TSMC (TSM) Q2 2026 earnings review
Guidance Jumps to 40% Growth as AI Demand Overwhelms Supply
TSMC delivered revenue of $40.2 billion, at the very top of its guidance range, up 33.7% YoY. Operating margin of 60.3% blew past the high end of guidance (58.5%), and net income surged 77% YoY to NT$706.6 billion. One caveat on earnings quality: NT$63.2 billion of that came from a one-time gain on the sale of Vanguard International Semiconductor shares - roughly 9% of net income. Excluding it, net income still grew about 62% YoY (derived). The bigger news is the outlook: full-year 2026 revenue growth was raised from 'above 30%' to 'slightly above 40%' in USD terms, the capital budget jumped from $52-56 billion to $60-64 billion, and TSMC announced an additional $100 billion investment in Arizona. The 2026 dividend rises 33% to NT$24 per share.
🐂 Bull Case
3Q26 guidance of $44.6-45.8B implies 37% YoY growth - faster than 2Q's 33.7%. The full-year 40%+ target implies a ~$50B fourth quarter, up nearly 50% YoY (derived). Management says supply stays tight through at least 2027, with demand strong 'all the way to probably 2029, 2030'.
Gross margin of 67.7% and ROE of 45.9% are at all-time highs, even while absorbing overseas fab dilution. TSMC funds a $60-64B CapEx year entirely from operating cash flow while raising the dividend 33%.
The rise of agentic AI is reviving CPU demand in data centers on top of GPUs/accelerators - and x86, Arm, and RISC-V CPU designers are almost all TSMC customers.
🐻 Bear Case
Accounts receivable grew 87% YoY against 36% revenue growth, with receivable days up from 23 to 29 in a year. Free cash flow has declined for two straight quarters despite record profits, as CapEx outruns operating cash flow growth.
Today's CapEx builds capacity for 2028-2029. CEO Wei himself admits customer forecasts are inflated: 'all the truths together is not a truth.' If AI infrastructure spending normalizes before then, TSMC owns overbuilt capacity at structurally higher overseas costs.
3Q26 gross margin is guided down 1.7pp to 66% as the N2 ramp dilutes margins by 3-4pp in H2 - wider than the 2-3pp full-year dilution guided in April. Overseas fab dilution widens to 3-4pp in later years.
⚖️ Verdict: 🟢🟢
Bullish. A guidance raise of this magnitude - from 30%+ to 40%+ mid-year - backed by a beat above the top of the margin range is rare for a company this size. Even stripping the one-time VIS gain, underlying earnings grew ~62%. A grade of 4 is defensible given the receivables spike, declining free cash flow, and the sheer scale of the CapEx bet, but the operating evidence this quarter is unambiguous.
Key Themes
HPC Hits 66% of Revenue - the Mix Shift Keeps Accelerating
HPC revenue grew 20% QoQ and now makes up two-thirds of the company, up from 60% a year ago, while Smartphone fell to 22% (-4% QoQ). This trend is accelerating: TSMC's revenue is increasingly a direct function of AI data center buildouts. Automotive was a quiet bright spot, up 15% QoQ.
N2 Arrives: 3% of Wafer Revenue in Its First Meaningful Quarter
2nm contributed 3% of wafer revenue in 2Q26 (from 0% in 1Q26), with a steep ramp guided for 3Q26. Management expects N2 to be a bigger node than N3 was at the same stage and a 'larger and longer-lasting node than 3nm.' The ramp is also why inventory days jumped 7 days to 87 - wafers in progress ahead of shipments. Meanwhile 3nm rose to 30% of wafer revenue, and TSMC is adding three new 3nm fabs (Taiwan, Arizona, Japan) plus converting 5nm tools to meet multi-year N3 demand.
Agentic AI Revives the CPU
Management called out a resurgence of CPU demand in AI data centers driven by agentic AI workloads - incremental silicon demand on top of AI accelerators. TSMC's position is architecture-agnostic: whether x86, Arm-based, or RISC-V, 'they are almost all TSMC's customers.' The company is allocating leading-edge supply to balance CPU, GPU, and XPU demand. This was a named reason for the CapEx raise to $60-64 billion.
A14 Family Expands: A13 and A12 Unveiled
The A14 node (2nd-gen nanosheet) is on track: internal test vehicles show close to 90% device performance and ~90% yield on 256Mb SRAM, customer tape-out activity is ahead of schedule, pre-production starts 2027, volume in 2028. Two new derivatives were introduced: A13 (over 6% die area saving via a 97% optical shrink, backward-compatible design rules) and A12 (brings Super Power Rail to the A14 platform), both slated for 2029. Management expects the A14 family to be an even larger, longer-lasting node than N2.
Receivables Growing Far Faster Than Revenue
The one data point that contradicts the otherwise flawless quarter: accounts receivable rose 87% YoY to NT$440.9 billion against 36% revenue growth, and receivable days climbed from 23 a year ago to 29. Possible explanations include timing of billings around the quarter-end revenue surge and growing concentration among a few very large HPC customers with stronger negotiating power on payment terms. Management did not address it. Not alarming at 29 days in absolute terms, but the trajectory needs monitoring - this is how customer-concentration risk shows up in the balance sheet first.
N2 Margin Dilution Widens to 3-4pp
In April, management guided N2 to dilute full-year 2026 gross margin by 2-3 percentage points. Now the H2 dilution alone is guided at 3-4pp - the ramp is steeper and more margin-expensive than planned three months ago. Add the standing overseas-fab dilution (2-3pp early stage, widening to 3-4pp later) and 3Q26 gross margin is guided down 1.7pp sequentially to 66% at midpoint. Strong leading-edge demand and cost improvement partially offset, but the direction for H2 margins is down, reversing four straight quarters of expansion.
Consumer End Markets Under Pressure (Macro)
Management flagged that consumer and price-sensitive segments are 'being challenged' by rising component prices - chiefly memory - and macroeconomic uncertainty, and said it is 'being prudent' in business planning. Smartphone revenue fell 4% QoQ and its share hit 22%, the lowest in the data set. TSMC's high-end exposure insulates it partially, but a consumer recovery is clearly not part of the 40% growth story - this is now an almost pure AI/HPC thesis.
The CapEx Cycle Rests on Judgment, Not Contracts
Year-to-date, 2026 CapEx guidance has moved from $52-56B (January) to the high end (April) to $60-64B (July), and management said the next three years will be 'even more significantly higher' than the past three years' $101B. This capacity lands in 2028-2029. Wei was candid about the epistemics: every customer gives aggressive forecasts, and 'you put all the truths together, it's not the truth.' TSMC says it audits AI data center construction progress to ensure its chips don't sit in inventory - a sensible check, but the downside scenario (AI capex normalizes before 2028) would leave overbuilt capacity at structurally higher overseas cost. Part of the CapEx raise is also simply inflation: 'now we buy the tools with inflation price.'
Mature Nodes Bifurcate: Power ICs and Sensors Tight, Consumer Soft
Within mature nodes, only AI-adjacent segments are in shortage: power management ICs (every AI data center needs them) and sensors. Commodity consumer demand remains weak, and TSMC continues to focus mature capacity on higher-value segments - JASM Fab 1 in Japan for CMOS image sensors, ESMC in Germany for automotive/industrial. Two related portfolio moves this quarter: a preliminary strategic partnership with Sony on next-generation image sensors (May 8), and the sale of an 8.1% stake in Vanguard International Semiconductor - a further step back from commodity mature-node exposure.
Advanced Packaging So Tight That Competition Is Welcome
Packaging capacity 'is so tight that now it's limiting my customers' growth,' per Wei. Asked about Intel's EMIB-T gaining traction, he welcomed it - rival packaging capacity lets more of TSMC's front-end wafers ship, which is the majority of the business. He drew a hard line, though: back-end competition is not a gateway to front-end share loss ('two different things'). COUPE, TSMC's co-packaged optics platform, has started production and is expected to become 'a very important technology' as data centers chase lower power and higher bandwidth. A CoWoS cost-reduction pilot line needs about one more year to mature.
Other KPIs
The line that flatters reported EPS. NT$63.20 billion came from disposal and mark-to-market gains on Vanguard International Semiconductor shares - a one-time item worth ~9% of net income. Excluding it, non-operating income of ~NT$32.6 billion (derived) is in line with 1Q26's NT$28.8 billion, and net income still grew roughly 62% YoY (derived). The effective tax rate also rose to 18.1% from 16.8%, reflecting tax on undistributed retained earnings. Operating income - the clean measure - grew 65% YoY.
Decelerating for a second straight quarter (NT$368.6B in 4Q25, NT$348.2B in 1Q26) even as profits set records - the CapEx ramp (NT$496B this quarter, +41% QoQ) is outrunning operating cash flow growth. OCF of NT$783.4B was itself a record, so this is deliberate investment, not deterioration, but H2 CapEx of $33-37 billion (derived from the $60-64B budget minus $26.8B spent YTD) means free cash flow stays compressed while the dividend commitment grows.
Up 11.1pp YoY and 5.4pp QoQ - extraordinary for a company carrying NT$3.5 trillion in cash and marketable securities and 69% equity funding. Net cash reserves reached NT$2.49 trillion. Wafer shipments grew 16.6% YoY against 36% NTD revenue growth, meaning roughly half the growth came from richer mix and pricing, not volume.
Guidance
Accelerating. The $45.2B midpoint implies +12% QoQ and +37% YoY, faster than 2Q26's 33.7% YoY. Drivers: the steep N2 ramp and continued leading-edge demand. The exchange rate assumption is 32 NTD/USD, slightly weaker than 2Q's 31.60 average.
Accelerating, and a major raise: guidance has moved from 'close to 30%' (January) to 'above 30%' (April) to 'slightly above 40%' now. The arithmetic is demanding: with $76.1B booked in H1 and ~$45.2B guided for 3Q, hitting 40% growth implies a ~$50B fourth quarter, up roughly 48% YoY (derived). Management's basis: extremely robust AI demand, the agentic AI/CPU wave, and the N2 ramp. That said, prior-year comparisons ease in 4Q (4Q25 grew slowest of 2025), which helps the optics.
Decelerating/Reversing. Gross margin midpoint of 66% is down 1.7pp from 2Q's 67.7%, ending four quarters of expansion. The cause is the N2 ramp diluting H2 gross margin by 3-4pp - wider than the 2-3pp full-year dilution guided in April - partially offset by strong demand, productivity gains, and cross-node capacity optimization. Overseas fab dilution remains a structural 2-3pp headwind, widening to 3-4pp in later years.
Accelerating sharply. Raised from $52-56B in January (already a step up from 2025's $40.9B). With $26.8B spent in H1, H2 implies $33-37B - a quarterly run-rate near $17-18B versus $15.7B in 2Q. Allocation: 70-80% advanced process, ~10% specialty, 10-20% packaging/testing/mask. Reasons given: demand ('the gap is very big') and tool-price inflation. Separately, TSMC announced an additional $100 billion Arizona investment (total now $265B), adding roughly four more fabs for 2nm-and-below plus advanced packaging - no firm schedule disclosed.
Raised again - the 4Q25 call had promised 'at least NT$23.' The quarterly dividend steps up to NT$7.00 starting with the 1Q26 payout (distribution October 8, 2026), from NT$6.00. Management committed to continued increases in 2027. NT$467B was paid out in 2025, up 28.6% YoY. No buybacks - TSMC returns cash exclusively via dividends.
Key Questions
What Is Behind the Receivables Spike?
Accounts receivable grew 87% YoY against 36% revenue growth, and receivable days rose from 23 to 29 in four quarters. Is this billing timing, extended payment terms for the largest AI customers, or something else? Management offered no explanation.
Why Did N2 Dilution Widen From 2-3pp to 3-4pp in Three Months?
April guidance put N2's full-year 2026 gross margin dilution at 2-3pp; now H2 alone is 3-4pp. Is the ramp simply steeper (bullish volume signal), or are N2 yields and costs tracking worse than planned?
What Does the $100 Billion Arizona Commitment Assume?
No timeline, no funding schedule, no margin-dilution quantification was given for the roughly four additional fabs. At what point does the Arizona cluster's cost structure - and any premium pricing for geographic flexibility - get disclosed in numbers rather than direction?
How Concentrated Is AI Revenue Now?
An analyst noted top-5 customer exposure is likely the largest in TSMC's history. Management dismissed the concern by pointing to new AI players, but disclosed nothing. What share of revenue do the top two customers represent, and how much of the 2028-2029 capacity is underwritten by them?
What Would Make TSMC Slow Down?
Management checks AI data center construction progress to ensure chips don't pile up in inventory. What specific leading indicators - data center completions, power availability, customer inventory - would trigger a CapEx deceleration, and how quickly can $60B+ of annual spending be throttled?