TSMC Sales Rise 45% as AI Hardware Demand Defies Market Jitters

TSMC reported July revenue of NT$467.58 billion, about $14.5 billion, despite renewed doubts about the scale and durability of AI spending. Sales increased 44.7% from a year earlier, extending a run that has made the company a central measure of AI hardware demand.

The result matters because financial markets have grown less willing to reward AI investment on promises alone. Cloud providers face questions about rising capital expenditures, delayed returns, and the useful life of expensive accelerators. TSMC’s factories provide a harder signal because customers must commit to physical production capacity well before their chips reach data centers.

The July number does not settle the debate over an AI investment bubble. One month of sales can move with production schedules, currency changes, and customer launches. However, the increase reinforces evidence from TSMC’s latest quarter that advanced chip demand remains stronger than recent market volatility suggests.

TSMC’s July Sales Turn AI Spending Into Factory Revenue

The 45% increase shows that large AI infrastructure commitments are still reaching the manufacturing floor.

TSMC’s monthly sales data showed consolidated revenue of NT$467.58 billion for July. That was 44.7% above the same month in 2025 and roughly 5.6% above June 2026 revenue.

The year-over-year comparison is the more important figure. Monthly changes can reflect shipment timing, but annual growth near 45% points to a much larger production base. TSMC is manufacturing more high-value silicon while its newest processes carry a growing share of wafer revenue.

Bloomberg highlighted the result as evidence that demand for AI hardware remained firm despite market jitters. Those jitters include falling valuations for some AI-linked companies and growing scrutiny of cloud capital expenditures.

TSMC does not report AI revenue as a separate monthly category. Its sales span smartphones, automotive products, consumer electronics, and high-performance computing. High-performance computing includes AI accelerators, server processors, and other demanding chips, so it offers the closest available business segment.

That category accounted for 66% of TSMC’s second-quarter revenue. The figure makes AI infrastructure an important driver, even though it does not provide a clean measurement of accelerator sales.

TSMC’s broader quarterly results support the July reading. Second-quarter revenue reached NT$1.27 trillion, or $40.20 billion, according to its earnings filing. Revenue increased 36% in New Taiwan dollar terms and 33.7% in US dollar terms from a year earlier.

Net income rose 77.4% to NT$706.56 billion. The company also recorded a 67.7% gross margin and a 60.3% operating margin. Those margins suggest that demand is not growing only through discounted or less profitable production.

The process mix provides another useful clue. Technologies at 7 nanometers or smaller generated 77% of total wafer revenue in the second quarter. Three-nanometer production contributed 30%, while 5-nanometer production contributed 33%.

Two-nanometer chips contributed another 3% during an early production ramp. A process node describes the manufacturing generation used to build a chip, with newer nodes generally improving density, performance, or energy efficiency.

These advanced processes are essential for leading AI accelerators. They also serve premium smartphone processors and other demanding designs, so the figures should not be treated as pure AI sales.

Still, the combination is difficult to dismiss. Rapid monthly growth, record quarterly revenue, high margins, and an advanced-node-heavy mix all point in the same direction. AI demand has moved beyond speculative orders and into substantial manufacturing volumes.

Why AI Hardware Demand Is Holding Up

AI chip demand remains strong because the largest buyers are building multiyear computing platforms, not filling a single replacement cycle.

An AI data center requires more than graphics processors. It needs networking chips, central processors, high-bandwidth memory, storage, power systems, and advanced packaging. TSMC participates in several parts of that chain through its manufacturing and packaging services.

Nvidia remains the most visible source of accelerator demand. Its systems depend on advanced logic manufacturing and sophisticated packaging that combines processors with nearby memory. AMD, Broadcom, and custom-chip programs from large cloud companies add further demand.

This diversity matters. TSMC is not relying on one model developer or one accelerator design. It manufactures chips for customers pursuing different approaches to training, inference, networking, and general-purpose computing.

Training involves adjusting an AI model using large datasets and extensive computing resources. Inference is the process of using a trained model to generate answers or complete tasks. Both require hardware, although their performance, memory, and cost priorities differ.

Demand has also expanded from model training into AI services used every day. Search features, coding assistants, media generation, advertising systems, and enterprise agents increase inference workloads. Those workloads can grow with usage even when model development slows.

Agentic AI adds another possible demand source. These systems perform multistep tasks by calling tools, checking results, and revising their actions. A single user request can trigger several model operations, increasing computation compared with a basic chatbot exchange.

TSMC Chairman and CEO C.C. Wei has described AI demand as structural and multiyear. Company guidance still requires careful treatment because customers can revise orders. However, capacity commitments and capital spending show that management expects demand to persist.

The company raised its full-year 2026 revenue outlook to growth slightly above 40% in US dollar terms. It also increased planned capital expenditure to between $60 billion and $64 billion.

Capital expenditure funds factories, manufacturing equipment, and advanced packaging capacity. These projects take years to plan and complete, making the budget a stronger commitment than a short-term sales forecast.

TSMC expects third-quarter revenue between $44.6 billion and $45.8 billion. The midpoint implies another sequential increase from the second quarter. It expects a gross margin between 65% and 67%, even as new overseas facilities and the 2-nanometer ramp add costs.

The July result makes that guidance appear more achievable, but it does not guarantee the quarter. Production schedules can shift between months, while changes in exchange rates affect reported US dollar revenue.

Demand also reaches beyond TSMC. ASML, the main supplier of extreme ultraviolet lithography systems used for advanced chip production, raised its 2026 outlook in July. Its updated forecast followed strong orders linked to the expansion of AI manufacturing capacity.

The company reported second-quarter revenue of €9.33 billion and net income of €2.92 billion, according to its capacity outlook. Strong business at both TSMC and a critical equipment supplier indicates that expansion extends across multiple layers of the supply chain.

This does not mean every semiconductor category is booming. Consumer electronics, industrial chips, and automotive components can follow different cycles. AI hardware demand is strong enough, however, to reshape the overall revenue mix at the industry’s most important foundry.

The Real Contest Is Demand Versus Capacity

TSMC’s biggest near-term challenge is not finding AI customers, but converting demand into finished systems without creating new bottlenecks.

Advanced AI chips require more than leading-edge wafers. After fabrication, processors must be packaged with memory and other components. Advanced packaging connects those parts with the bandwidth and energy efficiency required by large AI systems.

CoWoS, short for Chip-on-Wafer-on-Substrate, is one of TSMC’s main packaging technologies for AI accelerators. It places logic chips and high-bandwidth memory together on a large package. That design shortens data paths but requires specialized equipment and production capacity.

Capacity has remained tight because packaging demand grew faster than earlier infrastructure plans anticipated. Building more wafer capacity alone cannot solve that mismatch. An accelerator wafer has limited commercial value until its chips are packaged, tested, and integrated into a working system.

TSMC has said it is expanding advanced packaging and welcomes qualified outside suppliers that can share the load. That position illustrates the unusual scale of the current cycle. A dominant manufacturer has an incentive to support additional packaging capacity because shortages constrain its own front-end growth.

Memory is another possible bottleneck. AI accelerators need high-bandwidth memory, or HBM, which stacks memory dies to deliver much higher data throughput. Samsung, SK Hynix, and Micron must expand qualified HBM output alongside processor production.

Power delivery and networking can also limit deployment. A cloud operator cannot use accelerators effectively without suitable switches, optical connections, cooling systems, and electrical infrastructure. Delays in any one component can postpone revenue across the chain.

This turns the AI hardware race into a coordination problem. Nvidia and other chip designers need enough foundry capacity. TSMC needs packaging materials, equipment, and memory supplies. Cloud companies need buildings, power, cooling, and network connections.

The result is a production pipeline with long lead times. Customers reserve capacity before demand becomes visible in public product sales. That timing makes TSMC revenue an early indicator, but it can also obscure the final use of manufactured chips.

Some orders placed today support services that will not launch for several quarters. Others may replenish limited capacity rather than signal stronger end-user demand. Customers can also order conservatively high volumes when they fear losing access to scarce production.

TSMC’s 2-nanometer ramp adds another layer. The company said this generation accounted for 3% of second-quarter wafer revenue and expected a steep expansion during the third quarter. New nodes initially bring higher costs and operational complexity before yields improve.

Yield is the share of chips on a wafer that meets production requirements. A low yield raises the cost of each usable processor. Improving yield is therefore essential for turning advanced designs into profitable, high-volume products.

TSMC has significant experience managing such transitions. Still, growing 2-nanometer output while expanding packaging and overseas manufacturing makes execution more difficult. Strong orders cannot remove the engineering and construction constraints.

This is why the July sales increase is both encouraging and incomplete. It confirms that demand is reaching TSMC, but the next test is whether the entire supply chain can deliver usable computing systems at the expected pace.

Nvidia, Samsung, and Intel Still Shape the Pressure

TSMC’s sales strength widens its lead, yet customers and governments continue building alternatives to reduce dependence on one manufacturer.

Nvidia benefits directly from TSMC’s manufacturing scale, but it also creates concentration risk. A large portion of premium AI infrastructure spending flows through Nvidia’s accelerator roadmap, making production timing at one customer important to several suppliers.

TSMC does not publicly identify every customer’s monthly contribution. Investors therefore cannot determine how much of July’s growth came from Nvidia, Apple, AMD, Broadcom, or custom silicon programs.

That uncertainty matters because customer concentration can magnify product-cycle swings. If a major accelerator launch moves between quarters, TSMC’s monthly revenue may rise or fall without changing the long-term demand picture.

Samsung offers the broadest alternative because it operates advanced logic foundries, memory production, and packaging businesses. It has pursued 2-nanometer customers while working to improve yields and attract more external designs.

Its position creates a potential advantage for customers seeking a second source. However, advanced chip manufacturing depends on process maturity, design tools, intellectual property libraries, packaging, and predictable volume execution. A nominally comparable node does not automatically provide an interchangeable product.

Intel Foundry represents another alternative, especially for US and European policymakers seeking more geographic diversity. Intel has invested in advanced manufacturing and packaging while trying to serve external chip designers.

The company must prove that it can deliver competitive technology on schedule and support customers accustomed to TSMC’s foundry model. That effort remains strategically important even while TSMC’s sales continue rising.

TSMC’s expansion outside Taiwan responds partly to this pressure. The company plans additional advanced manufacturing in Arizona, alongside facilities in Japan and Germany. These projects bring production closer to major customers and government partners.

TSMC said in July that it would commit another $100 billion to US expansion, bringing its planned American investment to $265 billion. The plan includes additional advanced fabrication, packaging, and research capacity.

The US expansion can reduce some geographic concentration, but it does not quickly recreate the scale of Taiwan’s semiconductor cluster. Construction, workforce development, supplier localization, and equipment installation require years.

Overseas facilities can also carry higher costs. TSMC has warned that early stages of international expansion can dilute margins. Its third-quarter margin guidance remains high, yet future results must absorb more depreciation and startup expenses.

Customers face a difficult choice. Concentrating orders with TSMC offers access to proven advanced processes and packaging. Diversifying across foundries can improve resilience, but it may require redesigned chips, added engineering work, and different performance tradeoffs.

Governments face a similar tension. Subsidies can encourage domestic factories, but policy cannot compress every stage of technical qualification. A facility becomes strategically useful only after it produces competitive chips reliably and at sufficient volume.

The July sales result therefore pressures rivals without eliminating their opportunity. TSMC’s growth demonstrates the value of scale and manufacturing consistency. At the same time, that success gives customers and governments a stronger reason to fund alternatives.

What the 45% Increase Does Not Prove

A strong month confirms current production demand, but it cannot prove that every AI investment will generate an acceptable return.

The largest uncertainty sits beyond the chip factory. Cloud providers must turn accelerators into services that attract customers, reduce costs, or improve existing products. Manufacturing revenue arrives before that business outcome becomes clear.

Capital spending can remain strong even when expected returns fall. Companies may continue investing because competitors are doing so, because capacity is scarce, or because missing an AI platform shift appears more dangerous than overspending.

That behavior can support TSMC for several quarters. It can also create excess capacity if product adoption disappoints. Semiconductor history contains repeated cycles in which shortages encouraged aggressive expansion before demand weakened.

AI hardware has characteristics that distinguish it from a simple commodity cycle. Leading accelerators evolve quickly, software ecosystems affect hardware choices, and the largest buyers possess substantial cash flow. These factors support continued investment, but they do not remove cyclical risk.

The useful life of AI hardware creates another question. New processors can deliver higher performance per watt, making older systems less competitive before they physically wear out. Rapid replacement helps chip sales but can weaken the economics of previous purchases.

Power availability may slow deployments even when processors are ready. New data centers require grid connections, substations, cooling systems, and regulatory approval. A chip order does not guarantee that the customer can operate the hardware immediately.

Geopolitics remains a larger structural risk. Most leading-edge TSMC production still sits in Taiwan, which makes cross-strait stability important to the global technology sector. Overseas expansion reduces concentration gradually rather than immediately.

Trade restrictions can also change customer access and product designs. US controls limit shipments of certain advanced AI chips to China. Suppliers may create modified products, but policy changes can still alter volumes and development priorities.

Costs present another uncertainty. TSMC has identified possible pressure from materials, overseas manufacturing, and the initial 2-nanometer ramp. High demand can offset those costs, although maintaining margins above 60% indefinitely should not be assumed.

The July comparison also benefits from the timing and size of the previous year’s base. Growth rates naturally change as revenue expands. A 45% annual increase becomes harder to repeat when the comparison period already includes large AI orders.

Monthly sales offer little visibility into order cancellations or inventory held elsewhere in the supply chain. Customers may have binding capacity commitments, yet finished systems can still accumulate if deployment schedules slip.

For these reasons, the result should not be framed as proof that concerns about an AI bubble are mistaken. It is better read as evidence against an immediate collapse in hardware demand.

The distinction is important. Current orders can remain strong while long-term returns remain uncertain. Investors, developers, and enterprise buyers should separate physical deployment from successful adoption.

A data center full of accelerators measures installed capacity. It does not measure whether users value the resulting services, whether inference costs fall fast enough, or whether customers will pay for new features.

TSMC’s numbers answer a manufacturing question with unusual clarity. Chip production is still expanding. They cannot answer the final economic question, which is whether AI applications will justify the investment required to build that capacity.

Three Signals That Will Test the AI Hardware Boom

The next evidence should come from TSMC’s monthly trajectory, customer spending plans, and the delivery rate of advanced capacity.

The first signal is TSMC’s August revenue, scheduled for release in September. Another strong annual increase would show that July was not simply a shipment-timing event. A sharp sequential decline would not prove that AI demand has collapsed, but it would make the quarterly mix more important.

Investors should compare the August figure with TSMC’s third-quarter revenue guidance of $44.6 billion to $45.8 billion. Meeting the upper half of that range would strengthen the argument that advanced-node demand remains broad and sustained.

The second signal is capital spending from major cloud operators. Microsoft, Alphabet, Amazon, and Meta are among the companies building large AI computing fleets. Their results should reveal whether spending plans are increasing, holding steady, or moving into a slower phase.

The most useful information will go beyond total expenditure. Investors need to know whether new capacity is becoming available on schedule, whether AI services are producing revenue, and whether management expects similar investment next year.

Lower spending would not affect TSMC immediately because semiconductor orders carry long lead times. It would, however, weaken expectations for future capacity reservations. Continued increases paired with measurable AI revenue would support the current manufacturing outlook.

The third signal is execution across 2-nanometer production and advanced packaging. TSMC expects a steep 2-nanometer ramp, while packaging remains essential for converting processors and memory into AI systems.

Rising 2-nanometer revenue with stable yields and margins would show that the company can expand its newest technology without losing financial discipline. Persistent packaging shortages would indicate strong demand, but they would also cap system deliveries.

The balance matters more than a single growth statistic. TSMC must add capacity quickly enough to serve customers without building far beyond durable demand. Its customers must deploy that hardware quickly enough to turn reserved capacity into useful services.

Developers and enterprise buyers should watch these signals because hardware availability shapes product economics. More capacity can lower inference costs, improve service limits, and make advanced models available to a wider range of applications.

It also affects product planning. Teams building AI features need to distinguish temporary compute scarcity from lasting cost structures. Decisions about model size, latency, data handling, and vendor dependence can change as new hardware reaches production.

TSMC’s July sales make one conclusion reasonable: the AI hardware cycle has not stalled under recent market volatility. Orders continue moving through the most important advanced foundry at substantial scale.

The harder question now is whether infrastructure growth produces equally strong usage. Watch the next monthly sales report, cloud spending commitments, and advanced packaging output. Together, those indicators will show whether the 45% rise marks durable expansion or the strongest stage of a cycle nearing its limits.