P Equity Research Analyst: Hyperscalers' 10-Year Compute Demand Visibility Is a 'Lie'; Memory to Stay Cyclical

On Sept. 26, Mr. P, an analyst at P Equity Research, said on the Logan Jastremski podcast that hyperscalers' claims of 10-year demand visibility are unreliable, as they cannot predict demand even two years out. He estimated hyperscaler capital expenditure could reach $1.1 trillion to $1.2 trillion in 2027, with memory spending alone at $500 billion to $700 billion.

Mr. P argued that memory will remain cyclical, and that long-term agreements (LTAs) may be overhyped. He noted Samsung has received requests for 10-year contracts, but industry experts warn such deals could be quietly canceled if demand deteriorates. Citing rental prices, he said compute remains tight, but the real bottlenecks are power, advanced packaging and memory, not raw GPU counts.

In the bill of materials, ABF substrates, PCBs and memory stand out with significant price increases. Mr. P said ABF substrate shortages could last beyond 2028 or even into the 2030s. Gas turbine backlogs are already booked into the 2030s. On optics, near-package optics (NPO) may appear in 2027, while CPO capacity ramps in 2028-2029, but CPO dominance may not arrive until after 2030.