Opinion: The so-called "10-year AI demand visibility" is a lie; storage, ABF carrier boards, and power are the real hidden bottlenecks.

Beating AI Express News: P Equity Research analyst Mr. P stated in a September 26 podcast that the widely cited claim that hyperscale cloud providers have "10-year demand visibility" is unconvincing, noting cloud firms can barely predict demand changes two years out, and AI infrastructure investment will ultimately remain constrained by cyclical spending. In the near term, AI inference demand will continue driving growth in demand for storage products including HBM, DRAM, and NAND. Mr. P projects next year’s capital expenditure (capex) by hyperscale cloud providers will reach $1.1 trillion to $1.2 trillion, with storage spending accounting for 50% to 60% of that total—equivalent to roughly $500 billion to $700 billion; UBS even forecasts this figure could hit $900 billion. Mr. P added that the current AI computing power bottleneck has shifted from a simple shortage of GPUs to other areas: power, advanced packaging, storage, and ABF substrates. Older-generation GPUs like the H100 still maintain high secondary market prices, while rental costs for B-series GPUs continue to rise. Supply tightness for ABF substrates is expected to persist until 2028 to 2030 or later, and orders for gas turbines from makers including Mitsubishi, Siemens, and GE Vernova are already booked past 2030. On the data center interconnect front, Mr. P predicts copper cables and optical communications will coexist for years to come. NPO is likely to see expanded adoption starting in 2027, while CPO may begin ramping up in 2028 to 2029, though mainstream adoption will not arrive until after 2030.