SemiAnalysis's Patel Says Packaging Substrates, Not Power, Are AI Infrastructure's Biggest Bottleneck

Packaging substrates are the tightest constraint on AI infrastructure expansion, ahead of memory, optical interconnect and power, SemiAnalysis founder Dylan Patel said at the CASPA annual CEO summit on October 23. He said underinvestment and conservative capacity expansion by Japanese materials suppliers could limit AI accelerator and CPU output for the next two years. Patel said power draws excessive market attention because operators can add gas turbines, reciprocating engines and fuel cells on site, and power capacity can be raised faster and with less complexity than semiconductor supply chains. Bernstein analyst Stacy Rasgon said high-bandwidth memory consumes three to four times the wafer capacity of standard DDR5, tightening supply across the industry. He said memory makers trade at forward price-to-earnings multiples of two to four times, reflecting fears that new capacity in 2028 will trigger a price decline. Patel estimated nearly 50 gigawatts of data center capacity will come online globally next year, about 35 gigawatts of it in the U.S. and 7 to 8 gigawatts in China.