Table of Contents
TSMC projects 1.9x tool demand within months
In sum – what we know:
- A fivefold build-out – TSMC is building 25 wafer-fabrication and advanced-packaging facilities worldwide this year, including 13 in Taiwan, at roughly five times its historical pace.
- Tool demand nearly doubles – TSMC’s 2026 equipment-tool projection went from 1.5x its end-of-2025 baseline after Q1 to 1.9x by July, while 2026 capex rose only about 15% at the midpoint.
- Packaging is the pinch point – CoWoS, scaling from 5.5-reticle today to 14-reticle in 2028, plus COUPE optics in 2026, show the constraint is packaging and interconnect, not wafer starts alone.
TSMC is set for a dramatic expansion. News reports say the company is building 25 wafer-fabrication and advanced-packaging facilities around the world this year, including 13 in Taiwan. Speaking at SEMICON Taiwan in Taipei, TSMC Senior Vice President and Deputy Co-Chief Operating Officer Cliff Hou said the company is expanding capacity at roughly five times its historical pace, yet AI demand continues to exceed what it can supply. To be clear, not all 25 are fabs — there are 13 fabs in Taiwan plus five or six overseas, and the higher figure folds in advanced-packaging sites as well.
The scale of the acceleration shows up most clearly in TSMC’s own equipment math. Hou said the company’s end-of-2025 forecast for how much semiconductor-production equipment it would need in 2026 was already revised to 1.5 times that baseline after the first quarter, then to 1.9 times by July — nearly double in six months. He characterized the moment as demand growth at a scale and frequency the chip industry hasn’t seen in roughly 30 years.
The constraint isn’t limited to wafer starts, either. AI accelerators need leading-edge logic, high-bandwidth memory integration and advanced packaging, which means the squeeze extends across packaging, materials, tools, interconnects and energy efficiency. Taiwan remains the center of the build-out, with Arizona, Japan and Germany illustrating localization and geopolitical diversification — though skilled-labor shortages and a strained equipment supply chain are adding friction on every front.
Tooling demand
That 1.9x figure refers to TSMC’s internal projection of how many production tools it will need in 2026 relative to its forecast at the end of 2025 — 1.5 times the baseline after the first quarter, 1.9 times by July. Sourcing an almost-doubled volume of equipment is a genuine problem even for the world’s largest foundry, simply because the absolute quantities involved are so large.
Capital spending has risen too, but far less dramatically. TSMC raised its 2026 capex guidance in July to a range of US$60 billion to US$64 billion, up from the US$52 billion to US$56 billion it guided in January. That’s roughly a 15% bump at the midpoint, which doesn’t match the near-doubling in projected tool count — tool count and tool cost aren’t directly comparable, and the two figures measure different things.
TSMC reported 2Q 2026 revenue of NT$1.270 trillion and net income of NT$706.56 billion for the quarter ended June 30, 2026.
Driving forces
AI training and inference systems need three things at once — leading-edge logic chips, high-bandwidth memory integration and advanced packaging. That’s why TSMC’s push spans both wafer fabrication and packaging rather than process-node capacity alone. A faster transistor doesn’t help much if there’s nowhere to package it.
CoWoS is the clearest example. The technology (chip-on-wafer-on-substrate, TSMC’s method of placing compute dies and memory stacks side by side on a silicon interposer) is what lets an AI accelerator sit next to its HBM at all. TSMC is producing 5.5-reticle CoWoS packages today and, per its April technology symposium materials, plans a 14-reticle version for 2028 that could integrate about 10 large compute dies and 20 HBM stacks in a single package.
The same materials say TSMC’s co-packaged-optics technology, COUPE on substrate, is scheduled to begin production in 2026. TSMC positions it as improving power efficiency and latency for data-center interconnects — a vendor framing, but a telling one. The bottlenecks in AI capacity have spread well beyond wafer starts, into packaging, substrates, specialized tools, interconnects and the energy budget of the data center itself.
Geographic diversification
Taiwan still anchors the expansion, with 13 of the facilities Hou described under construction there. TSMC has previously said it’s preparing multiple phases of 2nm fabs in Hsinchu and Kaohsiung and will keep investing in both leading-edge and advanced-packaging capacity at home.
Arizona is the biggest bet overseas. TSMC’s 2025 annual report says its first Arizona fab entered high-volume production in the fourth quarter of 2024 with good yield, its second is expected to reach high-volume manufacturing in the second half of 2027, and construction of a third began in 2025. The company has framed the site as an eventual independent “GIGAFAB” cluster serving leading-edge smartphone, AI and HPC customers, part of a previously announced U.S. investment framework — up to US$165 billion covering six fabs when announced in March 2025, extended on July 16, 2026 by an additional US$100 billion to a total of up to US$265 billion covering 10 fabs, two advanced-packaging facilities and an R&D center. That framework predates the September remarks — it’s not a new commitment.
Japan and Germany fill out the map, with different jobs. TSMC’s first JASM specialty fab in Kumamoto began volume production at the end of 2024, and the annual report says the second JASM fab is planned for 3nm process technology, with volume production reportedly targeted for 2028. The Dresden ESMC project, by contrast, is aimed at mature automotive and industrial technologies — it diversifies the supply chain, but it isn’t a leading-edge AI facility and shouldn’t be read as one.
The overseas build-out is driven by customer demand, government pressure for localized production and plain geopolitical risk management. Foxconn Chairman Young Liu, speaking at the same event, described the shift as moving from “made in Taiwan” to “made with Taiwan.” The trade-off is cost. Overseas fabs generally run more expensive and are harder to coordinate than capacity concentrated in Taiwan, and the skilled-construction-labor shortage Hou flagged affects both Taiwan and the U.S.
