"I've Never Seen a Demand Cycle Like This" | Top Semis Analyst Stacy Rasgon on AI Chips

Rasgon says the top five hyperscalers will probably spend $1.5 trillion next year, making Jensen Huang’s $3 trillion to $4 trillion annual AI-infrastructure projection by decade-end seem less implausible, though the scale is making investors nervous. He points to early signs of returns—including Neocloud capacity rentals he estimates at $30 billion per gigawatt and roughly one- to one-and-a-half-year infrastructure payback—and says deployed AI hardware is being used immediately rather than stockpiled.

Rasgon argues that construction and power constraints could prolong the AI investment cycle: large deployments require advance coordination, while grids may not support additions of tens or hundreds of gigawatts annually. He forecasts wafer fabrication equipment spending rising from about $154 billion in 2026 to $204 billion in 2027 and roughly $256 billion to $258 billion in 2028, while saying clean-room and supplier capacity could limit installation timing and that unconstrained demand could support more than $300 billion in 2027.

Rasgon sees strong and strengthening semiconductor order visibility even though companies have little direct view of end-market demand, and he cautions that customer double ordering can make demand for 100 chips appear to suppliers as 1,000. He illustrates the limits of forecasting with the industry’s earlier projection of $1 trillion in revenue by 2030 versus his claim that it is on track for $1.5 trillion this year.

Rasgon says memory supply could be tighter in 2027 and 2028 than it is now, while whether new capacity is absorbed depends on real demand; he notes that memory accounts for about 80% of the wafer area in an Nvidia NVL72 rack. He links tight supply in part to HBM production, which requires three or four times as many wafers per gigabyte as standard DRAM.