Deputy Co-COO Cliff Hou put a number on the AI build-out at SEMICON Taiwan, and it is not a flattering one for anybody waiting on capacity.
TSMC is buying fab tools far faster than it told itself it would. Speaking at SEMICON Taiwan, Deputy Co-COO Cliff Hou said the company’s quarterly requirement for semiconductor production equipment has reached roughly 1.9 times what it forecast back in December 2025. TSMC equipment demand had already climbed to about 1.5x the original estimate during the first quarter, then kept going, hitting 1.9x by July.
That is a forecast miss of a kind you rarely see from a company this disciplined about capital planning. TSMC does not guess at tool orders casually; the lead times on advanced lithography and deposition equipment run into years, and the December number was supposed to be the conservative version of a boom.
Twenty fabs, and not enough people to build them #
The company is putting up close to 20 wafer fabs simultaneously, thirteen of them in Taiwan and another five or six spread across overseas sites. Hou flagged two bottlenecks, and neither is money. One is equipment delivery: the tool vendors cannot ship fast enough. The other is more mundane and harder to fix, a persistent shortage of construction workers. You cannot install an EUV scanner in a building that has not been poured.
The capital number moved to match. TSMC lifted its 2026 capex outlook to $60 billion to $64 billion, up from earlier guidance of $52 billion to $56 billion, which makes it the largest single-year revision in the company’s modern history. Between 70 and 80 percent of that is earmarked for advanced nodes, with 3 nm taking the biggest share.
Why a PC builder should care #
Because it is the same fabs. The AI accelerators, advanced logic and packaging capacity soaking up this spend come off leading-edge lines that also make the CPUs and GPUs in gaming PCs, and packaging capacity in particular has been the constraint on high-end graphics parts for a couple of years now. Add the memory situation that has already doubled some consumer SSD prices this year, and the picture for 2027 desktop pricing is not encouraging.
There is a second-order effect worth watching too. If tool deliveries and construction are the limit rather than demand, then every foundry customer is queuing behind the same physical constraint, and the ones with the deepest pockets and the biggest volume commitments get served first. Consumer silicon has historically not been at the front of that queue.
The caveats #
Capex is not capacity. Tools ordered are not tools installed, and a fab shell finished in 2027 does not produce wafers until it is qualified, which takes the better part of another year. Hou’s 1.9x figure describes demand for equipment, not output, and the gap between those two is exactly the problem he was describing.
It is also worth remembering that TSMC’s own guidance has swung before, in both directions, and that a revision this large is as much a statement about how wrong December looked in hindsight as it is a prediction about next year. What the number does establish is that the company with the best visibility into leading-edge demand in the industry did not see this coming nine months ago.
Sources: Tom’s Hardware and SemiMedia.