TSMC’s October 8 revenue disclosure supplied a physical checkpoint in the AI-investment debate. Third-quarter sales reached about NT$1.49 trillion, a record above the roughly NT$1.46 trillion market forecast. Bloomberg calculated year-on-year growth near 51%; Reuters rounded it to approximately 50%. Those descriptions concern the same strong quarter, not conflicting operational outcomes.
The result shows customers continuing to convert computing plans into manufactured silicon. It does not prove the services eventually using those chips will earn adequate returns. That distinction matters for a foundry serving Nvidia, AMD, Apple and custom-chip developers: its broad customer base measures advanced-chip production better than it measures the entire AI economy.
September completed a quarter of sustained growth
September revenue was NT$511.86 billion, up 54.6% from a year earlier and down 0.6% from August. July grew 44.7% year on year and August 53.3%, so the quarter was not sustained by just one exceptional month. Sales for the first nine months reached NT$3.899 trillion, up 41.1%.
The quarterly total was roughly 18% above the second quarter’s NT$1.270 trillion. News reports translated third-quarter sales to about US$46.7 billion, above management’s US$44.6–45.8 billion guidance. New Taiwan dollar reporting and US-dollar guidance use exchange-rate assumptions; currency can alter comparisons without changing physical demand.
The monthly figures are unaudited. At the time of this sales release, full third-quarter profit, margins, expenses and cash flow were not yet disclosed. The financial calendar then scheduled the earnings conference for October 15 and the October sales release for November 10. The sales evidence should not be confused with later results absent from this disclosure.
Monthly reporting does not identify every customer or separate accelerators, smartphone processors, networking and other products. High-performance computing—AI accelerators, data-center processors and related silicon—has been the largest disclosed growth engine, but inventory before smartphone launches also supported utilization. The entire 51% increase cannot be assigned to generative AI.
Orders commit more than an optimistic forecast
Advanced chip production requires capacity reservations, completed designs, masks, packaging coordination and system schedules months before an accelerator reaches a data center. TSMC revenue therefore follows physical manufacturing commitments rather than merely interest in a software subscription.
AI processors also demand advanced packaging that links compute dies with high-bandwidth memory. CoWoS is one method of making those bandwidth- and power-sensitive connections. Constraints can move beyond leading-edge wafers to packaging, testing, substrates, memory, networking, power equipment, cooling and final assembly.
The preceding quarter already showed the cycle’s scale: NT$1.270 trillion revenue and NT$706.56 billion net income. Revenue rose 36%; net income and diluted earnings per share each rose 77.4%, according to the second-quarter release. These are second-quarter figures, not third-quarter profits inferred from record sales.
Management subsequently raised the full-year revenue-growth outlook to slightly above 40% in US dollars and capital spending to US$60–64 billion. Planning, equipping, qualifying and scaling fabs takes years; that investment assumes demand beyond one month or generation, but management’s expectation is not a guarantee of future utilization.
TrendForce forecast 24.8% foundry-revenue growth in 2026, citing Nvidia and AMD GPUs alongside custom chips from Google, AWS, Meta and AI startups. This is an outside forecast, not realized industry revenue. Competing architectures can broaden TSMC’s opportunity because all still need fabrication, packaging and memory integration.
The buildout is strong even while its returns remain open
The criticism of AI spending concerns capital committed before commensurate service revenue. TSMC’s sales show that an expected retreat had not reached this manufacturer by the reported quarter. Current demand can remain high even if some future data centers prove uneconomic.
The foundry aggregates competing systems without selling a dominant AI model or operating a major cloud platform. Nvidia, AMD, custom silicon, phones and networking all contribute. September’s acceleration and consecutive record quarters support a multiyear hardware-cycle interpretation, not a final verdict about customer monetization.
Training creates concentrated projects; inference expands with service usage. Customers may increasingly optimize power, memory bandwidth, networking, availability and cost rather than only peak benchmark performance. Efficiency can reduce hardware for a fixed workload while making more applications viable and increasing aggregate demand.
Samsung’s stronger memory results and Foxconn’s server momentum provide supporting supply-chain context beyond logic wafers. They do not prove the entire AI economy healthy, but together suggest activity extending into memory and assembly. The harder test remains whether installed capacity produces services worth computing, power and cooling costs.
Record sales leave several financial risks unresolved
More revenue does not automatically protect gross margin. Advanced nodes, packaging and overseas fabs need upfront investment, followed by depreciation even when equipment is underused. A US$60–64 billion capital budget raises the hurdle for future demand.
Expansion in the US, Japan and Europe diversifies a network concentrated in Taiwan, but labor, construction, utilities, supplier depth and scale differ. TSMC previously estimated overseas ramps would dilute gross margin by two to three percentage points initially, with greater dilution later as expansion grows. That is an acknowledged resilience cost, not evidence that diversification has failed.
The October sales release could not resolve how stronger utilization and pricing offset those costs. Margins and forward guidance were separate checkpoints at the then-scheduled earnings event. A weaker margin would not erase orders; it would show that producing record sales had become more expensive.
Serving many designers also does not remove concentration in a few exceptionally large AI customers. A delayed chip, supplier change or reduced capital plan can move schedules. Tight supply can encourage double ordering; contracts, deposits and long planning cycles constrain casual reservations without eliminating forecasting mistakes.
Smartphone seasonality and currency add interpretation risks. Future quarters need to distinguish continuing AI production from temporary inventory building, and physical growth from exchange-rate effects. Customer usage can grow while unit economics remain difficult; revenue momentum is not proof that every reserved fab or data center is justified.
Supply access and credible second sources become strategic
Finished processors require coordinated wafers, packaging, HBM, substrates, networking and assembly. A good design has limited commercial value without volume delivered on schedule. Large cloud companies can make multiyear commitments across chips, facilities and power; smaller AI firms face harder financing choices, reinforcing demand for cloud access and custom silicon.
TSMC must allocate across different nodes, packages and schedules while avoiding idle capacity when the cycle cools. Custom processors broaden its base but increase execution complexity. Competition among designers can support the foundry without making its own technology lead permanent.
Samsung and Intel remain potential alternatives in process and packaging, both seeking outside foundry customers alongside internal operations. Winning selected programs could establish credibility without replacing TSMC everywhere. Second-source qualification may take generations, giving customers leverage and resilience gradually rather than instantly.
Rivals must show yield, performance, efficiency, reliability and delivery, not merely a node’s marketing name. TSMC’s supplier network, operating knowledge and customer trust are difficult to replicate while its revenue finances further expansion. Samsung can benefit from memory even before securing logic-foundry wins; Intel’s foundry roadmap still requires execution.
Equipment and power create further limits. Lithography, deposition and inspection suppliers have their own lead times. Finished buildings are not qualified high-volume output. Subsidies can accelerate construction, but cannot instantly create experienced workers or mature supplier networks; fabs and AI data centers also compete for dependable electricity.
The next evidence concerns profits and customer behavior
At the time of the October 8 disclosure, three follow-ups were relevant: margins at the scheduled October 15 conference, fourth-quarter guidance and packaging/utilization commentary beyond seasonal phones, and customers’ subsequent capital decisions. The November 10 sales release was another scheduled checkpoint, not data already available in the reported figures.
Continued accelerator designs reaching volume would broaden the base; delays or lower capital budgets could warn before they fully appear in sales. Adequate service revenue, usage and returns alongside expanded manufacturing would make this record an early milestone of a sustainable cycle.
The opposite outcome—monetization trailing spending—would leave the same record as evidence of an expensive investment stage. TSMC can show that customers were still building. Whether that capacity earns its cost remains a separate question for those customers, not a conclusion supplied by one revenue release.
Sources
TSMC September revenue · TSMC second-quarter earnings · TrendForce foundry forecast · Quarterly revenue reporting · TSMC financial calendar · Bloomberg quarterly report