AI chips drive around a third of TSMC revenues in Q2

Taiwan Semiconductor Manufacturing Corporation (TSMC), the world's largest foundry, delivered record revenues of $40.2 billion（約6.4兆円） in its second quarter ended June 2026, up 33.7 percent year-on-year and 12 percent sequentially. Net income accelerated even faster, climbing 74.7 percent year-on-year to $22.37 billion（約3.6兆円）, representing 55.6 percent of revenues. The earnings surge was driven primarily by artificial intelligence (AI) chip demand: an estimated $13.31 billion（約2.1兆円） in AI training and inference revenue represented roughly a third of total TSMC sales and grew 68.3 percent year-on-year—outpacing the company's prior forecast of mid-to-high 50 percent compound annual growth through 2030.

During its earnings call, CEO CC Wei signaled confidence in the AI opportunity but declined to update near-term projections, saying only that growth expectations were "stronger and stronger and stronger." Wei acknowledged the challenge of capacity planning when customer demand signals are unreliable: every customer, he explained, provides its "truth" about future demand, but collectively those truths do not form a single coherent truth because each customer inflates its requests to avoid missing AI opportunities. TSMC's capex guidance reflects this uncertainty and confidence combined: the company raised its 2025 capex forecast to $60–$64 billion（約10兆円）, a 56.5 percent increase at the high end from the earlier $52–$56 billion（約9兆円） range. This spending builds on a 37.4 percent increase in 2024 capex (to $40.9 billion（約6.5兆円） from $29.8 billion（約4.8兆円） the prior year), yet annual spending remains a fraction of multiyear commitments because fabs take five to seven years to construct.

The most dramatic commitment is a $100 billion（約16兆円） expansion of TSMC's Arizona operations, raising total US investment to $265 billion（約42兆円） over an unspecified number of years (three to five, TSMC suggests, depending on customer ramp rates). The expansion will add four new fabs in Arizona, all focused on 2 nanometer and smaller process nodes—the most advanced and most profitable transistor geometries. Arizona is already slated to host a 3 nanometer foundry as part of the original plan, mirroring capacity in Taiwan and Japan. TSMC is also converting 5 nanometer facilities to 3 nanometers to boost 3 nanometer output. In Q2, 2 nanometer processes recognized revenue for the first time, totaling $1.21 billion（約1900億円）, driven by AMD's Venice Epyc 9006 server CPUs and MI450 data center GPUs. The ramp is faster than the 5 nanometer and 3 nanometer ramps that preceded it, signaling strong AI demand.

TSMC's operational metrics underscore the value of supply scarcity. Wafer shipments (measured in 12-inch wafer equivalents) totaled 4.34 million units in Q2, up 16.6 percent year-on-year and 3.9 percent sequentially from Q1 2026. Revenue per wafer reached $9,271, up 14.6 percent year-on-year and 7.8 percent sequentially—triple the level of seven years prior and double the level of four years prior. This reflects the shift in TSMC's product mix toward advanced processes and the pricing power afforded by sustained undersupply relative to demand. Wei joked that TSMC must "milk the cash cow but you can't rip the udders off"—a cautious acknowledgment that margin expansion through scarcity has limits. Yet the company's actions reveal comfort with shortages: packaging capacity for advanced interconnects (CoWoS interposers and the emerging COUPE co-packaged optics) remains constrained, and TSMC is not rushing to expand these bottlenecks; Intel's new EMIB-T interposer technology provides partial relief, but only partial.

Within TSMC's revenue segments, the shift toward AI and data center workloads is stark. The HPC (high-performance compute) segment, which includes all data center processors (CPUs, GPUs, and other accelerators) as well as desktop and laptop chips, drove $26.53 billion（約4.2兆円） in sales, up 47.1 percent year-on-year. This segment now far exceeds smartphone revenue, which historically was TSMC's largest business. Smartphone revenues cooled, declining 5.2 percent sequentially (though still up 8.9 percent year-on-year to $8.84 billion（約1.4兆円）), while all other segments—IoT, automotive, DCE (discrete chip electronics), military, and aerospace—contributed $4.82 billion（約7700億円）, up 23.4 percent year-on-year. Within HPC, non-AI chips (by TSMC's classification) represent an estimated $13.23 billion（約2.1兆円）, up 30.5 percent; the remaining $13.31 billion（約2.1兆円） is AI. If this estimate is correct, AI accounts for roughly half of HPC segment revenues.

TSMC's profit generation from AI is likely even more concentrated than its revenue share. AI chips, predominantly manufactured on the newest nodes (2 and 3 nanometers), command higher prices and better margins than mature processes. Wei noted that TSMC's capex will remain roughly 70 percent allocated to front-end chip production (etching and patterning), 10–20 percent to packaging, and 10–20 percent to specialty products, implying that the company views the current mix as optimal. This suggests little urgency to relieve supply constraints; in the semiconductor industry, when demand exceeds supply and prices are rising, the incentive to maximize profits exceeds the incentive to maximize volume. TSMC's $110.22 billion（約18兆円） cash position provides a buffer, and at the high end of the new capex forecast ($64 billion（約10兆円）), the company would spend more than half of that cash in a single year, though actual quarterly spend will be lower given the multi-year construction timeline. Rivals like Samsung Foundry and Intel Foundry may eventually challenge TSMC's dominance, but four to five years of lead time in fabs means TSMC is likely to remain capacity-constrained and highly profitable through the rest of this decade.

TSMC's exceptional Q2 performance reflects a structural imbalance in AI infrastructure investment: spending on AI systems remains "insatiable," in the words of one observer, while foundry capacity—especially for advanced nodes—lags behind demand. CEO CC Wei's lighthearted comment that "you put all the truths together, it's not the truth" underscores the opacity of customer demand signals; every hyperscaler, cloud builder, and AI model maker inflates its capacity requests to avoid missing future opportunities. TSMC must make "very careful judgment" with incomplete information, yet its willingness to commit $265 billion（約42兆円） in US capex and raise its annual spending forecast to $60–$64 billion（約10兆円） signals confidence that AI demand will justify these outlays through at least 2030.

The profitability calculus is equally revealing. Net income of $22.37 billion（約3.6兆円） on $40.2 billion（約6.4兆円） in revenue (55.6% net margin) is extraordinarily high, driven in part by supply constraints that permit price increases. Revenue per wafer has tripled in seven years, reflecting the shift toward advanced processes (3 nanometer and below) that command premium pricing. TSMC's AI segment is likely the most profitable portion of this mix, yet the company has limited transparency in disclosing AI's share of earnings. This opacity, combined with sustained supply tightness, gives TSMC pricing power but may also invite competitive entry: Samsung Foundry and Intel Foundry are well-capitalized, and China's SMIC, while currently representing only 6 percent of foundry business, has government backing and captive demand from Chinese chipmakers locked out of TSMC. However, meaningful competition is four to five years away, given fab build timelines.

No comments yet. Be the first to share your thoughts!