J.P. Morgan: SOCAMM Noise Leads to Oversold Memory Stocks; Pullback Offers Prime Buying Opportunity

Overview

J.P. Morgan released a memory market update clarifying concerns over NVIDIA's Vera Rubin platform SOCAMM memory spec reductions. Combining insights from Computex 2026 and the latest collaboration between NVIDIA and SK Hynix, the firm reiterates its optimistic view on the memory sector. The core conclusion is that recent spec adjustments are primarily due to rationing under supply constraints, not a deterioration in fundamental demand. The current ~11% pullback in memory stocks offers a good entry window for medium-term investors, with the logic of an AI-driven 'higher for longer' memory upcycle remaining intact.

Core views

The Truth Behind SOCAMM Spec Adjustments: Market concerns that the reduction in SOCAMM capacity for Vera Rubin NVL72 racks from 192GB to 96GB signals cooling CPU-related memory demand are misplaced. Supply chain checks indicate this change is mainly due to rationing adjustments caused by tight memory supply, not performance downgrades. In fact, demand for 96GB modules is growing, and overall SOCAMM bit procurement demand has not changed. This means the lower per-module capacity will be offset by an increase in module count, keeping total demand robust. The report emphasizes that CPU-driven memory demand remains the core driver of industry acceleration through 2027. Deepening NVDA-SK Hynix Ties: NVIDIA officially designated SK Hynix as its largest memory partner, signing a technology cooperation agreement for over two years (with renewal options). The scope extends beyond data center HBM to cover full-stack products including Vera Rubin supercomputers, Vera CPUs (SOCAMM2), RTX Spark PCs (LPDDR5X), and Jetson Thor robot platforms. The report believes this long-term agreement extends demand visibility to 2027 and, against a backdrop of limited supply, benefits other memory manufacturers as well. Furthermore, NVIDIA CEO Jensen Huang stated that plans to double DRAM wafer capacity over the next five years may still be insufficient, further validating the strong outlook for AI memory demand. Computex 2026 Industry Validation: The event released several positive signals. For SK Hynix, 12-layer HBM4 mass production progress met expectations, while 12-layer 48GB HBM4E gained attention for supporting 16 Gbps speeds. They also showcased new AI-N B product concepts combining HBM throughput with SSD capacity. Solidigm noted that storage has become critical infrastructure for AI data pipelines, with offloading KV Cache to NAND improving Time To First Token (TTFT) latency by up to 27x. Kioxia, through deep collaboration with NVIDIA on AiSAQ software, provides cost-effective solutions for hyperscale RAG servers. The report views this differentiated capability of combining hardware and software as a key advantage distinguishing it from competitors.

Analysis framework

The report employs a 'noise identification + supply chain cross-validation' analytical approach. First, regarding emotional market rumors about spec cuts, it looks beyond surface parameter changes to deeply verify 'volume-price ratios' and 'total procurement volumes' within the supply chain, distinguishing the essential difference between 'passive rationing' and 'active downgrading.' Second, it triangulates findings through on-site research at Computex and announcements of major strategic partnerships, placing single-product technical iterations (e.g., HBM4, eSSD) within the macro narrative of AI infrastructure construction to confirm whether technology rollout pace matches commercialization demand. Finally, by combining short-term stock price volatility (-11%) with medium-to-long-term fundamentals (demand +33%/34%), it derives a contrarian investment timing judgment.

Methodology notes

• Industry/Industrial Analysis FrameworkSupply-demand framework

  Distinguishing whether spec changes stem from supply constraints or demand contraction

  In strongly cyclical industries like semiconductors, when end-product specs are adjusted downward, one must judge if it is 'down-specification' due to weak demand or 'rationing' due to upstream shortages. If the latter, it often accompanies sustained high unit prices and unchanged or even rising total volume demand, signaling continued prosperity rather than a turning point. This report uses this logic to refute the peak-demand thesis.

• Competition and Strategy FrameworkMoat / competitive advantage

  Customer stickiness built through integrated hardware-software solutions

  In standardized hardware competition (e.g., NAND), competing solely on parameters easily leads to price wars. The report notes Kioxia's deep adaptation with NVIDIA via AiSAQ software. This 'software-defined storage' capability increases customer switching costs, upgrading hardware manufacturers from mere component suppliers to system-level solution partners, thereby gaining premium pricing power and share stability beyond peers.

• Event Game Theory and Behavioral FinanceExpectation Gap/Expectation Management

  Finding pricing discrepancies by exploiting the market's linear misreading of technical parameters

  Markets often overreact to explicit negative information (e.g., halved specs) while ignoring implicit hedging factors (e.g., increased module count). Institutions obtain more granular information than public news through frontline research, identifying expectation gaps where 'surface negatives are actually neutral or even positive,' and positioning contrarily when stock prices deviate from fundamentals due to emotional selling.

Asset mapping & comparison

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

• SK hynix (000660.KS)

  NVIDIA's largest memory partner, leading in HBM4 mass production, deeply bound to AI computing ecosystem across full-stack product lines

  Strengths

  2+ year exclusive-level cooperation agreement locks in demand; 12-layer HBM4 mass production progress meets expectations; technical reserves in frontier forms like AI-N B

  Comparison

  Compared to Samsung and Micron, it secured NVIDIA's full-stack cooperation endorsement first, offering the strongest demand certainty

  Risks

  HBM yield ramp-up slower than expected; geopolitical risks affecting supply chain

• Samsung Electronics (005930.KS)

  Global memory leader, benefiting from overall industry supply shortages and NVIDIA's diversified supply chain strategy

  Strengths

  No. 1 global production capacity; 12-layer HBM4E samples shipped to customers; vertical integration advantages across entire DRAM/NAND lines

  Weaknesses

  Did not receive equivalent priority designation in NVIDIA's latest cooperation list

  Comparison

  Although slightly behind SKH in HBM certification progress, its massive capacity makes it a core beneficiary amid supply shortages

  Risks

  Advanced process yield issues; weak consumer electronics recovery dragging traditional storage business

• KIOXIA Holdings (285A.T)

  Core NAND/eSSD supplier, entering AI RAG server supply chain through software differentiation

  Strengths

  AiSAQ software deeply adapted with NVIDIA, building a hardware-software moat; focused on NAND track with high proportion of enterprise SSDs

  Weaknesses

  Lacks DRAM/HBM business, unable to enjoy AI training-side memory dividends

  Comparison

  Compared to pure hardware NAND manufacturers, software cooperation gives it greater stickiness and bargaining power in the AI inference storage market

  Risks

  Severe NAND price volatility; operational coordination risks with WD joint venture factories

Key data

• DRAM Bit Demand Growth Forecast+33% / +34%J.P. Morgan's forecast for global DRAM bit demand YoY growth rates in 2026 and 2027
• Recent Decline in Memory Sector-11%Average pullback of memory stocks over the past week, compared to only a 6% drop in the Philadelphia Semiconductor Index (SOX)
• NVDA-SKH Cooperation Term2+ YearsSK Hynix designated as NVIDIA's largest memory partner; agreement includes renewal options, extending demand visibility to 2027
• Performance Boost from Offloading KV Cache to NAND27xSolidigm data shows that offloading KV Cache to NAND can make Time To First Token (TTFT) 27x faster compared to recomputing context
• Vera Rubin SOCAMM Baseline Configuration192GB → 96GBSpec adjustment of market concern, but report notes total bit procurement volume is unchanged, attributed to supply rationing

Impact & implications

For the memory industry, debunking the SOCAMM noise removes the biggest short-term uncertainty, while the long-term binding between NVDA and SKH provides a demand anchor spanning cycles. The report specifically points out that eSSD's status as critical infrastructure in the AGI era was validated at Computex, which holds significant implications for the medium-to-long-term valuation reshaping of NAND manufacturers. For investors, the current pullback is seen as a prime opportunity for medium-term positioning. It suggests focusing on the 'higher for longer' upcycle logic, particularly targeting Asian leaders with advantageous positioning in HBM, enterprise SSDs, and advanced packaging. Close monitoring of CSP long-term contract signings, actual shipment rhythms of Vera CPUs, and Q3 contract price trends is needed to verify demand realization.

Risks

• Escalation of geopolitical tensions leading to supply chain disruptions or increased trade restrictions
• Potential for additional memory content cuts or configuration optimizations in the future
• Slowing AI capital expenditure leading to memory demand missing expectations
• HBM4 and advanced process yield ramp-up slower than expected

What to watch

• Updates on Long-Term Agreements (LTAs) between Cloud Service Providers (CSPs) and memory suppliers
• Memory procurement and demand dynamics related to x86 and ARM CPUs (including Vera CPU)
• Results of memory contract price negotiations in Q3 2026
• Micron (MU) May quarter earnings and outlook (expected late June)