Nomura rebuts 2 bearish narratives on memory chip stocks

Nomura argues that fears over Korean memory oversupply and Meta's compute rental misread the structural demand picture for semiconductors.

Nomura said fears of a memory chip glut from Korea's massive expansion plan and Meta's decision to rent out idle computing power are overblown, arguing both narratives misunderstand the structural supply-demand dynamics in AI infrastructure.

"These concerns reflect a misunderstanding of both the timeline of capacity additions and the nature of AI compute demand," the Nomura team wrote in a July 2 research note.

Korea's ~4,800 trillion won investment plan — with ~3,700 trillion won earmarked for memory — would take 5 to 10 years to translate into actual production capacity, Nomura said. The Yongin Semiconductor Cluster, a mega-project launched nine years ago, expects its first clean room only by February 2027, with small-scale production by year-end. Meanwhile, high-bandwidth memory production is crowding out general-purpose DRAM supply, creating acute shortages rather than surpluses.

The stakes are significant for investors in Samsung Electronics, SK Hynix, and Micron Technology, whose shares have been rattled by supply fears. If Nomura is correct, the recent selloff — Micron fell more than 10% on July 1 alone — represents a mispricing opportunity rather than the start of a cyclical downturn.

Meta's Compute Rental Points to Maturation, Not Peak Demand

The second bearish narrative — that Meta's plan to sell idle GPU capacity indicates peak AI hardware demand — also misreads the situation, Nomura said. Meta is the only major cloud-scale operator without a cloud business, and renting out spare capacity mirrors Amazon's founding of AWS to monetize idle data center infrastructure.

Meta's reported "Meta Compute" unit, led by infrastructure head Santosh Janardhan alongside Meta Superintelligence Labs figure Daniel Gross and president Dina Powell McCormick, would sell excess AI compute to external clients. The company has guided to $115 billion to $145 billion in 2026 capital spending, with a 2,250-acre hyperscale campus in Louisiana and a gigawatt-scale data center under construction in the Midwest.

Rather than indicating demand exhaustion, Nomura argued the move could trigger the Jevons paradox — the economic principle where lower resource costs stimulate greater total consumption. By making token prices more affordable, Meta's spare capacity could unlock new AI use cases that ultimately increase demand for memory and compute hardware.

The Distribution Problem Creates Winners and Losers

Commentator Jay Yoon offered a complementary framing: the aggregate market remains undersupplied, but specific players who overbuilt relative to their own model demand are stuck with expensive silicon. "We are still massively short compute," Yoon said. "It's a compute allocation problem."

CoreWeave, the neocloud provider that counts Meta as a $21 billion customer, saw its stock drop 14% on the news. The company reported Q1 revenue growth of 111.6% year over year with a backlog near $99 billion, but its net loss widened to $740 million. If Meta's entry compresses pricing for neoclouds, CoreWeave's debt-heavy model faces pressure.

For investors, the key signal to watch is Meta's CapEx guidance. A raise would confirm the arms race narrative — bullish for NVIDIA and memory makers. A cut would validate the bear case. Alphabet's cloud revenue of $20.03 billion, growing 63% year over year with a backlog above $460 billion, provides the benchmark for what real hyperscale traction looks like.

NVIDIA shares have fallen 13.5% over the past month. Samsung and SK Hynix each dropped more than 7% and 9% respectively in early July trading, with South Korea's KOSPI triggering a trading halt. If Nomura's analysis holds, the selloff has overshot the fundamentals.

This article is for informational purposes only and does not constitute investment advice.