Bottom line — Two-thirds of TSMC’s revenue in a single quarter now comes from AI chips (HPC). In Q2 2026, the smartphone — long the face of this company — fell to 22%, and data-center chips took its place. This structural shift is the firmest evidence yet that ‘the AI cycle is real,’ and at the same time the invoice for a concentration risk in which earnings lean on a single source of demand. This piece breaks down the numbers in TSMC’s Q2 results, released July 16, and sums up what that signal means for Korean memory, which is tied to it through HBM.
1. The quarter the smartphone was pushed aside — start with the numbers
TSMC’s Q2 2026 revenue was $40.2 billion, up 36% year over year and above the top of the company’s own guidance. Net income hit a record quarterly high of TWD 706.56 billion (about $22 billion), a 77% surge from a year earlier. Double-digit profit growth has now continued for nine straight quarters.
By the numbers alone, you could file it under ‘another good quarter.’ But the real event this quarter lies in the composition of revenue. The high-performance computing (HPC) segment — the AI accelerators for cloud data centers — jumped 20% in a single quarter to take 66% of total wafer revenue. Meanwhile the smartphone, which was this company’s largest revenue source as recently as 2022, sank to 22%. Honestly, when I first saw the results table, the biggest surprise was how far the smartphone share had been pushed down.
For a sense of scale, $40.2 billion is one quarter’s revenue — three months, not a year. Nine straight quarters of double-digit profit growth reads as a signal that this company is in the middle of a structural upswing measured in years, not a one-off boom in a single quarter. And the core of this release is that the engine driving that rise has been completely swapped out — from the smartphone to AI.
| Item | Q2 2026 | Guidance / YoY |
|---|---|---|
| Revenue | $40.2B | YoY +36%, above top of guidance |
| Net income | TWD 706.56B | YoY +77%, record quarterly high |
| Gross margin | 67.7% | Above guidance 65.5–67.5% |
| Operating margin | 60.3% | Above guidance 56.5–58.5% |
| Return on equity (ROE) | 45.9% | High-profit structure maintained |
Key TSMC metrics, Q2 2026 (Source: TSMC earnings release)
2. Two-thirds of revenue is HPC — the platform landscape has flipped
Broken down by platform, the change is visible at a glance. HPC rose 20% from the previous quarter to a 66% share. Smartphone fell 4% to 22%, the Internet of Things (IoT) rose 4% to 5%, and automotive rose 15% but its absolute share stayed at just 4%. The remaining 3% or so is digital consumer electronics and the like. It means two-thirds of one company’s revenue is effectively tied to a single theme: the ‘AI data center.’
| Platform | Share | Scale (relative to max) |
|---|---|---|
| HPC (AI / data center) | 66% | |
| Smartphone | 22% | |
| IoT | 5% | |
| Automotive | 4% | |
| Others (consumer electronics, etc.) | ~3% |
Bar width is relative size vs. the maximum (HPC 66%) (Source: TSMC Q2 earnings slides)
Just four years ago, the largest revenue source and today’s largest revenue source have completely swapped places. In the semiconductor industry, a compositional shift at this speed is uncommon. The catch is that this shift happened less because ‘TSMC did well’ and more because ‘the AI investment cycle piled up that much.’ Good news and uneasy news sit inside the same number.
Compared with past cycles, the difference is stark. Semiconductor demand in the smartphone era was broad demand created by the replacement cycles of hundreds of millions of people worldwide. Today’s HPC demand, by contrast, is narrow and deep — placed in bulk by a handful of large cloud operators and AI firms. When the demand base is broad, one customer dropping out spreads the shock; when it is narrow, earnings can swing on a single investment decision by a few big customers. That is why 66% is both a strength and a vulnerability.
2-1. Why 60%-plus margins are exceptional
The quietly frightening part of these results is the margins. The 67.7% gross margin exceeded even the top of the company’s guidance (67.5%), and the operating margin of 60.3% was well above guidance (56.5–58.5%). A profit margin this high on selling a single wafer reads as a sign that customers have effectively set down their pricing leverage.
The reason is simple: right now there is essentially nowhere but this company that can print AI chips on leading-edge processes (the 3nm / 2nm-prep lines). Whoever holds the bottleneck sets the price. That said, the company warned that in Q3, margins could be slightly pressured as early 2nm mass-production costs are reflected. It is a signal that margins will not stay at this height forever.
For investors, the margin is a window into the quality of earnings. If revenue grows but the margin collapses, the meaning of that growth shrinks; when revenue and margin rise together, that growth is read as coming from pricing power. This quarter’s 60%-plus operating margin is closer to the latter. Still, how this metric moves from Q3 on is likely to be the first number the market checks next.
3. The $64 billion capex bet — laying down supply in advance
What the market watched even more than earnings was the two outlooks the company raised alongside them. TSMC lifted its 2026 annual revenue-growth forecast to ‘more than 40%’ in dollar terms — a big jump from the roughly 30% previously. At the same time, it raised this year’s capital-expenditure plan by 15%, from the prior $52–56 billion to $60–64 billion. The largest scale ever.
| Outlook item | Prior | Raised |
|---|---|---|
| 2026 revenue growth (USD) | ~30% | More than 40% |
| 2026 capex | $52–56B | $60–64B |
Magnitude of the 2026 guidance upgrade (Source: TSMC earnings release / TrendForce)
Raising capex is an act of proving in money the company’s conviction that ‘demand will continue for the next several years.’ CFO Wendell Huang cited as the basis for the upgrade “customers’ structural and sustained strong demand, especially the emerging agentic-AI market.” He also reaffirmed a plan for an additional $100 billion investment in Arizona, USA — a signal to lay down supply in advance. Put the other way, this company’s earnings now hold only if AI data-center investment does not bend.
The capex upgrade sends ripples to suppliers too. As leading-edge lines expand, demand for the materials, components, and equipment that go into them rises with them, and Korean and Japanese equipment/materials firms are partly hooked in here. But such trickle-down effects have a lag between the order point and actual revenue recognition, so it is hard to conclude that this announcement translates immediately into related firms’ earnings. The direction is favorable, but the speed needs confirmation.
4. The signal coming to Korea — HBM and the memory cycle
TSMC’s numbers do not end as the story of one Taiwanese company. On a single AI accelerator, alongside the logic chip that TSMC prints, high-bandwidth memory (HBM) is invariably stacked next to it. And the main HBM suppliers are Korea’s SK hynix and Samsung Electronics. TSMC’s HPC revenue jumping connects directly to HBM demand jumping.
The actual indicators point the same way. According to the Ministry of Trade, Industry and Energy’s June export-import trends, June semiconductor exports reached the $44.8 billion range — a record-high level. The government’s explanation is that strength in high-value memory such as HBM overlapped with a rebound in DRAM fixed-transaction prices. Cumulative first-half semiconductor exports also already surpassed the prior annual record. The picture TSMC sees in foundry and the picture Korea sees in memory point in the same direction.
One thing must be distinguished, though. TSMC’s 60%-plus margin and the persistence of Korea’s memory cycle are different in character. Foundry is close to a bottleneck-holding oligopoly; memory is a market where prices adjust quickly as supply rises. Even riding the same AI demand, the two industries may reach cycle peaks in different ways.
Going a little deeper, this memory strength has two axes. One is HBM, directly tied to AI servers, where SK hynix is ahead and Samsung is chasing. The other is a rebound in general-purpose DRAM fixed-transaction prices, read as the combined effect of capacity concentrating on HBM production and thereby leaving ordinary DRAM supply relatively tight. That is, demand pulled up directly by AI, and prices that rebounded because supply tightened due to AI, are working at the same time.
Honestly, at first it felt strange to see TSMC’s results and Korea’s export indicators move in such the same rhythm. But recalling that Taiwan’s logic chip and Korea’s memory sit side by side inside one object — the AI accelerator — makes it a natural picture. The question is whether this synchronization holds only in the upswing, or moves together in a correction too.
4-1. The invoice for a country called Taiwan
These results also mesh with the picture of Taiwan’s whole economy. The share that TSMC alone holds in Taiwan’s market cap and exports has long been overwhelming. The more the AI cycle lifts this company, the stronger the structure in which the growth rate, currency, and export statistics of a country called Taiwan are swayed by one company’s quarterly results. A blessing in the upswing, but also the invoice for a concentration in which a national economy is tied to a single industry.
It is not an unfamiliar picture in Korea either. Korea, where semiconductors carry a large axis of exports, likewise sees the whole nation’s trade balance sway with this industry’s cycle. That is why we cannot brush off TSMC’s 66% as someone else’s affair. One industry’s super-boom can also create a statistical illusion — overall exports may look good, but opening them up reveals a lean on a specific item.
5. The voice on the other side — the “glut in three years” concern
Not everyone reads these numbers as pure optimism. Parts of the market warn that today’s fever of AI capital investment could tip into an oversupply phase within about three years. The logic: once the stretch in which data-center operators race to buy chips ends, the large capacity laid down in advance instead comes back as a burden.
TSMC itself said Q3 margins could be pressured by early 2nm mass-production costs. On top of that, how fast back-end bottlenecks such as CoWoS — used in AI-accelerator packaging — unwind is a variable. Honestly, I see these results less as evidence that ‘AI is strong’ and more as a signal that ‘demand has been dumped onto one place, TSMC.’ The stronger the number, the more one must ask back what premise props it up. Market consensus weights continued demand; the opposing side weights the possibility of a peak in the investment cycle. The judgment, in the end, rests with the reader.
One more point: a glut concern does not immediately mean collapse. If demand continues in a stepwise fashion after capex is laid down, the surplus is absorbed gently. Conversely, if demand cools abruptly at a certain point, the capacity laid down in advance returns as a burden all at once. So what to watch is not ‘is demand strong’ but ‘how long, and how smoothly, does demand continue.’
6. A ten-year horizon — 2nm and agentic AI
Seen over a long span, this quarter is a point where two long-term flows cross. One is process miniaturization. As 2nm mass production ramps in earnest, performance-per-watt rises another notch, and AI-chip demand is likely to be reshuffled on top of that again. The other is a change in the character of demand. As ‘agentic AI’ — AI that carries out tasks on its own, as the CFO mentioned — spreads, the compute needed could rise to a different dimension from today.
If we bring this topic up again around 2030, the inflection points to check are roughly three. First, did AI data-center investment actually generate enough revenue to cover its depreciation? Second, did TSMC’s oligopolistic position hold amid new US and Japan plants and rivals’ pursuit? Third, did Korean memory’s profitability endure even after HBM supply expanded? Today’s figure of 66% is only an interim report card on those three questions.
The direction of technology is generally hard to reverse. There is little reason for work that has moved once to AI computation to go back to the old way. So in the long picture, what matters is speed and distribution more than direction. Who takes that demand at what margin, and whether the fruit stays in one place, Taiwan, or spreads widely to Korean memory too, will be the viewing points of the next few years.
FAQ — Frequently asked questions
Q1. If TSMC’s results are good, are Samsung and SK hynix unconditionally good too?
The direction is broadly the same, but the intensity differs. As AI-accelerator demand grows, the HBM demand attached beside it grows with it, which favors Korean memory. But foundry is a bottleneck-holding oligopoly, while memory is a market where prices are sensitive to supply expansion. Even riding the same cycle, the way profitability persists may differ.
Q2. Why is 66% of revenue being AI a risk?
Because it means a company’s earnings are effectively tied to a single theme (data-center investment). It is powerful while that investment continues, but if the investment pace bends, earnings volatility grows just as much. Concentration works as a lever in the upswing and a burden in the correction.
Q3. Is raising capex to $64 billion a good signal?
It reads as a signal that the company is confident in future demand. But capex is reflected in earnings through depreciation over several years. If demand does not continue as expected, the capacity laid down in advance can become a burden — worth watching too.
Q4. If Q3 margins are pressured, is that a sign earnings are turning down?
The company explained it is due to early 2nm mass-production costs. It is closer to a temporary factor commonly seen in the stretch of launching a new process. Still, one needs to clearly recognize that margins will not stay at this quarter’s level (60%-plus) forever.
Q5. How should I view semiconductor-related assets now?
This piece is not a specific-stock recommendation. TSMC’s results confirm the fact that ‘AI demand is still strong,’ but they do not tell you the timing of the cycle’s peak. It is safer to judge against your own investment horizon and weighting, keeping both scenarios — continued demand and oversupply — on the table.
In closing — the stronger the number, the more you ask back the premise
This piece is not a buy/sell recommendation — it is an analytical memo dissecting the structure of the reported results. TSMC’s Q2 was clearly its strongest quarter in history, and the fact that two-thirds of it was AI shows the cycle is still alive. At the same time, that strength came from a concentration on a single source of demand, and the company itself handed over the invoice too — Q3 margin pressure and massive capital investment.
Korean memory has the same-direction tailwind, but foundry and memory reach their peaks in different ways. What is needed when you meet a good number is not cheering, but asking how long the premise propping that number up will hold. Deciding carefully against your own investment horizon, funds, and weighting is what matters most.
Primary sources / Corporate IR
TSMC Q2 2026 earnings release and earnings call (investor.tsmc.com)
Ministry of Trade, Industry and Energy, June 2026 export-import trends (Korea Policy Briefing, korea.kr)
Reference media
Investing.com (TSMC Q2 2026 slides) · TrendForce (capex upgrade) · Taipei Times · TechTimes · Reuters
This article was automatically translated from Korean by AI. Please refer to the Korean original for the most accurate content.