Nomura: Concerns about global memory oversupply are exaggerated; AI demand remains the key support

Overview

The report focuses on two market concerns affecting memory stocks: first, whether the large long-term investment plans announced by Korean memory companies and the government will lead to supply overshoot; second, whether Meta’s announcement to sell surplus compute to external customers implies that AI-related memory demand has peaked. Nomura’s core view is that both concerns are being exaggerated by the market.

Core views

First, although Korea’s memory investment plan is very large in scale, it lacks a clear timeline, and building new clusters involves long lead constraints from land, electricity, water resources, and clean-room construction, so the actual supply impact is at least 5-10 years away. Second, current memory shortages are mainly driven by strong AI demand and priority production allocation to HBM, with commodity memory supply growth lagging, rather than by intentional supply restraint by producers. Third, Meta’s sale of surplus compute is more like a natural extension of a cloud-business model, which can improve resource utilization and reduce AI service costs, and should not be interpreted as AI hardware demand weakening.

Analysis framework

The report applies an integrated framework covering memory industry supply-demand cycles, capex construction cycles, government industrial policy, data center utilization, and AI compute economics, and addresses market concerns one by one. The emphasis is not on a single company earnings forecast, but on whether these events alter the directional supply-demand balance and AI hardware demand trend.

Methodology notes

• Industry supply-demandMemory cycle analysis

  Time mismatch between supply expansion and demand shocks

  The report attributes the current shortage to AI demand exceeding expectations, HBM being prioritized in production, and slower growth in commodity memory supply, and notes that past cycles have had both underinvestment leading to shortages and overinvestment leading to oversupply.

• Capital expenditureCapacity buildout lead-time analysis

  Several years lag from cluster investment to mass production

  Using the Yongin Semiconductor Cluster as an example, even nine years after project launch, the first clean room is expected to be completed only in February 2027, with limited production possibly starting by the end of 2027, indicating that large investments do not rapidly translate into supply.

• AI economicsJevons law

  Lower unit usage costs can stimulate total demand growth

  The report suggests Meta’s surplus compute sales may ease compute scarcity and help stabilize or reduce per-token prices, thereby creating new AI demand rather than lowering AI hardware demand.

• Corporate fundamentals & financialsROIC improvement logic

  Commercializing idle compute improves capital efficiency

  Selling idle compute that emerges outside peak-demand allocation is viewed as a reasonable way to improve data center capital return for Meta.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Global memory industry

  Primary coverage focus of the report

  Strengths

  Strong AI demand, high profitability of HBM, tight commodity memory supply, and companies are actively expanding capacity.

  Weaknesses

  The sector has historically shown cyclical overinvestment and price volatility.

  Comparison

  A key difference from past cycles is that AI demand growth, new mechanisms such as LTA, and more durable structural growth may reduce traditional oversupply risk.

  Risks

  If companies again over-expand capex, or if AI demand falls short of expectations, supply overshoot could still occur.

• Korean memory companies

  Entity driving market concerns over large investment plans

  Strengths

  Supported by Korean government policy and tax incentives, with a strong industrial foundation for long-term new cluster development.

  Weaknesses

  New clusters face constraints from land, electricity, water resources, and construction cycles, making short-term capacity contribution difficult.

  Comparison

  Comparable to advanced industrial support policies in the United States, Japan, and China.

  Risks

  If the investment plan is executed too aggressively during demand slowing, it may increase mid-to-long-term supply-release uncertainty.

• Meta (META US, Not rated)

  Subject of the surplus compute sale event

  Strengths

  Can improve data center utilization and ROIC, and provide compute sources for external AI customers.

  Weaknesses

  Core business is not cloud computing; external compute sales still carry execution and commercialization uncertainty.

  Comparison

  The report likens this to Amazon’s path of turning idle data center capability into AWS cloud services.

  Risks

  If the market misreads external compute sales as weaker internal AI demand, AI hardware sentiment could be pressured in the short term.

Key data

• Total Korea investment plan sizeKRW 4.8 trillionThe combined long-term investment plan announced by Korean memory companies, their affiliates, and the government.
• Memory-related direct investmentKRW 3.7 trillionThe portion of the total plan directly linked to memory.
• Yongin Semiconductor Cluster progressFirst clean room expected to be completed by February 2027; limited production may start by late 2027Used by the report to illustrate that the time from investment to output can exceed 10 years.
• Time until new capacity impactAt least 5-10 yearsThe report believes that even with accelerated execution, the supply impact of these investments will not materialize quickly.
• Meta and Amazon caseMETA US, AMZN US are both Not ratedThe report compares Meta’s external sales of surplus compute to Amazon’s early commercialization of idle data center capacity through its cloud business.

Impact & implications

The report’s implication for the memory industry is constructive: near-term concerns about supply overshoot and an AI demand turning point may be excessive, while the industry remains in an AI-driven environment of strong demand and supply tightness. If compute availability increases and lowers AI usage costs, AI application demand may expand and continue to support both HBM and memory demand.

Risks

• Excessive medium-to-long-term investment by memory firms leading to oversupply.
• AI demand growth below expectations, weakening demand for HBM and commodity memory.
• If Korea’s investment plan is accelerated by policy, it may increase uncertainty around future capacity releases.
• Antitrust investigations and price-collusion allegations could cause regulatory and sentiment disruptions.
• Meta’s monetization path, pricing, and customer demand for compute sales still contain uncertainty.

What to watch

• Specific timelines, site selection, policy subsidies, and tax incentives in Korea’s memory investment plan.
• Progress on clean room completion and limited production at the Yongin Semiconductor Cluster.
• Changes in capacity allocation between HBM and commodity memory.
• Whether commodity memory prices, inventories, and lead times continue to reflect shortages.
• The impact of Meta’s compute sales on per-token pricing and compute availability for customers such as Anthropic and OpenAI.
• Whether AI capex continues to support data centers, HBM, and memory demand.