Nomura believes market concerns about memory oversupply and Meta's external sale of computing capacity are exaggerated

Overview

This Nomura Quick Note focuses on two issues weighing on global memory sector share prices recently: the large-scale medium- to long-term investment plans of South Korean memory companies, and Meta's plan to sell idle computing capacity to external customers. The report's core judgment is that investor concerns about oversupply and an inflection point in AI hardware demand have both been overstated.

Core views

First, the capacity expansion of South Korean memory companies is not irrational investment forced by the government, but medium- to long-term preparation driven by severe shortages caused by AI, the early exhaustion of land/power/water resources at production sites, and the need for new cluster construction after 2035. Second, Meta's external sale of idle computing capacity does not mean declining demand for AI-related hardware; as data center scale expands, selling off-peak and seasonally idle capacity is a natural path to improving resource utilization and ROIC, and may create new AI demand by lowering the price per token.

Analysis framework

The report analyzes the issue through supply-demand cycles, capacity construction cycles, policy support, data center utilization, and analogies to cloud business models. The authors compare South Korea's investment plan with advanced industry incentive policies in the United States, Japan, and China, and compare Meta's external sale of computing capacity with Amazon's development of AWS and xAI's sale of computing capacity.

Methodology notes

• Industry supply-demand analysisMemory cycle and capacity construction cycle

  There is a long cycle from investment to mass production, so short-term announcements do not equal short-term supply release.

  The report points out that the first clean room in the Yongin Semiconductor Cluster is not expected to be completed until 2027年2月, nine years after project launch, and small-scale production may not be achieved until the end of 2027, so the impact of the new cluster on supply will take at least 5-10 years.

• Business model analysisCommercialization of idle computing capacity

  Data centers need capacity configured for peak demand, and idle resources during non-peak periods can be sold externally to improve utilization.

  The report compares Meta's external sale of computing capacity with AWS's early commercialization of excess data center capacity, arguing that this may improve Meta's ROIC rather than weaken AI hardware demand.

• Demand elasticity analysisJevons' law

  A decline in unit usage cost may stimulate growth in total demand.

  The report believes that an increase in external computing capacity supply may stabilize and lower the rising price per token caused by shortages, thereby creating new AI demand.

Asset mapping & comparison

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

• Global memory industry

  Core research subject

  Strengths

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

  Weaknesses

  Historically, the memory industry has experienced overinvestment and cyclical volatility, and rising prices may also suppress smartphone and PC demand.

  Comparison

  The current situation differs from past cycles driven purely by corporate overinvestment; the report emphasizes that AI demand, LTA, and more stable structural growth may reduce historical cycle risks.

  Risks

  If companies once again undertake record-high investment or demand slows, oversupply may still emerge.

• META PLATFORMS INC

  Relevant stock and observation target for AI computing demand

  Strengths

  It has large-scale data centers and internal AI/social network advertising computing demand, and externally selling idle computing capacity has the potential to improve ROIC.

  Weaknesses

  Its core business is not traditional cloud services, and there remains optionality and execution uncertainty over whether it will continuously sell computing capacity externally.

  Comparison

  The report compares it with Amazon AWS and xAI's commercialization of idle computing capacity, but notes that Meta was previously one of the few CSPs with both internal and external data center use cases that had not launched a cloud business.

  Risks

  If the market interprets the external sale of computing capacity as insufficient internal AI demand, it may weigh on sentiment in the AI hardware chain in the short term.

• AI computing capacity and cloud service demand

  Driver of memory demand

  Strengths

  AI service providers such as Anthropic and OpenAI that lack proprietary data centers have strong demand for computing capacity, and external supply can support enterprise AI services.

  Weaknesses

  There is uncertainty among computing capacity supply, token pricing, and end-application demand.

  Comparison

  The report believes Meta's external sale of computing capacity is more like AWS commercializing idle capacity than a turning point in AI hardware demand.

  Risks

  If AI service demand grows less than expected, incremental demand for computing capacity and memory may fall short of market expectations.

Key data

• Total South Korea memory-related investment planKRW4.8 quadrillionThe combined amount of the medium- to long-term investment plans jointly announced by South Korean memory companies, affiliated companies, and the government.
• Directly related memory investmentKRW3.7 quadrillionThe scale of investment directly related to memory within the above plan.
• Time of the first clean room in the Yongin Semiconductor Cluster2027年2月The report says the first clean room of this large project is expected to be completed nine years after launch.
• Small-scale chip production time2027年底The report expects the Yongin project may only achieve small-scale production by the end of 2027.
• Impact cycle of new investment至少5-10年The report believes that even if the plan accelerates, the material impact on supply will only come at least 5-10 years later.
• Nomura Buy rating share57%Nomura Group's disclosed distribution of global equity research ratings as of 2026-03-31.
• Nomura Neutral rating share41%Disclosed rating distribution as of 2026-03-31.
• Nomura Reduce rating share2%Disclosed rating distribution as of 2026-03-31.

Impact & implications

If Nomura's judgment is correct, South Korea's medium- to long-term investment announcements will have limited negative impact on near-term memory prices and the earnings cycle, while AI-driven shortages in HBM and commodity memory may continue to support industry strength. If Meta expands its external sale of computing capacity, it may improve utilization, lower AI service costs, and stimulate more AI application demand, which is not necessarily negative for AI hardware and memory demand.

Risks

• South Korean memory companies may still create future oversupply through excessive investment.
• If AI demand slows, shortages in HBM and commodity memory may ease, putting pressure on prices and earnings.
• End markets such as smartphones and PCs may see demand decline due to rising memory prices and production disruptions.
• Consumer antitrust litigation and controversy over supply control may increase regulatory and reputational risk for the industry.
• There is uncertainty around the commercialization pace, customer demand, and pricing impact of Meta's external sale of computing capacity.

What to watch

• Whether South Korea's KRW4.8 quadrillion investment plan will provide a clear timetable and project rollout pace.
• The progress of the Yongin Semiconductor Cluster toward clean room completion and small-scale production in 2027.
• Supply-demand gaps, price trends, and capacity expansion execution for HBM and commodity memory.
• The customers, pricing, scale, and ROIC impact of Meta's external sale of idle computing capacity.
• Whether the price per token stabilizes or declines as computing capacity supply increases, and whether this brings new AI demand.
• Whether new mechanisms such as LTA can reduce the risks of overinvestment and price cycles in the memory industry.