In Part I, I argued that the data center buildout is priced in and the real opportunity lives in the companies that compound after the buildout — the operational AI compounders with physical scale, operational complexity, and distribution moats that get deeper as AI gets better. CAT and freight as the worked examples.
This is the other side of the same coin. Same lens, opposite answer.
The question every company is being asked right now is: does AI compound inside this company, or does it replace what this company sells? If the answer is the first, you get a Swim company. If the answer is the second, or if the moat depends on assumptions that the next computing paradigm invalidates, you get a Sink company. And the market — which has been running the Mag 7 framework for so long that it now treats the framework as physics rather than as a hypothesis — is mispricing several names that fall on the wrong side of the question.
I want to be careful with the register here. The argument is not that any of these companies are about to fail. They will not fail. They are well-capitalized, dominant in their current categories, and operationally excellent. The argument is narrower and more useful: a company that consensus has priced as a secular growth compounder can still be a poor investment if the moat is structurally weakening, even when the company itself is fine. Multiple compression is a kind of sinking. It just looks like nothing happening for a long time, and then a re-rating that consensus calls a surprise.
Two anchor cases. Apple and TSMC. Both have moats. Both moats are weaker than consensus believes. The mechanism is different in each case, which is why both are worth examining.
The structural argument
Before the cases, the framework.
Moats come in two forms. Innovation moats are dynamic — the company stays ahead of competitors by being better at the thing the market actually wants, and the moat compounds because each generation of product extends the capability lead. Lock-in moats are static — the company captured the market at some critical moment and now keeps it through switching costs, ecosystem effects, or installed-base dependencies that are independent of whether the company keeps innovating.
These are not mutually exclusive. The strongest companies in history have both. Apple in 2010 had both — the iPhone was the best phone and the App Store and iCloud made leaving expensive. Microsoft Office in 2000 had both — Word was genuinely the best word processor and every business workflow assumed .doc files. The combination is what builds the dominant franchises.
But the two moat types decay differently, and that’s the part the market underweights.
Innovation moats can compound forever in principle, because each generation refreshes the lead. The risk is execution failure — the company stops innovating and the lead disappears in a few cycles. This is a discrete risk. Either the next product works or it doesn’t.
Lock-in moats decay continuously, but slowly, and only become visible when a new paradigm offers users an exit. The risk is that the exit appears before the company has built a position in the new paradigm. This is a structural risk, not an execution risk. The company can keep doing exactly what it did before, and the moat still erodes, because what changed is the world around it.
The dangerous configuration — and this is the central claim of the piece — is a company whose innovation moat has decayed but whose lock-in moat is doing all the work of holding the position, while the next paradigm shift offers the exit ramp. That company looks dominant for years. Quarterly numbers are fine. Margins are healthy. Then a new paradigm catches and the moat starts leaking from a place no one was watching, because the lock-in was the load-bearing piece and the lock-in only worked as long as there was no escape.
That’s the lens. Now the cases.
Case one: Apple
Apple had a record 2025. Total revenue hit $416 billion, up 6% year-over-year. Services revenue crossed $109 billion at a 76.5% gross margin. The iPhone 17 cycle had demand outstripping supply through Q1 2026, and management guided for a December 2025 quarter that would be the best in company history. The active installed base passed 2.5 billion devices.
By every conventional metric, Apple is doing fine. This is exactly the configuration where a multiple compression argument is hardest to make and most useful to consider — because if the argument were easy to see in the current numbers, it would already be in the multiple.
Here’s what’s actually happening underneath the headline.
The innovation moat decayed somewhere between the introduction of the AirPods in 2016 and the introduction of Vision Pro in 2024. That window includes Apple Watch (a real product, but a watch is not an iPhone-class category creation), AirPods (genuinely excellent, also not a category creation — true wireless earbuds were already happening), Apple Silicon (the most important product accomplishment of the period, but a component rather than a new category), and Vision Pro (the swing at the next computing paradigm that didn’t connect — held back by the $3,499 price tag, limited use cases, and the fundamental problem that nobody figured out what spatial computing is for).
The point is not that any of these were failures. AirPods alone is a top-50 global business. The point is that the company has not introduced a category-defining product since the iPhone in 2007, and the gap between iPhone-class categories has stretched from 4-7 years (Cook’s preferred framing) to potentially indefinite. The Watch and AirPods are accessories to the iPhone. Vision Pro was the swing at being something else, and it didn’t land. Apple is, structurally, an iPhone company with high-margin services attached to the iPhone installed base. Everything else is decoration.
So what’s holding the moat? The lock-in. iMessage, iCloud, the App Store, FaceTime, AirDrop, family sharing, payment friction, the whole accumulated weight of switching costs. The Services number — $109 billion, growing 14% — is the lock-in monetized. App Store fees, advertising on the App Store, iCloud storage, Apple Music, Apple Pay — these are rents on the installed base, not new category creation. The 76.5% gross margin tells you exactly what kind of business this is. It’s a tax on people who can’t easily leave.
This is fine. As long as there’s no exit ramp, this business runs forever. The argument for Apple at current multiples is precisely that the lock-in is permanent, the iPhone is forever, and Services compound off the installed base into perpetuity.
The argument against is the AI capex line.
In 2025, Apple spent roughly $14 billion on AI infrastructure. In the same period, Amazon spent $200 billion. Google spent $180 billion. Microsoft and Meta were in similar ranges to the latter two. Apple’s AI capex is, by an order of magnitude, the smallest in the Mag 7. The strategic framing the company has chosen — privacy-first, on-device AI, with selective partnerships for the heavy cloud inference (the Gemini partnership has been telegraphed) — is internally coherent. It’s also a bet that the next computing paradigm doesn’t require owning the foundation model layer.
If that bet is right, Apple is fine. The iPhone remains the default computing surface, AI is a feature on top of the iPhone, and the lock-in continues to monetize.
If that bet is wrong — if the next computing paradigm is an agent layer that runs above the device and treats the iPhone as a commodity input rather than a privileged platform — the lock-in starts leaking. Not because users leave Apple. Because the value of being on Apple’s platform compresses. iMessage doesn’t matter as much when an agent handles your messaging. The App Store doesn’t matter as much when agents transact directly with services. Apple’s 30% take on app revenue doesn’t survive a paradigm where apps are not the unit of consumption.
Notice what this argument is not. It’s not that Apple shutters or fades. It’s that Apple becomes a premium consumer hardware company with a healthy services business, instead of a secular growth compounder with infinite operating leverage. That’s a different multiple. The current multiple — Apple trades at a premium to the S&P even after recent compression — assumes the secular growth framing. The risk isn’t existential. The risk is that the framing flips, and a company whose earnings might still grow at 6-8% gets re-rated from a 30x multiple to a 20x multiple. That’s a 33% drawdown without a single quarterly miss.
The TAM tell is the AI capex. Companies that believe they need to own the next paradigm spend like they need to own it. Apple’s $14 billion against Amazon’s $200 billion is either a brilliant capital-discipline call that will look prescient in 2030, or it’s the company already conceding the next platform layer to others while extracting maximum rents from the current one. Either reading is possible. The market is currently pricing the first reading. The second is at least as plausible, and it’s the one where the multiple compresses.
The honest version of the bull case is: Apple’s installed base is so large and so sticky that even a partial loss of paradigm leadership leaves a fantastic business. The honest version of the bear case is: that’s true at a 20x multiple. It’s not true at a 30x multiple. The space between those multiples is the trade.
Case two: TSMC versus ASML
The TSMC case is structurally different from Apple. Apple is about whether innovation has decayed under a healthy lock-in moat. TSMC is about whether a moat that consensus considers impenetrable is actually a series of papercut-vulnerable positions stacked on top of a single real monopoly that TSMC doesn’t own.
Set up the contrast. Most investors think of TSMC and ASML as both being monopolies in semiconductor manufacturing. They are not the same kind of monopoly.
ASML is a literal monopoly on extreme ultraviolet lithography. There is exactly one company on Earth that makes EUV machines. Each machine costs $200-400 million depending on the configuration. The lead time on a new tool is years. The R&D investment to replicate the system is measured in decades. Nikon and Canon, the other major lithography vendors, both abandoned the EUV development path in the 2010s after concluding the engineering problems were not solvable on a competitive timeline. The Chinese state has been openly trying to build an EUV alternative since at least 2018 and is, optimistically, ten years away from a usable system. ASML’s moat is not “we are very good at this.” ASML’s moat is “physics, supply chain, and twenty years of accumulated process knowledge make us the only option, and there is no realistic timeline on which that changes.”
TSMC’s moat is different. TSMC is the best at what they do. They have the highest yields, the most mature process, the deepest customer relationships, the best ecosystem, the most experienced engineering organization. They captured 71% foundry market share in Q3 2025 and they make over 90% of the world’s most advanced chips. Apple, AMD, Nvidia, Qualcomm, Broadcom, and most of the AI accelerator startups all manufacture there.
But “the best” and “the only” are entirely different moats. And TSMC’s moat is the first kind, not the second.
Here’s the state of play in late 2025 / early 2026 on the leading edge.
TSMC put N2 (their 2nm node) into volume production in Q4 2025. That’s the current frontier. Around the same window:
Intel’s 18A node — comparable to N2, with RibbonFET (their version of gate-all-around) and PowerVia backside power delivery — is in early production. Yields are reportedly 55%, rising to 65-70% by end of 2025. Panther Lake, the first 18A consumer product, is being built at Fab 52 in Arizona. There are credible rumors that IBM is moving from Samsung to Intel 18A. Intel is targeting 100,000+ wafers per month at 18A by 2027.
Samsung is putting SF2 (their 2nm node) into production in 2025-2026. Tesla has signed a 2nm contract with Samsung. The yield is behind TSMC’s, but Samsung has anchor customers and U.S. fab capacity.
Rapidus, the Japanese government-backed foundry founded in 2022, is prototyping 2nm GAA transistors at its IIM-1 fab in Hokkaido. Early electrical characteristics are reportedly meeting targets. They received their first EUV machine from ASML in December 2024 and are aiming for trial production by mid-2025, volume production targeted for 2027. They’re collaborating with IBM and imec on the process.
The Chinese mature-node push (SMIC and others) is not at the leading edge but is aggressively expanding 14nm/7nm capacity, eating the trailing-edge volume that TSMC used to harvest at high margin.
Now here’s the question that actually matters: does any one of these competitors need to displace TSMC for the moat to leak?
No. They don’t. That’s the death by 1000 papercuts.
Intel doesn’t need to take Apple. Intel takes IBM, takes some defense and government work, takes a slice of the AI accelerator market that wants U.S.-soil manufacturing. Samsung doesn’t need to take Nvidia. Samsung takes Tesla, takes some of the second-source allocations from customers who want supply chain redundancy. Rapidus doesn’t need to take TSMC’s whole book. Rapidus takes the Japanese domestic and friendly-Asian customers who want supply chain sovereignty. The Chinese mature-node fabs don’t need to take 2nm. They take the trailing-edge volume that funds TSMC’s leading-edge R&D.
Each of these is a small bleed. Together they compress TSMC’s pricing power, force higher capex to maintain the lead, and shrink the share of the market where TSMC is the only viable choice rather than the best viable choice.
And the structural problem underneath all of this is that TSMC’s moat depends on every one of its customers preferring TSMC’s execution to the alternatives, every year, forever, on every node. That’s a probability distribution where each draw is favorable but the cumulative product across decades is not 100%. ASML’s moat doesn’t have that problem. ASML’s moat is that the customer has no alternatives at all on the question that matters most, which is whether the chip can be manufactured in the first place.
Here’s the cleanest expression of the asymmetry. If you imagine 10 years out, what does TSMC look like? Probably: still the largest foundry, still the highest-volume leading-edge manufacturer, but with 50-60% leading-edge market share instead of 90%, with margins that have compressed from current levels because Intel and Samsung took enough volume to force pricing concessions, with a stock that has compounded at a lower rate than the Mag 7 because the moat narrative collapsed even as the business stayed great.
What does ASML look like 10 years out? Probably: still the only EUV vendor on the planet, with the same 100% share of the leading-edge lithography market, with pricing power that has gotten better because High-NA EUV is even harder to compete with than standard EUV, and with a customer base that includes whoever wins the foundry war — TSMC, Intel, Samsung, Rapidus — because all of them buy from ASML. ASML’s revenue grows whether TSMC keeps its share or loses it. ASML wins the foundry war regardless of which foundry wins.
This is the structural distinction the market currently underweights. Both stocks trade like monopoly compounders. Only one of them actually is one. The question isn’t whether TSMC is a great business — it is. The question is whether the multiple correctly distinguishes “great business with a leaky moat” from “monopoly with no realistic competitor on a 20-year horizon.” Right now, it doesn’t.
What this means
Three things.
First, the Mag 7 framework is not physics. It is a hypothesis — that the seven dominant tech franchises of the last decade will continue to compound at premium multiples through the AI transition because their moats are durable. That hypothesis is true for some of them. It’s not obviously true for all of them, and the current multiples assume it’s true for all of them. The framework is wrong about Apple specifically, and the asymmetry on TSMC versus ASML is one of the cleanest underpriced distinctions in the market.
Second, Sink and Swim are the same question. Does AI compound inside this company, or does it replace what this company sells? The Swim companies are the ones where AI deepens the existing moat. The Sink companies are the ones where AI offers the exit ramp from the existing moat. If you can answer this question for any company in your portfolio, you know whether the next decade rewards or punishes the position.
Third — and this is the part that’s harder to write about because it requires some intellectual honesty — every name on either side of this analysis can still be a great business in absolute terms while being a poor investment in relative terms. Apple will be fine. TSMC will be fine. The question is whether they keep their multiples, and the multiple is where most of the return lives. The line between “great company” and “great investment” is exactly the line drawn by whether the moat is compounding or decaying. That distinction does not show up in the current quarter. It shows up over a decade.
The market is paying for the next decade. Make sure you know which one of those it’s actually paying for.
This concludes the two-part Sink or Swim series. The framework — physical scale, operational complexity, distribution moat for Swim; innovation decay, lock-in dependency, paradigm exit ramp for Sink — is built to be reusable.