Weekly Chart Pick · 2026-09-11
This company makes every AI chip that matters and 73% of the world’s outsourced chips — sales are up 53% in a month, and the stock is coiled in a triangle whose apex is this week
TSMC is the factory behind Nvidia, Apple, AMD, Broadcom and MediaTek: 73% of the pure-foundry market, more than nine-tenths of the leading edge, a 2-nanometre node in volume that no one else can sell at scale, and advanced packaging booked out for more than a year. August revenue rose 53% from a year ago, the first month ever above NT$500 billion; the company guides 2026 growth above 40%. The stock has spent ten weeks digesting a 22% drop from its 30 June record — a fall that followed the best quarter in its history — and has built an ascending triangle whose descending line off the top and rising line off the January low meet this week at about $435, against a pivot cluster our support scanner puts at $432.46. On Friday afternoon it trades at $433.11 with the stochastic at 80 and rising. Twelve-month view: higher, in our view, with the 5-year channel’s upper boundary near $510–530 as the first measured objective; invalidation for the setup is a close below $408.
- The frame is a strategic position, not a trade: TSMC is the only company on earth that manufactures leading-edge logic at scale for other people. 73% of the pure-foundry market in the second quarter, against 7% for Samsung, the next largest; nodes of 7 nanometres and below are 77% of wafer revenue; the 2-nanometre node has been in volume since late 2025 with all 2026 capacity booked, more than half of it by Apple, and output heading to 100,000 wafers a month by year-end. At the leading edge there is no second source.
- The growth is not a forecast, it is in the monthly numbers: August revenue NT$514.8 billion, up 53.3% from a year earlier and 10.1% from July, the first month above NT$500 billion; January to August +39.3%. Second-quarter revenue $40.2 billion, up 36%, net income up 77%, gross margin 67.7%, operating margin 60.3%. Guide for the third quarter $44.6–45.8 billion and for the year ‘slightly above 40%’ growth; capital spending $60–64 billion.
- The chart is a coil under a record. From the 30 June high at $477.57 the stock fell 21.5% to $374.67 on 29 July, then built higher lows — $410 on 24 August, $414 on 1 September — under a descending line off the top. That line, the horizontal at about $435 that has capped every rally since August, and the rising line from the January low converge this week. Price trades at $433.11 on Friday afternoon, on the apex, with our support scanner’s pivot cluster at $432.46. A coil resolves; this one resolves within days.
- The oscillator is high and rising, not low and turning: slow stochastic 80.0 against 76.3, K above D, in the upper band. That is the profile of a breakout, not a bottom — breakouts from ascending triangles usually happen with the oscillator already elevated — and it is also the profile of a failure if the line holds. The stochastic pinned above 80 while price sits under resistance is the one number that decides which of the two this is.
- The count we carry: the advance from the October 2022 low near $58 has run three waves — a first to mid-2024, a second to the April 2025 low near $135, a third to the June top at $477 — and the July drop is a shallow fourth, retracing about 30% of the third. A fifth wave would carry the stock to the upper boundary of the 5-year channel, near $510–530 today and rising toward roughly $600 over twelve months. The alternative is that June was the end of the cycle and this is a corrective bounce; a close below the July low at $374.67 is where that alternative wins.
- The risk is the one every reader already knows and the two that most do not. Taiwan is a single point of failure in a year in which a Gulf war has put oil at $107 — geopolitics is not abstract in 2026. Less discussed: the market sold TSMC’s record second quarter because operating margin was guided from 60% to 57% on 2-nanometre and Arizona ramp costs, and our print record shows the tape has paid only two of ten double beats. The stock is graded on capex and margin, not on the beat. A close below $408 ends the setup; below $374.67 the count.
TSM Taiwan Semiconductor Manufacturing (TSMC) last close 428.03 (2026-09-10)
The Setup
The claim: the most strategically important manufacturing company in the world has corrected 22% on a margin guide while its sales accelerated to +53%, and the correction has taken the shape that resolves upward more often than not — an ascending triangle under a record high, at its apex, with the oscillator elevated.
Start with why the stock fell, because it is the whole setup. On 16 July TSMC reported the best quarter it has ever had and guided operating margin from 60% to 57% for the next one, on the 2-nanometre ramp and the cost of building in Arizona. The market read ‘revenue growing at twice the pace of profit’ and sold the ADR from $477 to $375 in a month. That is a valuation adjustment for lower incremental margins, not a change in the demand picture — and the demand picture has since been confirmed monthly: July +45%, August +53%. A stock that falls on margin and then watches revenue accelerate for two months is a stock whose correction has run out of reasons.
The chart says the same thing in its own language. Since the 29 July low the stock has made three higher lows — $374.67, $410.12, $414.00 — under a descending line off the June top that now sits at about $435, coincident with the horizontal that has capped every rally since mid-August and with the $432.46 pivot cluster our support scanner identifies. Three lines meeting at one price is a coil, and a coil after a 22% correction inside a five-year uptrend is the pattern that has preceded each of this stock’s last three legs higher. The stochastic at 80 and rising says the buyers are already in; the question is only whether the line gives.
Above the line, the objectives are measured, not imagined. The June high at $477.57 is 10% up. The upper boundary of the five-year channel sits near $510–530 today and rises with time; a fifth wave from the July low that matches the character of the earlier legs reaches it inside the twelve-month window, and the channel line itself points toward roughly $600 a year out. Below the line, the setup has two exits, and we take both seriously: $408, where the scanner’s invalidation sits and the higher-low sequence breaks; and $374.67, where the fourth-wave reading fails and the alternative — that June was the cycle top — takes over.
Why this name and not one of its customers: TSMC is the one company in the AI build-out whose revenue is agnostic to who wins. If Nvidia’s Rubin dominates, TSMC makes it. If Broadcom’s custom accelerators take share, TSMC makes those. If Apple’s on-device AI drives an iPhone cycle, TSMC makes the A20. It is the toll road every design must cross, at 68% gross margin, with the packaging capacity that nobody else has. That is what Thomas means by a strategic position like Nvidia’s: not a bet on a product, but on a chokepoint. The correction has priced the cost of widening the road; the coil is the market deciding whether it has priced enough.
In the Closelook readings
Said plainly: TSMC is a constituent of the Rubin Build-Out 100 in the Foundry & Integration sector, Fabrication layer — the layer that fell 1.9% on Thursday when all eight Rubin layers were red and the Foundry sector −2.6%, before Friday’s rebound. Our fundamentals scoring engine ranks it fifth of 383 scored names on the semis-cyclical weight set with a composite of 71 (quality 86, growth 72, momentum 66) and a forward multiple our snapshot carries at under 20 times against trailing 32. The pattern engine reads it at 48.96 on the directional-flow scale, ‘decelerating-up’, heat neutral, with the 21-day delta at −8.3 — the flow cooled through the July drop and has not yet re-accelerated. Our support-confluence scanner has one of six filters engaged at the $432.46 pivot cluster and declines to give odds: ten comparable episodes against a gate of twelve, so the machine says ‘not enough history’ rather than pretending. The print-record card carries nine EPS beats and ten revenue beats in ten prints against a 2-5-3 split — two paid, five flat, three sold — the most under-rewarded record of any large chip name we track; the next print is in mid-October.
52-week range (closes): 258.91 – 477.57
The latest print
The second quarter, reported 16 July, was the best in the company’s history and the stock fell 5% on it. Revenue NT$1,270 billion, $40.2 billion, up 36% from a year earlier and 12% from the first quarter, at the top of guidance; net income NT$707 billion, up 77.4%; earnings $4.31 per ADR against $3.87 expected; gross margin 67.7% against a guide of 66–68%; operating margin 60.3%; net margin 55.6%. High-performance computing — the AI accelerators, CPUs and networking chips — was 66% of revenue, up 20% quarter on quarter. Advanced nodes were 77% of wafer revenue, and the 2-nanometre node contributed 3% for the first time.
What the market sold was the guide, not the quarter. Third-quarter revenue $44.6–45.8 billion, 37% growth at the midpoint; gross margin 65–67%, down about 1.7 points on the 2-nanometre ramp; operating margin 56–58% against 60.3% just reported. Capital spending for 2026 raised to $60–64 billion, and a further $100 billion committed to Arizona under the US–Taiwan tariff framework that lifts the American programme toward $465 billion and eleven fabs. Management said overseas fabs dilute gross margin by two to three points in their early years. Revenue growing at twice the pace of profit was the sentence that sent the ADR from $477 to $375 in a month.
Since then the monthly sales have said the demand is real. July revenue rose 45% from a year earlier; August NT$514.8 billion, up 53.3% on the year and 10.1% on the month, the first month above NT$500 billion and the fourth straight monthly increase. January to August is NT$3,387 billion, up 39.3%, which means the ‘slightly above 40%’ full-year guide is a description, not an ambition. The third-quarter print lands in mid-October — our print record shows the stock has beaten on both lines in nine of ten quarters and been paid for it twice.
The fundamentals
What the business is. TSMC manufactures chips designed by others and owns none of the designs: Nvidia’s Blackwell and Rubin accelerators, Apple’s A- and M-series, AMD’s Instinct and Epyc, Broadcom’s custom accelerators for Google and Meta, MediaTek’s and Qualcomm’s modems and phone chips. It is paid per wafer and per package, so its revenue is the sum of the AI build-out’s unit volume times the price of the most advanced silicon on earth — and it sets that price, because there is no alternative supplier at the leading edge.
The monopoly, in numbers. 73% of the pure-foundry market in the second quarter of 2026, a share it has held or grown for six consecutive quarters; Samsung Foundry is second at 7%, still working on yields for its 2-nanometre process; Intel’s foundry has no marquee external customer at volume. At 3 nanometres and below the share is above 90%. The 2-nanometre node entered volume production in late 2025; all 2026 capacity is booked, more than half by Apple for the iPhone 18 Pro chips that ship next week, the rest by AMD, Nvidia, MediaTek and Qualcomm. Output is planned to rise from 50–60 thousand wafers a month in the first half to 100 thousand by year-end, alongside 180 thousand at 3 nanometres.
The second moat: packaging. CoWoS, the advanced packaging that stacks memory beside the accelerator, is the physical bottleneck of the AI build-out. Both families were reported fully booked into 2027 with lead times of a year or more; capacity is being expanded toward 130–150 thousand wafers a month by late 2026, and Nvidia is reported to hold about 60% of it for the Rubin generation. A customer cannot switch foundry without also finding packaging, and there is no one else with the packaging.
The financial shape. Gross margin 67.7% and operating margin 60.3% on $40 billion of quarterly revenue are the margins of a monopoly with pricing power, not of a contract manufacturer. Capital spending of $60–64 billion a year is funded from operating cash flow with a quarterly dividend of NT$7 a share still rising. Our scoring snapshot carries the ADR at about 32 times trailing and under 20 times forward earnings, with a PEG below 0.8 — the market is paying a lower multiple for 40% growth at TSMC than for most of its customers. The costs are the overseas fabs — two to three points of margin dilution in the early years — and the concentration: Nvidia and Apple together are a large share of revenue.
The chart, read by hand
Trend: Two channels and a triangle. The first channel rises from the October 2022 low near $58; its median line was the ceiling in 2023 and the floor from 2025, and its upper boundary — untouched since the channel began — sits near $510–530 today. Inside it, a steeper channel from the April 2025 low near $135 carried the 2025–26 advance; its upper boundary was tagged exactly at the 30 June record of $477.57, and the July drop returned price to the steep channel’s median. The one-year window shows the correction’s shape: a descending line from the June high through the 17 August high at $431, a horizontal at about $435 that has capped September, and a rising line from the January low near $275 through the 29 July low at $374.67 and the 24 August low at $410. The three lines meet this week at $433–435 — where the stock trades on Friday afternoon. The 50-day average is at $419, rising, and the stock reclaimed it on 4 September; the 200-day is at $375, 14% below price and rising.
Oscillators: Slow stochastic 80.0 against 76.3 on Friday afternoon — K above D, both in the upper band and rising. The last two turns up from the low band, in late July and late August, each produced a rally of 10% or more; this is not one of those. It is an oscillator already high while price sits on resistance, which is how ascending triangles look in the sessions before they break — and how they look when they fail. RSI is at 54, neutral, with room in either direction. The tell is the stochastic holding above 80 as price closes above $435: that is a breakout with momentum. A K-under-D cross from here with price still under the line is the failure signal and points back toward the $410–419 zone.
Elliott Wave count: Our count, offered for correction. The cycle from the October 2022 low near $58 reads as an impulse: wave 1 to the mid-2024 high near $190, wave 2 to the April 2025 low near $135 (a deep second, about 30% in price), wave 3 to the 30 June 2026 record at $477.57 — an extended third, three and a half times the first in price terms — and wave 4 the July drop to $374.67, a shallow fourth retracing about 30% of the third, alternating in character with the deep second as the guideline prefers. On that reading wave 5 is either underway from the July low or begins on the break of $435; its first objective is the channel’s upper boundary near $510–530, with the line rising toward $600 over twelve months. The alternative count is that June completed the impulse and the July–September action is wave A and B of a larger correction; it becomes the primary count on a close below $374.67. We do not dress this up as a formal labelling — the fourth wave is short in time for a third of that size — but the levels the two counts disagree at are the same levels the trend lines give, which is why we use them.
Key levels: $477.57 — the 30 June record and the steep channel’s upper boundary; the first objective above the coil · $510–530 — the five-year channel’s upper boundary today, rising toward roughly $600 over twelve months · $439–440 — the 8 September high close and the breakout confirmation above the descending line and the $435 horizontal · $432.46 — the support scanner’s pivot cluster; Friday afternoon’s $433.11 sits on it · $419 — the rising 50-day average, reclaimed on 4 September · $408 — the scanner’s invalidation and the level where the higher-low sequence breaks · $374.67 — the 29 July low; the count’s invalidation and the alternative’s confirmation
What the machines say
Technical snapshot (as of 2026-09-10) — vs 20d +3.3% · vs 50d +3.6% · vs 200d +16.4% · off 52w-high -9.1% · stochastic K 86.7 / D 72.3 · Reclaimed its 50-day
Pattern engine (scan 2026-09-10) — directional-flow 49/100, neutral, decelerating up (5d -1.5%, 21d -8.3%)
Closelook scores (as of 2026-09-10) — composite 71/100 · momentum 66/100
Machine blocks render only where our engines actually cover the name — nothing is padded.
The risks, equal billing
The risk everyone names, and it is real. About 90% of the world’s most advanced logic is made on one island, and 2026 has shown what a regional war does to a commodity everyone depends on — oil settled at $107 on Thursday because of the Gulf. Taiwan is a single point of failure for the entire AI build-out and for this stock; the Arizona programme, now $465 billion and eleven fabs under the tariff framework, reduces that risk over a decade, not over the twelve months we write about. No chart level protects against it.
The risk the market actually trades. The July drop was not about demand; it was about margin. Operating margin guided from 60.3% to 56–58%, gross margin down 1.7 points on the 2-nanometre ramp, two to three points of dilution from overseas fabs, capital spending at $60–64 billion. Every quarter from here carries the same test: revenue at +40% is known, the question is how much of it reaches the operating line. Our print record says the tape has paid only two of ten double beats — five flat, three sold, including a 4.5% fall after the January print. A record quarter in mid-October could be sold again if the margin guide slips.
The chart risk is symmetric. An ascending triangle at its apex either breaks out or breaks down within days, and the oscillator at 80 under resistance means the failure would start from a high, not from a low. Our pattern engine reads the name as ‘decelerating-up’ with a 21-day delta of −8.3, and the support scanner refuses odds for lack of history — the machines are not confirming the chart, they are neutral on it. A close below $408 ends the setup; a close below $374.67 ends the count. We would write both.
Concentration and currency. Nvidia and Apple are a large share of revenue and hold most of the 2-nanometre and CoWoS capacity between them; a pause in either’s build-out would show up in monthly sales within a quarter. The ADR is priced in dollars for a company that reports in New Taiwan dollars, and a stronger Taiwan dollar is a margin headwind management has flagged before. And the Fed: a hike on Wednesday is 82% priced, and this is a $2.1 trillion growth stock that has just spent four days falling with the long bond.
Our forecasts
12 months from now: Higher, in our view. Demand is supply-constrained and visible in monthly revenue running 40–50% above last year; the leading edge has no second source and the packaging bottleneck is booked into 2027; the stock trades under 20 times forward earnings for that growth; and the ten-week correction has resolved into an ascending triangle at its apex inside an intact five-year channel. The measured objectives are the June record at $477.57, then the channel’s upper boundary near $510–530 and rising toward $600 over the window. A diary view, not advice.
The next few weeks: The coil resolves. A weekly close above $440 — through the descending line, the $435 horizontal and the 8 September high close of $439 — confirms the breakout and puts the June high in play into the mid-October print. A close below $408 breaks the higher-low sequence and voids the setup; the stock would then be trading the alternative count toward the $374.67 July low, and we would say so.
Diary views, not advice — held publicly alongside the scored record.
Why publish this name? Because our Directional Flow readings are what let us say it: stocks with this or a similar setup have often — historically — done well over the following twelve months. That is the engine's backtested observation about the past, not a forecast; past performance may not be an indication of future performance. Which is exactly why every pick is scored publicly against the tape from today, at one week, one month, three, six and twelve. Twelve months. A diary view, not advice, and not an entry price. Thomas’s charts lead; the count and levels are ours and are open to correction.
Closelook publishes an investment research diary, not investment advice. Every pick is scored publicly against the tape at 1w/1m/3m/6m/12m from publication — the record, not the single call, is the product.
Open the live TSM terminal → All picks & the running record →
A note from the editor: the Weekly Chart Pick is free. Its price is one click — add Closelook as a preferred source on Google, and you support this work in Google Discover and Search every time you read.
Google opened in a new tab. Add Closelooknet there, then come back to this tab — the Chart Pick unlocks the moment you return.
The preferred-source setting lives in your Google account — if Google asks you to sign in first, you will land on the selection page right after. Google does not tell us whether you saved the selection. Clicking “Done” unlocks the Weekly Chart Pick on this device for 12 months.