As demand for artificial intelligence chips soars, Taiwan Semiconductor Manufacturing Co. is stepping carefully, signaling that it will avoid the kind of aggressive spending that in past upswings left chipmakers exposed. The company, the world’s largest contract chip producer, is weighing new investment plans in Taiwan, the United States, and Europe while keeping a close eye on the risk of excess capacity.
The timing matters. Orders linked to data centers and AI accelerators continue to climb in 2024 and 2025, yet leaders inside and around the company point to hard lessons from earlier cycles. The message is simple: growth is real, but it may not be linear.
“The Taiwanese chipmaking giant has been burned by previous investment booms.”
History of Booms and Busts
Semiconductor investment has long moved in waves. The dot-com crash in 2001 left foundries with idle tools. A smartphone super-cycle later spurred a rush to expand, only to be followed by inventory corrections mid-decade. The crypto-mining spike in 2017 and 2018 briefly lifted demand for graphics chips before collapsing.
TSMC, which supplies Apple, Nvidia, AMD, and others, has navigated each turn. It raised capital spending during the pandemic to ease a chip shortage, then slowed some plans as consumer electronics cooled in 2023. Analysts say the company wants to avoid repeating periods where billion-dollar fabs came online just as orders dipped.
Current Bet: AI, With Guardrails
Today’s scramble for AI capacity is unlike past cycles in one respect: customers are locking in long-term supply. Cloud providers and chip designers are committing to advanced nodes for accelerators and high-bandwidth memory interfaces. That has given TSMC some visibility into future demand.
At the same time, costs for advanced nodes and overseas fabs are high. The company has warned that building outside Taiwan can carry a meaningful cost premium, and any project delays can compound that. Management has guided capital spending within a range rather than a surge, signaling discipline.
Industry estimates suggest AI-related work was a mid-single-digit share of revenue in 2023 and is rising. But executives have cautioned that traditional segments—smartphones, PCs, and automotive—still swing quarterly results and can blunt the pace of any AI-led upturn.
Competing Views on How Much to Build
Investors are split. Some argue the company should add capacity fast to lock in market share and meet urgent demand for accelerators. They point to record orders at leading AI customers and multiyear roadmaps for new chips.
Others urge restraint, citing cyclical risk, project execution outside Taiwan, and geopolitics. They note that a pullback in consumer electronics or a pause in data-center spending could quickly flip the supply picture.
- Upside case: AI demand stays tight through 2026, supporting premium pricing.
- Downside case: Customer overordering and new fabs create a glut by 2025.
- Execution risk: Higher costs and delays at overseas sites weigh on margins.
Signals in the Numbers
Guidance in recent quarters reflects a cautious stance. Capital expenditure has been framed as a band, not a big step-up, allowing room to adjust. The mix is shifting toward advanced nodes used in AI chips, while mature nodes grow more slowly. Utilization rates have improved from the 2023 slowdown, but the company has avoided declaring a new super-cycle.
Suppliers report steady orders for lithography, packaging, and advanced testing tied to high-performance computing. TSMC’s advanced packaging capacity—needed for chiplet designs—remains tight, yet management has paced expansions to match longer-term commitments rather than short spikes.
What It Means for the Industry
The company’s stance pressures rivals to be careful with their own expansion plans. Overbuilding could reset pricing across nodes, while shortages could hand share to competitors. Customers will likely continue to prepay or sign take-or-pay deals to secure supply, shifting some risk from the foundry to chip designers and cloud operators.
For policymakers courting fabs, the message is also clear. Subsidies help, but training, permitting, power reliability, and supply chains for chemicals and tools matter just as much. Without those, projects can slip and costs can rise, making cautious budgeting more likely.
The takeaway is measured optimism. AI demand is real and still growing, but memory of past cycles is fresh. TSMC is prepared to expand, yet it is building in checkpoints to avoid another painful glut. Investors should watch capital spending guidance, advanced packaging capacity, and the order cadence from major AI customers. If those hold, the company can grow with the trend without repeating old mistakes.
