J.P. Morgan: Memory Stock Pullback Offers Prime Buying Opportunity
AI summary card
J.P. Morgan: Memory Stock Pullback Offers Prime Buying Opportunity
SOCAMM specification adjustments stem from supply constraints rather than weakening demand; NVDA-SKH multi-year partnership locks in demand through 2027; recommend accumulating Asian memory leaders on dips.
- Reduction in SOCAMM capacity from 192GB to 96GB is driven by supply shortages; total bit procurement demand remains unchanged
- NVIDIA designates SK hynix as its largest memory partner in a multi-year collaboration covering full-stack products
- DRAM bit demand projected to grow 33% and 34% in 2026 and 2027, respectively
- Computex confirms smooth HBM4 mass production; eSSDs emerge as critical infrastructure for AI inference
- Memory sector pulled back 11% over the past week, viewed as a medium-term allocation window
- Top picks: Samsung Electronics, SK hynix, Kioxia (all rated OW)
Report interpretation
Overview
This report addresses recent market concerns regarding reduced SOCAMM memory capacity in NVIDIA's Vera Rubin platform. Incorporating insights from Computex 2026 and the latest strategic partnership between NVIDIA and SK hynix, it reiterates a bullish stance on the memory semiconductor industry. J.P. Morgan believes current specification adjustments are primarily due to memory supply constraints rather than performance downgrades. Overall demand remains robust, and the recent stock price pullback presents an attractive entry point for medium-term investors.
Core views
The Truth Behind SOCAMM Specification Adjustments: The market fears that reducing SOCAMM capacity in Vera Rubin NVL72 racks from 192GB to 96GB will weaken CPU-related memory demand. Supply chain checks indicate this change is mainly due to tight memory supply, not product performance degradation. In fact, demand for 96GB modules is growing, and overall SOCAMM bit procurement demand remains unchanged. This implies reduced per-unit capacity will be offset by increased module counts, keeping CPU-driven memory demand a core driver of industry acceleration in 2027. Deepening NVIDIA-SK hynix Ties: The parties announced a multi-year technology partnership, officially designating SK hynix as NVIDIA's largest memory supplier. Collaboration extends beyond data center HBM to include Vera Rubin supercomputers, Vera CPUs, RTX Spark PCs, and Jetson Thor robotics platforms. This full-stack deep cooperation extends demand visibility through 2027 and benefits other memory manufacturers as supply tightness is unlikely to ease soon. NVIDIA CEO Jensen Huang noted that even doubling DRAM wafer capacity over the next five years may still fall short of meeting AI demand. Computex 2026 Industry Validation: The event sent positive signals. SK hynix confirmed 12-high HBM4 mass production is on schedule, with strong market interest in 12-high 48GB HBM4E supporting 16Gbps. Solidigm emphasized storage has become critical infrastructure for AI data pipelines; offloading KV Cache to NAND improves Time To First Token (TTFT) by 27x. Through AiSAQ software collaboration with NVIDIA, Kioxia offers differentiated solutions for hyperscale RAG servers, where software ecosystem stickiness distinguishes it in NAND competition.
Analysis framework
The report employs a 'supply chain cross-validation + sensitivity analysis' methodology to counter market noise. First, frontline supply chain research distinguished 'specification downgrades' from 'supply rationing,' confirming demand remains intact. Second, sensitivity analysis modeled bit demand under various SOCAMM configurations, quantitatively proving total demand remains constant even with all-96GB adoption. Finally, combining firsthand expo intelligence with long-term agreements among tech giants validates the 'Higher for Longer' upcycle logic across both technological evolution and commercial binding dimensions.
Methodology notes
Differentiating Specification Adjustments Under Supply Constraints
When end-product specifications are scaled back, it is essential to distinguish between 'configuration cuts due to weak demand' and 'passive adaptation due to supply shortages.' By verifying unchanged total bit procurement volume, this report determines the SOCAMM capacity reduction falls into the latter category, concluding demand has not weakened.
Bit Demand Sensitivity Analysis
When hardware specs change, analysis should not focus solely on individual capacity changes but calculate total bit demand as 'quantity × capacity.' By modeling different 96GB/192GB mix ratios, the report proves that even if unit capacity halves, increased module counts fully offset any loss in total bit volume.
Integrated Hardware-Software Ecosystem Barriers
In standardized hardware competition, software ecosystems are key differentiators. The report highlights Kioxia-NVIDIA SSD software collaboration, arguing such deep integration beyond pure hardware supply significantly enhances customer stickiness, creating asymmetric advantages versus competitors.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- SK hynix (000660.KS)NVIDIA's largest memory partner; joint full-stack product development
- Strengths
- Leading HBM4 mass production progress; secured NVDA multi-year exclusive-tier partnership endorsement; demand visibility extended to 2027
- Comparison
- First to obtain NVDA full-stack memory cooperation certification vs. Samsung and Micron
- Risks
- HBM yield volatility; geopolitical risks
- Samsung Electronics (005930.KS)Major memory supplier; beneficiary of overall supply tightness and AI factory buildout
- Strengths
- Large-scale capacity; participation in Korea AI factory infrastructure buildout; HBM4 certified
- Comparison
- Core NVDA supplier alongside SK hynix, though slight differences exist in certification pace for specific high-end products
- Risks
- Cyclical volatility in legacy memory business
- KIOXIA Holdings (285A.T)Key NAND beneficiary; AI inference storage infrastructure provider
- Strengths
- Deep AiSAQ software integration with NVDA; RAG server solutions offer high stickiness and cost advantages
- Comparison
- Software ecosystem collaboration creates differentiated moat vs. pure-play NAND hardware vendors
- Risks
- Intensifying NAND market competition; AI inference demand rollout below expectations
Key data
- DRAM Bit Demand Growth Forecast+33% / +34%J.P. Morgan's forecast for global YoY DRAM bit demand growth in 2026 and 2027
- Recent Memory Sector Decline-11%Average pullback in memory stocks over the past week; Philadelphia Semiconductor Index (SOX) fell 6% over same period
- NAND KV Cache Offload Performance Gain27xOffloading KV Cache to NAND yields 27x faster Time To First Token (TTFT) vs. recomputing context
- NVDA-SKH Partnership Duration2+ YearsBase contract term for SK hynix designated as NVIDIA's largest memory partner, including extension options
Impact & implications
For the memory semiconductor industry, recent sentiment volatility has obscured strong fundamentals. The report argues SOCAMM spec adjustments reflect necessary industry responses to supply shortages, not peak demand signals. A series of partnerships between NVIDIA and Korean tech giants further solidifies Asia's central role in global AI infrastructure. Investors should ignore short-term noise, focus on the AI-driven 'higher for longer' memory upcycle, and use pullbacks to accumulate leading companies with core technological moats and ecosystem advantages.
Risks
- Escalating geopolitical tensions could disrupt supply chain stability
- Further memory content reductions or optimizations cannot be ruled out
- Actual realization of AI memory demand may lag expectations
What to watch
- Updates to 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)
- 3Q26 contract price negotiations
- Micron Technology (MU) May-quarter earnings and guidance (expected late June)