundamentally different compounding curve.
The second question I'd ask: what makes my position genuinely hard to replace? Not hard to compete with. Hard to replace. There's a difference. A competitor with more funding can match features. Nobody rebuilds 15 years of CUDA developer adoption in a funding round. Nobody rebuilds Mold-Tek's in-house tooling capability and client calibration overnight. The switching cost has to be structural, not just inconvenient.
And if I can't honestly answer both questions in a way that satisfies me - if I'm in a position where I'm one of many players fighting for the same consumer's attention and wallet - I'd at least want to go in with my eyes open about what kind of war I've signed up for.
The real money, almost always, is one layer further back than where everyone is looking. It's in the company making the containers that every paint brand ships in. It's in the machines that print the chips that run the models that everyone argues about. It's in the runtime that every AI application is built on top of, not the application itself.
It's quiet there. Nobody writes about it. The founder doesn't get invited to podcast panels to talk about disruption. The company doesn't trend on any platform.
It just compounds. Quietly. Reliably. For decades.
That's the business I'd want to build.