On July 16, 2026, TSMC — the company that manufactures the most advanced chips for nearly every major AI designer — reported the best quarter in its history. Second-quarter 2026 revenue reached a record NT$1,270.38 billion (US$40.20 billion), up 36.0% from a year earlier, and net income climbed to NT$706.56 billion, up 77.4% year over year [source: TSMC, 2026]. Two-thirds of that revenue traced to a single category: high-performance computing (HPC), the platform that includes AI accelerators, accounted for 66% of quarterly sales [source: TSMC, 2026].
One company's results rarely define an industry. But when the world's leading contract chipmaker posts a record quarter, ties two-thirds of it to AI-linked computing, and raises its 2026 capital spending to as much as US$64 billion, the numbers describe something larger than one balance sheet. They describe a semiconductor "supercycle" pulled by artificial intelligence. This article looks at what that boom is actually built from — leading-edge foundry manufacturing, high-bandwidth memory (HBM), and a capital-spending race — and, on the other side of the ledger, at whether the demand underneath it can last.
A word on method before the numbers. Throughout this piece we try to keep three kinds of statements apart: what has been measured (audited quarterly results), what has been projected (company guidance and management outlooks), and what has been reported (market-tracker estimates and press accounts that have not been independently verified). The supercycle looks different depending on which of those you are reading, and most of the confusion around it comes from mixing them.
Table of Contents
- The evidence: what TSMC's numbers show
- Why the foundry sits at the center
- HBM: the memory pillar of the boom
- The capex race — and what US$64 billion signals
- Is the demand real? Two readings
- Conclusion — what to watch
The evidence: what TSMC's numbers show
The measured quarter
Start with the measured facts, because they are unusually clear. In the second quarter of 2026, TSMC's revenue rose 36.0% year over year and 12.0% from the prior quarter to US$40.20 billion, while net income grew 77.4% year over year — and 23.4% from the prior quarter — to NT$706.56 billion [source: TSMC, 2026]. Diluted earnings per share reached NT$27.25, equal to US$4.31 per American depositary receipt, also up 77.4% from a year earlier [source: TSMC, 2026]. Profitability set records across the board: gross margin of 67.7%, operating margin of 60.3%, and net margin of 55.6% were all all-time highs [source: TSMC, 2026]. That net margin implies net income of roughly US$22 billion for a single three-month period — a figure derived directly from TSMC's own reported numbers, not an outside estimate.
It is worth pausing on what those margins mean, because they are the part of the report that is hardest to fake. A gross margin near 68% says that for every dollar of chips TSMC sold, only about 32 cents went to the cost of making them. Companies reach that kind of profitability only when they can charge a premium the market has no easy way to refuse — when customers need what they are buying and have few places else to buy it. Record revenue can be bought with price cuts; record revenue and record margins together are a different signal. They say demand was strong enough that the seller set the terms.
A business now led by AI computing
The composition of that revenue matters as much as its size. HPC made up 66% of sales and grew 20% quarter over quarter, while smartphones — for years TSMC's largest segment — slipped to 22% and declined 4% [source: TSMC, 2026]. The mix has tilted: data-center and AI computing, not phones, is now the primary engine of the business. That shift is the single clearest signal in the report that this cycle is different from past ones.
Consider what the two directions mean together. In the same three months, the segment TSMC had long leaned on — smartphones — contracted, while the computing segment expanded by a fifth [source: TSMC, 2026]. A shift that large and that one-directional in a single quarter is hard to read as a seasonal wobble; it is a change in what the largest advanced-chip maker in the world is mostly building for. The center of gravity has moved from the phone to the data center.
One caveat that keeps the reading honest
One caveat keeps the reading honest. "HPC" is TSMC's own platform category, and it bundles AI accelerators together with conventional CPUs and networking chips. It is a strong proxy for AI demand, not a pure "AI revenue" line item. So the correct statement is not "66% of TSMC's revenue is AI," but "66% came from the platform where AI accelerators sit, and that platform is growing fastest." The distinction is small, but it is the difference between a measured fact and an overstatement — and it is the kind of distinction the rest of this article depends on.
The near-term guidance points the same way
TSMC's own outlook for the next quarter points in the same direction, though it belongs to a different layer of certainty. The company guided to third-quarter 2026 revenue of US$44.6-45.8 billion, with gross margin of 65-67% and operating margin of 56-58% [source: TSMC, 2026]. The low end of that revenue range already sits above the US$40.20 billion the company had just recorded, so the near-term expectation is not a plateau but continued sequential growth, and the guided margins step down only modestly from the second quarter's records. Two things are worth holding onto here. First, guidance is a projection, not a measured result: it is management's expectation, and it can move. Second, even as a projection it is unusually confident — a company does not guide to another record-territory quarter unless its order book already points that way. Keep the label "projection" attached, and the guidance still tells you the near-term picture the company itself paints is one of a business still climbing.
Why the foundry sits at the center
What a foundry actually does
To see why one company's earnings carry industry-wide weight, it helps to define what a foundry does. TSMC is a contract manufacturer: it does not design chips of its own but fabricates the designs of "fabless" companies — the firms behind most AI accelerators, graphics processors, and smartphone chips. A fabless company draws the blueprint; the foundry turns silicon into the finished part. When those designers want the fastest, most power-efficient silicon, they compete for space on the same leading-edge production lines.
That division of labor is why a single manufacturer's results can speak for many brands at once. The logos on AI accelerators differ, but a large share of the most advanced ones are printed in the same fabs. When you read TSMC's quarter, you are reading, indirectly, the order books of the designers who depend on it — which is exactly why a contract manufacturer's earnings became a barometer for the entire AI hardware build-out.
Where the boom concentrates: the leading edge
Those leading-edge lines are where the boom concentrates. In the second quarter, 7-nanometer-and-below processes accounted for 77% of TSMC's wafer revenue, with 5-nanometer at 33%, 3-nanometer at 30%, and the newest 2-nanometer node at 3% [source: TSMC, 2026]. "Node" here is shorthand for a generation of manufacturing process; the newer, more advanced generations are the ones AI accelerators are built on almost exclusively. So the platform that makes them is also the platform in shortest supply. The bottleneck of the AI boom is not ideas or capital; it is advanced manufacturing capacity.
The concentration matters because advanced capacity cannot be conjured on demand. A leading-edge fab is an enormous fixed investment, and bringing a new one online takes years, not months. That is what makes the leading edge a genuine chokepoint: when demand for the newest nodes jumps in a single quarter, supply cannot answer in the same quarter. The queue simply lengthens.
Why one supplier can stand in for an industry
That scarcity is why a single supplier's numbers can stand in for an industry. At the leading edge, alternatives are few, and TSMC's own management has said plainly that demand is running ahead of what it can currently produce — the reason it is raising spending rather than sitting on record profits [source: TSMC, 2026]. When the company at the center of the supply chain cannot keep up, the pressure ripples outward to everyone who needs advanced chips.
The pressure is visible from the customer side too, not just the supplier's. In press reports, at least one major chip designer — Broadcom — has said publicly that foundry capacity constraints are limiting how many AI chips can be supplied. That is a reported account rather than an audited disclosure, so it should be read as one company's characterization, not a verified market figure. But it is a telling one: it is comparatively rare for a buyer, rather than a seller, to be the party describing the ceiling. When the constraint is loud enough that customers name it out loud, the shortage is not a rounding error in a forecast; it is the shape of the market.
HBM: the memory pillar of the boom
Why memory is the second pillar
Foundry capacity is only half the story. An AI accelerator is useless without somewhere to move data fast enough to feed it, and that job falls to memory. Modern AI accelerators need enormous memory bandwidth — the rate at which data can be shuttled in and out of the processor — and ordinary memory cannot supply it. AI chips therefore rely on high-bandwidth memory (HBM), a design that stacks DRAM chips vertically and places them right beside the processor, so that far more data can flow to the chip than a conventional memory layout allows. HBM has become the second indispensable pillar of the supercycle.
The reason the two pillars rise together is that they are complements, not substitutes. A faster accelerator with too little memory bandwidth starves; abundant bandwidth with too little compute sits idle. Every advanced AI processor coming off the leading-edge lines needs its matching stack of HBM, which is why a boom in foundry output and a boom in high-end memory are, in practice, the same boom seen from two sides.
The evidence: SK Hynix's record quarter
The clearest evidence sits in the memory makers' own results. SK Hynix, the largest HBM supplier, reported first-quarter 2026 revenue of KRW 52.58 trillion (about US$35.5 billion), up 198% year over year — its first quarter ever above KRW 50 trillion — with operating profit of KRW 37.61 trillion (up 405%) and net profit of KRW 40.35 trillion (up 398%) [source: SK Hynix, 2026]. The company attributed the surge to demand for high-value products including HBM, high-capacity server DRAM, and enterprise SSDs [source: SK Hynix, 2026].
Those growth rates deserve a second look, because they are of a different order than the foundry side. Revenue nearly tripling and operating profit multiplying roughly five-fold in a single year is not the arithmetic of a steady components business; it is the arithmetic of a product that went from useful to indispensable almost at once. Semiconductors are a cyclical industry, and a year like this is what a sharp up-swing looks like when a scarce new product — high-end HBM — lands in the middle of it. The scale of the jump is itself a measure of how suddenly the demand arrived.
Reading the market structure carefully
The market's structure deserves a careful note, because here the data quality changes. Market trackers estimate SK Hynix holds roughly 56-57% of the HBM market, ahead of Samsung and Micron — but that is an outside estimate, not an audited company disclosure, so it should be read as directional rather than exact. Samsung, for its part, has been reported to have passed Nvidia's qualification for its 12-layer HBM3E memory and to have sold out its 2026 HBM capacity [source: KED Global, 2026]. Those reports come from the press and market researchers rather than audited filings, so the honest framing is "reported and widely expected," not "independently verified."
The distinction is not pedantry; it changes what you are entitled to conclude. A company's own audited revenue is a fact you can build on. A market-share estimate and a "sold-out" report are useful signals about direction, but they carry an asterisk, and stacking conclusions on them as if they were audited numbers is how confident narratives outrun their evidence. Read this way, the memory story is strong on the parts that are measured — SK Hynix's revenue and profit — and merely indicative on the parts that are estimated.
Concentration is also a risk
What the memory side confirms is the same message the foundry side sends: demand is outstripping supply, and it is showing up in record revenue rather than in projections. But the concentration also plants a risk. A market this dependent on a few suppliers and a few buyers is sensitive to any one of them changing course — a single large customer trimming orders, or a rival's capacity coming online, can move prices in a way that a broad, fragmented market never could. The same narrowness that makes the boom so visible in a handful of income statements is what makes it fragile.
The capex race — and what US$64 billion signals
A spending number, not a profit number
The most forward-looking number in TSMC's quarter was not a profit figure but a spending one. The company raised its 2026 capital-expenditure guidance to US$60-64 billion — a midpoint near US$62 billion — up from an earlier US$52-56 billion and more than 50% above its roughly US$40.9 billion of actual spending in 2025 [source: TSMC, 2026]. It also lifted its full-year revenue growth outlook to "slightly above 40%" in US-dollar terms, from an earlier "above 30%," and announced an additional US$100 billion investment in Arizona, bringing its total US commitment to about US$265 billion [source: TSMC, 2026].
Capital expenditure is the closest thing a company has to a signed statement about the future. Operating profit reflects demand that has already arrived; capex reflects demand a company expects to arrive, because the plants being funded now will not produce revenue for years. Raising that budget by more than half in a single year, and raising the revenue outlook alongside it, is management telling the market it does not expect this to be a one-quarter spike.
Capital spending as conviction
Chief executive C.C. Wei described AI-related demand as "extremely robust" and said it reinforces the broader AI investment cycle, expecting demand to stay strong through the end of the decade [source: TSMC, 2026]. Building fabrication plants takes years, so a capex surge of this size is a bet that today's demand is not a spike but a plateau — that the orders will still be there when the new capacity comes online in 2027 and beyond. The Arizona commitment underlines the point: an additional US$100 billion, on top of prior pledges, is not the kind of money a company spends to serve a demand it thinks is temporary.
It is worth labeling that outlook precisely. "Extremely robust" and "strong through the end of the decade" are management's characterization and projection — official, and coming from the company with the best view of its own order book, but forward-looking statements rather than audited facts. They belong on the projection layer, next to the guidance, not on the measured layer next to the quarterly results.
The same spending is the clearest risk
Yet the same spending that signals confidence is also the clearest source of risk. Fabs are enormous fixed investments that are hard to unwind; capacity built for a boom becomes a burden if demand cools before it is filled. Semiconductor history is a history of cycles, and heavy capital commitments made near the top have, in past cycles, turned into oversupply on the way down. That is not a prediction — it is the reason the capex figure cuts both ways.
The mechanism is worth spelling out, because it is what links a confident spending plan to a future glut. A fab authorized in 2026 to meet 2026's demand will not reach full output until later, by which point demand may have moved. If it has grown, the new capacity is filled and the bet pays off; if it has stalled, the plant still has to be paid for, and its fixed costs press on margins precisely when revenue is softest. The very scale that makes TSMC's spending a vote of confidence is what would make a miss expensive. Confidence and exposure, in this business, are the same line item.
Is the demand real? Two readings
The bullish reading
Because the stakes are high, the sustainability question deserves both sides fairly stated. The bullish reading rests on measured facts: TSMC and SK Hynix both report demand exceeding supply, they are raising spending and selling out capacity, and forecasters broadly expect strong semiconductor growth in 2026 [source: TSMC, 2026][source: SK Hynix, 2026]. In this view, agentic AI, rising compute per query, and continued data-center buildouts make the current pace a floor rather than a ceiling.
The strongest point in the bullish column is that its core evidence has already happened. The record quarters are not forecasts of demand; they are demand that has been booked, billed, and audited. When two of the industry's pivotal suppliers both say, in the same season, that they cannot make enough, and both back the claim with record income statements and larger budgets, the burden of proof shifts. The bull case does not need to predict a boom — it can point to one already on the books and ask what, specifically, is expected to stop it.
The cautious reading
The cautious reading does not deny the records but questions how much of the demand is durable. Industry analysts warn of "double-booking" — customers placing overlapping orders with several suppliers to secure scarce capacity, which can inflate apparent demand — and note that if AI capital spending slows, an inventory correction could follow. Forecasts for 2026 semiconductor growth span a wide range, from roughly 12% to above 40%, and some scenarios warn of an outright downturn; the gap itself signals how uncertain the outlook is. A recurring question is when AI's revenue will catch up with the scale of investment now flowing into it.
Double-booking is the sharpest of these worries because it attacks the bull case at its strongest point — the order book itself. If a customer, unsure it can get enough chips, places the same order with two suppliers intending to keep whichever arrives first, then some of the "demand" both suppliers report is the same underlying need counted twice. Nobody outside the buyers knows how much of the current backlog is real versus defensive, which is exactly why the practice worries analysts: it would not show up until the shortage eased and the duplicate orders were cancelled — the point at which a boom can turn into a correction. Set that against the open question of AI monetization — whether the revenue AI products generate will grow fast enough to justify the capital now being poured into serving them — and the cautious case is not a bubble call so much as a demand for proof that today's orders are tomorrow's, too.
Keeping the layers separate
The two readings are easier to hold together if the layers are kept separate. The record quarters are measured facts. The 40%-plus growth outlook and the "robust through 2030" comment are conditional projections — official, but forward-looking and subject to change. The bubble-versus-boom debate is largely about sentiment and interpretation. Conflating them is how a strong quarter gets mistaken for a guaranteed future, or a legitimate caution gets mistaken for a crash call. The disciplined position is to treat each layer as what it is.
Done honestly, that discipline dissolves much of the apparent contradiction between the two camps. The bulls are mostly right about the measured layer: the quarters are real and the shortage is real. The cautious side is mostly right about the projection layer: the outlooks are confident but unproven, and the forecast range is wide enough to hold very different futures. The two are not so much disagreeing about the same fact as emphasizing different layers of the same picture. Seen that way, "record boom" and "watch for a correction" are not opposites — they are the measured present and the uncertain future, stated at the same time.
Conclusion — what to watch
The supercycle is not a slogan; it is visible in audited numbers. TSMC posted the best quarter in its history with AI-linked computing at its core, SK Hynix crossed KRW 50 trillion in a single quarter, and the whole structure rests on three pillars: leading-edge foundry capacity, high-bandwidth memory, and a capital-spending surge [source: TSMC, 2026][source: SK Hynix, 2026]. What the numbers cannot tell us is whether the present, measured so precisely, will hold.
So the things to watch are concrete. First, whether AI's end demand and monetization keep pace with the capital now being spent to serve it. Second, how the wave of foundry and HBM capacity coming online in 2026 and 2027 lands — meeting demand, or overshooting it. Third, whether any double-ordering unwinds and how memory prices move when it does. Fourth, whether Samsung and Micron's HBM ramps erode SK Hynix's lead and, with it, pricing power. None of these will be settled by a single quarter. What this moment calls for is not a verdict but the discipline to watch the results, the capacity, and the demand together, at the same eye level — crediting the measured present without borrowing certainty from it for the future.