P Equity Research analyst Mr. P said on a Sept. 26 podcast that the market's claim that hyperscale cloud providers have "10 years of demand visibility" is not credible. He said cloud companies have difficulty forecasting demand changes even two years ahead, and AI infrastructure investment will ultimately remain tied to cyclical spending constraints.
Memory demand is still expected to rise with inference growth
Mr. P said near-term AI inference demand should continue to drive demand for memory products, including HBM, DRAM and NAND.
He estimates hyperscale cloud providers will spend $1.1 trillion to $1.2 trillion in capital expenditure next year. Of that, memory could account for 50% to 60%, or about $500 billion to $700 billion. UBS has gone further and estimated the number could reach $900 billion.
Bottlenecks are moving beyond GPU counts
In Mr. P's view, the main constraint in AI compute is no longer just the number of GPUs available. The bottleneck is shifting toward power, advanced packaging, memory and ABF substrates.
He said older-generation GPUs such as the H100 still command high prices in the secondary market, while rental rates for B-series GPUs continue to move higher. ABF substrate supply tightness may last from 2028 to beyond 2030. On the power side, order books for gas turbine makers including Mitsubishi, Siemens and GE Vernova are already filled past 2030.
Copper and optical links are likely to coexist for years
On data center interconnects, Mr. P said copper cables and optical communications are likely to coexist over the next several years.
He expects NPO to scale first in 2027. CPO may begin ramping in 2028 to 2029, but it may not become mainstream until after 2030.