Today TSMC printed a record quarter. Revenue $40.20 billion against a $39.76 billion estimate, up 34% year over year. EPS $4.31 against a $3.77 estimate — a 14% beat. Gross margin 67.7%, up from 58.6% a year ago. Operating margin 60.3%, up from 49.6%. Profit up 77% from a year ago. Management now tracks full-year 2026 revenue growth above 40%, ahead of the "above 30%" guide they started the year with, and issued a bullish third-quarter outlook on top of it.

The stock fell 4.6%.

ASML raised guidance by €7 billion yesterday. HBM memory is sold out through 2027. Every fundamental print this month has said the same thing: demand is not the problem. And the whole semiconductor sector is red today anyway.

There are reasons for this selloff. Just not the ones that matter.

If you already hold TSM and watched it go red on your screen this morning, or you have been waiting for a pullback on this name that never seemed to come until today — this is for you.

The business, plainly

TSMC does not design a single chip that ends up in your phone, your car, or the data center training the model you used this morning. It manufactures almost all of them. Every leading-edge AI chip on Earth — Nvidia's, AMD's, Apple's, Broadcom's — is built in a TSMC fab, because nobody else can build it at the size and yield this requires.

That is the actual mechanism — not the vague "AI demand" headline. The chokepoint is advanced packaging — CoWoS — the process that stitches compute dies and memory into one chip. TSMC is targeting roughly 127,000 CoWoS wafers a month by the end of this year, and Nvidia alone has already locked up more than half of it through 2027. When your customers are pre-booking your factory two years out, you are not selling into demand. You are rationing it.

Here is the number that proves it. Wafer shipments this quarter: 4.336 million, up 17% year over year. Wafer ASP: $7,881, up 13% year over year. Say those two numbers out loud, 17% and 13%, and notice they don't fight each other. TSMC sold 17% more wafers and charged 13% more per wafer, in the same quarter. A company with real competition picks one: sell more, or charge more. TSMC did both, at once, because the customer on the other end of that order has nowhere else to go.

That is why today's print carried a 55.6% net margin. Net, not gross — more than half of every dollar TSMC books off the top line reaches the bottom line as pure profit. That is what it looks like when your customer has no second supplier to call.

Current price: $400.18. Down from a 52-week high of $478.89, up 32% since January. Dividend yield 0.83%.

Why the stock fell on a beat-and-raise

Here is the part worth sitting with, because it is the whole thesis.

The bears will point to capex: management now tracks $60-64 billion for the year, up from a prior $56 billion guide, and flagged that the 2nm ramp alone dilutes gross margin by 3 to 4 points in the second half. On top of that, TSMC just committed another $100 billion to Arizona — ten fabs and an R&D center, $265 billion total in the US. Free cash flow ticked down. That is a real, legitimate line item. It is also the cost of building the only advanced fab network outside a single geography, which is the thing that removes the biggest tail risk this stock has ever carried. And management did not touch the long-term revenue or AI growth targets while raising near-term spending — this is the plan working, not the plan changing.

On the earnings call, one analyst pushed CEO C.C. Wei to size the gap between demand for 3-nanometer-and-below capacity and what TSMC can actually supply — he floated 30 to 50%, and asked if it could be bigger. Wei didn't give a number back. He said: "The gap is very big." That is not a company estimating demand from outside forecasts. TSMC builds its roadmap several years ahead alongside the chip designers themselves, and talks directly to the cloud companies who will end up renting that compute. Advanced-node capacity now takes five to seven years to develop, build, and ramp. The capex raise is not TSMC chasing this quarter. It is TSMC building toward a gap its own CEO won't even put a ceiling on.

Wei went further on the same call. Asked how long the current demand holds, he said it stays strong "all the way to probably 2029, 2030." That is one of the longest visibility windows TSMC has ever given the market, and it lands almost exactly on the far edge of the three-year scenario table below. The bottleneck is not easing. That is bullish for the whole leading-edge and memory supply chain sitting behind it — NVDA, AMD, AVGO, MRVL, MU, and SK Hynix all read through on the same gap.

But that is not what moved semis red across the board today. Korea's Kospi dropped 6% on the same morning TSMC beat by double digits — you could smell the panic before you could find a fundamental reason for it. That is not a company-specific capex worry — that is a crowded trade unwinding. Everybody who wanted semiconductor exposure already owns the same five names. The trade got stretched after a year of enormous gains, a DeepSeek-builds-its-own-chip headline gave the sellers a story to point to, and the desks derisked into it regardless of what the actual earnings said.

This is not the first time this year. Micron. The Korea crash in the spring. The CPO scare. Each time, the fundamentals were fine within weeks, and the stock that got dumped on the loudest headline was the one worth buying into the panic. Same movie. This is week five of that same pattern this month alone — ASML, Meta, TSMC, the inflation prints, HBM sold out — and the tape went the other way on every single one.

When the fab that manufactures everyone's AI chips is this pricing-powerful and the stock falls anyway on a day it printed record numbers, the stock is wrong. Not the fab.

The honest counter-argument

I am not going to wave away the real risks, because a moat does not cancel a geography. Here is the bitter part to swallow before the sweet part.

Taiwan is Taiwan. That risk was true last year and it is true today, and no amount of Arizona fab progress erases it inside the next few years — it only shrinks it slowly, $100 billion at a time. The 2026 capex raise is running roughly in step with revenue growth, not ahead of it, but it is a five-to-seven-year bet on a demand gap Wei himself won't size — if that gap turns out to be smaller than "very big," TSMC still has to eat the depreciation on capacity nobody rents. And if AI infrastructure spending ever does digest hard the way SUNB's rental cycle is digesting right now, TSMC's margins compress with everyone else's. At a forward P/E near 27x, this is priced as a compounder already, not as a name nobody has discovered. You are not getting a hidden gem. You are getting a proven monopoly at a fair-to-good price on a scary day.

Five scenarios, 3 years out

| Scenario | Price target | Move from $400 | What has to happen | |---|---|---|---| | Super Bull | $1,119 | +180% | AI capex keeps compounding through 2029, CoWoS stays sold out, geographic de-risking finishes without permanent margin drag, and the market re-rates TSMC as AI infrastructure's toll road (32x). | | Bull | $780 | +95% | Growth decelerates gently but stays strong (~22% EPS CAGR), overseas dilution proves temporary, multiple holds near today's level (27x). | | Base | $532 | +33% | Growth normalizes to the mid-teens as the supercycle matures, margins stabilize after the overseas ramp, multiple compresses slightly toward TSMC's historical range (22x). | | Bear | $286 | -28% | Hyperscaler capex digests hard for a year, overseas dilution runs longer than guided, geopolitical friction resurfaces, multiple de-rates to 16x. | | Super Bear | $128 | -68% | A genuine AI capex bust or a Taiwan Strait crisis disrupts fab operations directly, growth reverses, the market prices in existential risk at 11x. |

I built this table from TSMC's own reported numbers and my own assumptions on EPS growth and exit multiple. It is a framework, not a forecast, and definitely not the number a Wall Street desk would defend on a call. For reference, the actual sell-side spread runs from a $423 average target up toward $600-625 on the highest estimates — my Base and Bull cases sit inside that real range, not off in fantasy land.

The expected return, honestly

Weight those five scenarios at 15%, 35%, 35%, 12%, and 3%. Those are my own odds, not a model's, reflecting a business with real pricing power and a real beat-and-raise quarter, discounted for the fact that nobody gets a 27x-forward-earnings monopoly stock for free.

Do that math and the expected return is roughly 17% a year, including the dividend, over the next three years.

That beats the Sunbelt math I ran two weeks ago by a wide margin, and it should — this is a business that just posted record results and raised guidance, not one that just cut its growth range. The shape of the distribution is also better: the downside scenarios require things that have not happened yet (a real capex bust, a geopolitical shock), while the upside scenarios require the current trend — sold-out capacity, hyperscaler capex — to simply continue.

What I am actually doing

Weekly RSI sits in the high-50s to low-60s, nowhere near the oversold reading of 30 I usually wait for before sizing into a position. If you have followed the Sunbelt call, you know I do not normally buy without that technical confirmation.

This is the exception, and I want to be precise about why. The RSI rule exists to stop you from buying a falling knife on a business that is genuinely deteriorating — Sunbelt's rule was different because its own guidance was getting worse, so waiting for oversold made sense. TSMC's business did not get worse today. It printed the best quarter in its history and raised the outlook. A 4.6% drop on a beat-and-raise, on a day the whole sector was being sold on positioning rather than fundamentals, is not a technical setup. It is a mispriced entry point on a company whose numbers are improving faster than its stock price today reflects.

So I am adding to the Compounding Portfolio at $400, not waiting for a technical signal that this business hasn't earned. If the panic deepens and weekly RSI does eventually break 30, that is simply a better price on the same thesis, and I will add again.

TSM also goes on the watchlist for anyone building out their own Compounding Portfolio, next to the same AI-infrastructure names worth tracking through this cycle: NVDA, AMD, AVGO, ASML, MRVL, MU.

Good business. Real moat. The panic is the catalyst, not the warning.

You can run this same five-scenario test on anything sitting in your own watchlist before the market prices it for you. That is exactly what I built the Compounding Portfolio system inside the Sprint Club to do: https://www.strategysprints.com — and I will run it again the next time panic misprices a name worth owning.

Happy hunting. Simon & The Sprinters 🐬⚡️🐆