C.C. Wei, the chairman of TSMC and the man whose factories make essentially every interesting chip you have heard of in the last three years, said on Wednesday that AI chip demand is still outpacing his ability to supply it. His phrasing, which is the kind of phrasing you get from an executive who has been asked the same question every quarter for two years, was: “Customer demand is so high, and we can only support so much. We are already working very hard.”
The customers in question are the ones whose stock tickers move when Jensen Huang says the word “platform” in any sentence. Nvidia and AMD are the two big AI-accelerator buyers competing for TSMC’s leading-edge nodes. Apple is also in there, eating its allocated wafers for the M-series chips and the A-series silicon in the iPhones that subsidize every other line item in Cupertino’s R&D budget. The combined demand is, again, larger than the supply, which is the same thing TSMC said last quarter and the quarter before that.
TSMC’s published answer to this problem is more fabs in more places. The company is expanding capacity in Taiwan, building out the Arizona facility that was the centerpiece of the 2022 CHIPS Act announcement, and breaking ground in Japan. None of this is fast. Leading-edge fab construction is a roughly three-year cycle from groundbreak to first wafer-out, and the equipment ramp behind that is its own multi-year process. The capacity coming online in 2028 was decided in 2025. The capacity coming online in 2029 is being decided right now in conference rooms in Hsinchu.
The structural takeaway is one TSMC has been quietly making at every earnings call: it is the bottleneck on AI compute, full stop. Every hyperscaler capacity-planning slide, every Nvidia revenue model, every frontier-lab training run is in the end a bet that TSMC builds enough fab capacity at the right node fast enough. Wei is not exaggerating when he says they are working very hard. He is also, structurally, asking the customers to be patient. They will not be.