TSMC: A Record Quarter, and a Capex Bill That Keeps Climbing

TSMC 2Q 2026: A Record Quarter, and a Capex Bill That Keeps Climbing

A reactive read of Taiwan Semiconductor Manufacturing's second-quarter print, TSMC's first appearance in our coverage. Every profitability line landed at or above the top of guidance and management raised the 2026 revenue growth outlook to over 40%. The stock fell roughly 7% in the two sessions that followed, amid a broader semiconductor sell-off in which investors were also weighing another capex increase and the first sequential gross margin guide-down in several quarters.

Executive Summary

TSMC (2330.TW) reported 2Q 2026 on July 16, and the headline numbers are unambiguous. Net revenue reached NT$1,270.38bn (US$40.20bn), up 36.0% year-over-year and 12.0% sequentially, landing at the very top of the US$39.0-40.2bn guided range. Gross margin hit 67.7%, above the 65.5-67.5% guide, and operating margin reached 60.3%, well above the 56.5-58.5% guide. Net income attributable to shareholders was NT$706.56bn, up 77.4% year-over-year, and diluted EPS of NT$27.25 (US$4.31 per ADR unit) also grew 77.4%. Management raised full-year 2026 revenue growth guidance to "slightly above 40%" in US dollar terms.

The stock did not celebrate. TSM closed near $426 the session before the print and traded down to roughly $398 by July 17, a decline of about 7% over two sessions, before stabilizing near $404 as of July 20 (stockanalysis.com). The decline came amid a broader semiconductor sell-off, and within that backdrop, two disclosures within an otherwise clean beat appear to have weighed on the stock specifically.

Three takeaways:

1. The operating results are as good as TSMC has ever reported, and the beat is genuine. Every profitability metric cleared its guided range, wafer shipments rose 16.6% year-over-year, and Taiwan's most important company just posted a 55.6% net profit margin. This is not a company running into demand headwinds.
2. Capex guidance has moved through three successive guide points this year, and management still cannot tell you where it stops. January established a 2026 capital budget of US$52-56bn, April moved expectations toward the top of that range at roughly US$56bn, and July formally raised it to US$60-64bn, alongside a fresh US$100bn commitment in Arizona. CEO C.C. Wei told analysts the demand-supply gap is "very big" without offering a number, and CFO Wendell Huang said the next three years of capex will be "even more significantly higher" than the last three, which were already a record.
3. Gross margin just reached a record high, and is now guided down. Third-quarter gross margin guidance of 65-67% (66% midpoint) is 170bps below the 2Q26 actual, driven by a 3-4 percentage point drag from the steep ramp of 2nm technology, a headwind management says will persist through the second half and will be compounded over the coming years by widening dilution from overseas fab ramps.

Financial Highlights

Headline results:

2Q26 results against guidance:

Revenue landed exactly where guidance capped it. CEO C.C. Wei described the quarter as "at the high end of our guidance," which is the accurate read rather than a blowout beat. The margin lines are where the real upside showed up, both finishing meaningfully above their guided ceilings on cost improvement and higher capacity utilization.

Revenue by technology node (advanced):

2nm went from zero to 3% of wafer revenue in a single quarter, the leading edge of what management called a "steep ramp-up." Advanced technologies (7nm and below) now account for 77% of wafer revenue, up from 74% a year ago, and the mix is shifting fast enough within the advanced bucket, 3nm up 6 points year-over-year while 5nm and 7nm cede share, that the margin dynamics of the node mix are moving quarter to quarter, not just year to year.

Revenue by platform:

HPC, which spans AI infrastructure, CPUs, and other high-performance compute applications rather than AI alone, grew 20% sequentially and now represents two-thirds of TSMC's revenue, up from 60% a year ago. Smartphone shrank both in share and in absolute sequential terms. This is the clearest read on where TSMC's growth is actually coming from, and it is not diversified.

3Q26 guidance:

The revenue guide is another sequential acceleration. The margin guide is the first sequential step-down in several quarters, and management was explicit about why: a 3-4 percentage point gross margin drag from the 2nm ramp, only partially offset by strong demand and continued cost improvement.

A note on quality of earnings. NT$63.20bn of the NT$95.83bn in total 2Q non-operating income, 66%, came from the disposal and mark-to-market gain on TSMC's stake in Vanguard International Semiconductor, sold down in May. Measured against the sequential increase rather than the total, the gain is even more concentrated: non-operating income rose NT$67.00bn from 1Q26's NT$28.83bn, and the Vanguard gain accounts for roughly 94% of that increase. Management quantified the contribution to EPS directly on the call: NT$2.24 of the NT$27.25 in diluted EPS, roughly 8% of the headline number. Strip it out and adjusted EPS is closer to NT$25.01, still up 63% year-over-year against a NT$15.36 base, still a record, just a less dramatic one. The NT$63.20bn itself is a mix of a realised disposal gain and a non-cash mark-to-market component, not one or the other, and neither piece is operating income, so it should not be extrapolated into the run-rate. Separately, days of inventory rose 7 days to 87, which management attributed to the N2 ramp rather than to slowing sell-through, a pre-build ahead of volume production that is worth tracking alongside the margin dilution story rather than reading as a demand-quality signal on its own.

Balance sheet and capital returns:

Free cash flow of NT$287.36bn arrived even after a record NT$496.00bn capex quarter, funded by NT$783.36bn of operating cash flow. Management confirmed the 2026 cash dividend per share will rise to NT$24, up 33% from NT$18 in 2025, and said it expects "continued and increasing" dividends per share in 2027. The capital intensity is rising, but so is the cash generation underneath it.

Transcript and Management Commentary

Capex guidance has moved through three successive data points this year, and management will not put a ceiling on it. CFO Wendell Huang confirmed the progression directly in Q&A: January established a range of US$52-56bn, April moved expectations toward the top of that range at roughly US$56bn, and the budget now stands at US$60-64bn. Asked for a three-year capex outlook similar to the one TSMC gave during the COVID cycle, Huang declined to give a number but said "the CapEx in the next three years will be even more significantly higher than the past three years." CEO C.C. Wei was more candid about the uncertainty underneath the spending than the raised numbers might suggest. Asked whether he weighs competitive capacity-building alongside customer demand signals, Wei said every customer tries to tell the truth about their own demand, "but I believe every customer tell me the truth, everyone. You put all the truth together, it's not a truth." Asked directly how large the demand-supply gap is, Wei declined to size it beyond calling it "very big," adding he did not want to compare it to memory's own well-known shortage.

The US$100bn Arizona announcement is the single largest disclosure of the quarter. Wei announced an additional US$100bn investment in Arizona, bringing the total commitment there to US$265bn, covering "several more" 2nm-and-below logic fabs plus advanced packaging capacity, with "probably additional four or more fabs" combining frontend and backend. He offered no firm schedule, saying timing "depends on the market situation," though TSMC is "trying to move as fast as we can." Alongside the US expansion, TSMC is building 13 leading-edge and advanced packaging fabs in Taiwan over the next several years and executing on three additional 3nm fabs, in Taiwan, Arizona, and Japan.

The margin story for the second half is now explicit, and it is a headwind, not a tailwind. Huang laid out the mechanics directly: a 3-4 percentage point gross margin drag from the 2nm ramp through 2H26, and a structural, multi-year dilution from overseas fab ramps that he sized at 2-3 percentage points in early stages, widening to 3-4 percentage points in later stages "in the next several years" as Arizona and Japan scale. Partial offsets include strong leading-edge demand, productivity gains, and cross-node capacity optimization, but Huang did not suggest these fully close the gap. This is the first sequential margin guide-down in several quarters.

Agentic AI is reviving CPU demand, and management is unwilling to update its five-year AI CAGR. Wei described a "resurgence in the role of CPUs in AI data centers," noting that x86, ARM, and RISC-V approaches are "almost all TSMC's customers," which broadens the addressable base beyond GPUs and AI accelerators. Pressed by Morgan Stanley's Charlie Chan on whether the January guide of a mid-to-high-50s% five-year AI revenue CAGR needs updating, Wei declined to give a number: "it's stronger and stronger and stronger... we don't know how to answer this question." On customer concentration, Arete's Jim Fontanelli raised the risk of TSMC's largest customers becoming a larger share of the business directly. Wei dismissed the framing, arguing the AI industry has "a lot of new players," not simply more concentration among existing ones.

On competition, Wei leaned on trust and lead time rather than price. Asked about Samsung Foundry's memory-funded margins and Intel ($INTC)'s US policy backing, Wei acknowledged both directly, "I'm jealous" of Samsung's profitability, and confirmed TSMC receives undisclosed government support of its own, before falling back to a five-year technology-qualification lead time as the real moat: choosing a foundry partner "is not buying milk from 7-Eleven." On the emerging EMIB-T backend packaging competition from Intel, Wei's answer was notably unguarded: TSMC's own advanced packaging capacity is "so tight that now it limits my customers' growth," so a competitive alternative that absorbs backend overflow is welcome, since it protects the larger frontend wafer business. On mature nodes, Wei drew a sharp line, power management ICs and CMOS image sensors tied to AI infrastructure are in genuine shortage, while general commodity mature-node demand remains soft on weak consumer spending.

Implications and Outlook

TSM traded down roughly 7% in the two sessions after a quarter that beat every profitability line and came with a raised full-year growth outlook. That decline came amid a broader semiconductor sell-off, so it would overstate the case to pin it entirely on TSMC's own print. Within that backdrop, two things stand out as disclosures investors appear to have weighed specifically: an escalating capex commitment with no stated ceiling, and a margin structure now guided into multi-year dilution.

The bull case rests on the cash generation underneath the spending. TSMC produced NT$287.36bn of free cash flow in a quarter where it also spent a record NT$496.00bn on capex, and still raised the per-share dividend 33% for 2026. Operating cash flow is funding record capex, a growing dividend, and a NT$2,486.33bn (~US$79bn) net cash position simultaneously, even as management admits it cannot fully quantify the demand it is chasing.

The bear case, or at least the caution case, is that TSMC's own visibility is the tell. A company with five-plus years of technology-development and capacity lead time, and a disciplined top-down and bottom-up capacity planning process Wei described in detail, still would not size the demand-supply gap beyond "very big," would not update its five-year AI CAGR beyond "stronger and stronger," and would not provide longer-term capex or deployment timing despite repeated analyst questions. That is either admirable candor about a genuinely unprecedented cycle, or evidence that even the company with the best forward order book in semiconductors is planning capacity off aggregated customer forecasts it has explicitly said it does not fully trust. The margin guide is where that uncertainty first shows up in the numbers: 2nm dilution of 3-4 points is a known, bounded cost of ramping TSMC's next leading-edge node, but the overseas dilution, widening from 2-3 points to 3-4 points over several years as Arizona and Japan scale, is a structural cost of the geopolitical diversification TSMC has taken on, not a temporary ramp cost that resolves itself.

At a forward P/E in the 18.8x-19.1x range and a market capitalization around US$2.07tn (July 17 close), TSM trades well inside the multiples the broader AI infrastructure complex has commanded through this cycle, and Street consensus, a "Strong Buy" average rating with a price target near $520, implies meaningful upside from current levels (stockanalysis.com). Whether the pullback proves durable depends less on the broader sector sell-off it arrived alongside and more on whether 3Q26 confirms the margin trajectory management just guided, or whether cost improvement and utilization gains claw back more of the 2nm dilution than the 66% midpoint implies.

Where This Touches Coverage

TSMC's numbers this quarter land inside several names already in coverage, not just alongside them. AMD ($AMD) is the clearest case. It is fabless, and the HPC-driven revenue mix TSMC just reported, along with the leading-edge capacity it says demand still outstrips, is substantially AMD's own supply chain. The relationship shows up directly in how little capital AMD itself has to carry. AMD's capex ran $389m against $10,253m of revenue in its most recent quarter, a capex intensity of roughly 3.8%. TSMC's own capex intensity, by contrast, was 39.0% of revenue in 2Q26 (NT$496.00bn of capex against NT$1,270.38bn of revenue). AMD is outsourcing much of its manufacturing capital intensity to TSMC rather than carrying it on its own balance sheet. The two names also price very differently for that arrangement. TSMC's enterprise value works out to roughly 13.9x trailing revenue, an EV of approximately US$1.99tn against US$142.9bn of revenue over the last four reported quarters (market cap of approximately US$2.07tn at the July 17 close, net of TSMC's own net cash position), per TSMC's own quarterly releases (3Q25, 4Q25). AMD trades at roughly 22x on the same basis, per our reverse DCF valuation tab. The market pays a meaningfully higher revenue multiple for the company that has to win an allocation of scarce capacity than for the company that owns the capacity and carries the capital intensity behind it, worth holding against TSMC's own refusal this quarter to size how large that scarcity actually is (see our AMD primer).

The capex side connects to Microsoft ($MSFT) more directly than to any other name in coverage, and the connection has real timing behind it, not just a shared theme. Our capex conversion primer tracks Microsoft's capex intensity rising from 25.0% of revenue in fiscal 1Q26 (roughly July-September 2025) to 36.8% in fiscal 2Q26 (October-December 2025), a 47% increase rather than a doubling, then holding at 37.2% in fiscal 3Q26 (January-March 2026). That climb sits inside the same window across which TSMC's own 2026 capital budget moved through its three guide points, January's US$52-56bn range, April's drift toward roughly US$56bn, and July's formal raise to US$60-64bn.

Neither figure fully explains the other. Microsoft's capex covers land, buildings, leases, and networking that never touches a TSMC fab, and a meaningful share of it is not AI-related spend at all, so it is not a clean read on demand addressable to TSMC specifically. But the direction and the timing line up closely enough to matter, and the scale gap is the more important number: Microsoft, Amazon, and Alphabet are guiding to a combined US$570-580bn of CY2026 capex, against TSMC's own US$60-64bn, itself only partially addressable to any single supplier including TSMC. TSMC's inability to size its own demand-supply gap looks less like evasiveness once measured against a customer capex pool that large, even allowing for how much of it never reaches a TSMC fab.

Sony is the most direct link and the most concrete one. TSMC's own recap of the quarter's events discloses a preliminary agreement with Sony Semiconductor Solutions on next-generation image sensors, signed May 8, a live commercial relationship rather than an inferred one, and relevant to the semiconductor segment tracked as a condition in our Sony primer.

Micron Technology ($MU) sits in the same conversation for a different reason. C.C. Wei told analysts on the call he is "jealous" of memory's gross margins, and TSMC's own packaging capacity, tight enough that it is limiting customer growth by his own account, is an interlocking or sequential bottleneck alongside the HBM supply economics covered in our Micron piece, not literally the same capacity pool. Memory and logic are different businesses, built in different fabs, but they are drawing on overlapping parts of the same AI capital cycle and dependent on some of the same downstream packaging and substrate capacity.

What to Watch

• The 3Q26 margin print against the 66% gross margin midpoint. This is the first real test of whether the 2nm dilution management guided to actually lands as sized, or whether cost improvement claws back more of it than guided.
• Overseas fab dilution as Arizona and Japan scale. Management sized this at 2-3 points early, widening to 3-4 points in later stages "over the next several years." The pace of that widening is now a standing input to every future margin guide.
• Capex cadence into 2027-28. The 2026 budget has already moved through three successive guide points. Whether the next multi-year framework, which Morgan Stanley has separately modeled near US$200bn cumulative for 2026-28, gets confirmed or extended again is the single largest variable in the capital story.
• Arizona deployment schedule. The additional US$100bn and "four or more" fabs came with no firm timeline, tied explicitly to customer demand rather than a fixed build-out plan.
• Whether the five-year AI revenue CAGR gets a formal update. Management would not move off the January mid-to-high-50s% figure this quarter beyond qualitative language. A specific revision, up or down, at a future symposium or earnings call would be a material data point.
• A14 and its derivatives. A14 risk production starts in 2027 and volume production in 2028, with A13 and A12 extensions targeted for 2029. Internal product-like vehicles are already showing close to 90% device performance and SRAM yield, while customer tape-out activity is running ahead of schedule.
• Mature-node bifurcation. Power management ICs and CMOS image sensors tied to AI infrastructure remain in shortage while general commodity mature-node demand stays soft, a split worth tracking for read-through to the broader analog and sensor supply chain.

Closing

TSMC's second quarter was, by every conventional measure, the best in the company's history. Revenue landed at the top of guidance, both margin lines cleared their ceilings, net income grew 77% year-over-year, and management raised the full-year growth outlook to over 40%. None of that is in question. The stock fell anyway, amid a broader semiconductor sell-off in which two things sitting just beneath the headline appear to have weighed on TSMC specifically: a capex budget raised again this year with no ceiling management is willing to name, and a gross margin guide that, for the first time in several quarters, points down instead of up. Both are manageable inside a business generating close to NT$300bn of free cash flow a quarter even at record spending levels. Whether they are manageable inside the multiple the market has assigned to that business is the question TSMC now has to answer with results rather than guidance.