Stacy Rasgon: “Demand Is Off The Charts” in Semiconductors… And Set To Double Again Soon

Semiconductor fundamentals are improving sharply: sector earnings are up more than 100% year to date, while the sector index has risen about 80%. Forward earnings estimates have climbed since June even as share prices fell, compressing valuation multiples. Stacy Rasgon says investors’ concern that the cycle is near a peak may help explain why earnings beats are receiving less reward. He does not expect a peak this year or next, and says even 2028 or 2029 does not currently look likely, though visibility weakens further out.

Across AI-related companies, the recurring message is that customers cannot get enough compute. Rasgon says demand could support NVIDIA revenue growth of 100% next year, above the company’s expectation of at least 70%; he estimates NVIDIA could approach $700 billion in revenue. He says the company’s outlook reflects the land, power and data-center shells expected to be ready. Broadcom has forecast 100% growth next year and the year after, and expects $115 billion in AI revenue next year; Rasgon thinks both companies could exceed their outlooks if demand holds.

The five largest hyperscalers are likely to spend about $1.5 trillion next year, Rasgon estimates, making Jensen Huang’s forecast of $3 trillion to $4 trillion in annual AI infrastructure spending by the end of the decade less far-fetched. Building at that scale requires coordination across companies and supply chains, leading customers to order further in advance. Rasgon says infrastructure constraints could temper the pace of spending but extend the cycle. He sees growing evidence that investment is producing returns, including neoclouds renting capacity and rapid revenue growth at model companies, while customer willingness to pay remains a central unresolved question for skeptics.

Rasgon says suppliers’ position near the back of the supply chain limits their view of end demand, and tight supply and long lead times can prompt customers to overorder before canceling excess orders. He sees no evidence of widespread unused GPU stockpiles and believes capacity under construction is generally being used; equipment waiting for power or other infrastructure is mainly a timing issue. Power is a major constraint on adding tens or hundreds of gigawatts of capacity annually, and local opposition to data centers will persist in some places. He argues it will not stop construction, which can also take place outside the United States.

Memory supply is already tight in 2026, with companies indicating conditions could tighten further in 2027 and 2028. High-bandwidth memory uses stacked DRAM dies, with yield losses, and a gigabyte of HBM requires three to four times as many wafers as a gigabyte of DDR5, limiting capacity for other DRAM. AI infrastructure is also increasing NAND demand. After a severe downturn in which some memory companies had negative gross margins, the industry added little capacity and remains cautious; new factories take years, and the effect of material capacity additions expected from 2027–2028 on pricing is uncertain.

Rasgon forecasts wafer-fabrication-equipment spending of about $154 billion in 2026, $204 billion in 2027 and $256 billion to $258 billion in 2028, with further growth possible if data centers are built. Clean-room shortages and equipment-maker capacity may limit near-term shipments. He favors semiconductor equipment stocks, leaning slightly toward Applied Materials over Lam Research and KLIC because of Applied Materials’ greater DRAM exposure and lower valuation, while saying he could own any of them. He also favors traditional AI names and AMD. He expects investors’ rotation among memory, equipment, networking, optics, power semiconductors and CPUs to normalize, though NVIDIA’s size can constrain portfolio buying.

Competition is a frequent investor concern, but Rasgon says NVIDIA’s software, broad ecosystem and ability to invest in ecosystem participants make threats appear greater than they currently are relative to the opportunity. He feels better about Intel than he has in a long time, citing strong server demand, cost changes and CEO Lip-Bu Tan’s approach, but expects material foundry revenue to take years. High channel inventories and the sector’s long stretch without a typical cycle make traditional warning signs harder to interpret. Rasgon says today’s buildout is responding to existing demand, unlike the buildout ahead of demand around 2000; if there is a bubble, he sees more risk in unsustainable earnings than in unusually high valuation multiples.