Opinion: The So-Called '10-Year AI Demand Visibility' Is a Lie—Storage, ABF Substrates, and Power Are the Hidden Bottlenecks

动察 Beating AI News Flash: Mr. P, an analyst at semiconductor research firm P Equity Research, said on a September 26 podcast that the market's claim that hyperscale cloud providers have "10-year demand visibility" is not credible. Cloud providers find it very difficult to predict demand changes two years out, and AI infrastructure investment will ultimately still be constrained by cyclical spending.

In the short term, AI inference demand will continue to drive growth in demand for memory products such as HBM, DRAM, and NAND. It is expected that hyperscale cloud providers' capital expenditure next year will reach $1.1 trillion to $1.2 trillion, of which memory spending may account for 50% to 60%, about $500 billion to $700 billion. UBS even estimates that this figure could reach $900 billion.

Mr. P believes that the current AI computing bottleneck has gradually shifted from simply the number of GPUs to areas such as power, advanced packaging, memory, and ABF substrates. Older-generation GPUs such as the H100 still maintain relatively high second-hand prices, and rents for B-series GPUs also continue to rise. Supply tightness for ABF substrates may persist until after 2028 to 2030, and orders at gas turbine manufacturers such as Mitsubishi, Siemens, and GE Vernova have already been booked beyond 2030.

Regarding data center interconnection, Mr. P expects copper cables and optical communications to coexist for several years. NPO may take the lead in expanding applications in 2027, while CPO may begin ramping in 2028 to 2029, but truly becoming mainstream may have to wait until after 2030.