RESEARCH MEMO · AI Infrastructure | Memory and HBM
AI Needs More Than GPUs: Which Memory Layer Is Truly Scarce?
AI Memory Industry Deep Dive | 20 September 2026
HBM is the most constrained, technically difficult and pricing-power-rich layer of AI memory. Server DRAM remains tight as AI systems expand and HBM absorbs manufacturing resources. Enterprise SSD demand benefits from training data and inference caches, while consumer NAND may return to looser supply in 2027. China can already supply parts of the DRAM and NAND stack used in AI data centers; leading-edge HBM remains the clearest gap.
Samsung, SK hynix and Micron held about 87.6% of 2Q26 global DRAM revenue
TrendForce estimates 2026 DRAM demand modestly exceeds supply
New NAND capacity may restore pricing pressure in consumer markets
CXMT held roughly 7.7% of 2025 DRAM share, while leading-edge HBM remains unproven
HBM first | Server DRAM tight | NAND diverging
01 · RESEARCH BRIEF
The one-minute brief
Memory is no longer one homogeneous industry. The closer a product sits to the GPU, the higher its bandwidth, value content and qualification burden. HBM keeps accelerators fed; server DRAM provides the system workspace; enterprise SSDs hold models, datasets, checkpoints and colder data. The investment map must also be layered: SK hynix and Micron Technology have the most direct HBM and server-memory exposure; Samsung Electronics is the broadest full-stack supplier but offers lower AI-memory purity; Kioxia Holdings and SanDisk are closer to enterprise SSD and NAND cycles; CXMT and privately held YMTC represent localization, where technical progress, earnings conversion and valuation must be judged separately.
Audio transcript
This AI memory cycle cannot be treated as one industry. HBM is the scarcest layer, server DRAM remains tight, enterprise SSDs benefit from AI data growth, and consumer NAND may face looser supply in 2027. China is not absent from AI memory: commodity DRAM and NAND are scaling, while leading-edge HBM remains the clearest gap. For investment research, HBM is about technology and customers, NAND is about supply and pricing, and Chinese memory is about execution and valuation.
Known facts and open questions
- Scope
- HBM, server DRAM, enterprise SSDs, consumer NAND and China localization
- As of
- 20 September 2026
- Evidence
- Industry supply data, product milestones, production and listing disclosures
- Boundary
- Industry direction, corporate benefit and security valuation are separate questions
NAND and SSD | Data warehouse
Non-volatile storage for models, datasets, checkpoints and archives; largest capacity and lowest cost, but slowest speed
DRAM | System workbench
Fast working memory for CPUs and servers; mid-range speed and cost
HBM | High-speed feeder beside the GPU
Stacked DRAM placed close to accelerators; highest bandwidth, cost and technical barriers
Interconnect and packaging | Value multiplier
TSVs, advanced packaging, thermals and joint qualification determine whether HBM can ship at scale
02 · THESIS → EVIDENCE → UPDATE
What changed in the thesis?
The bottleneck shifts from chip count to data delivery
- Prior thesis
- AI infrastructure research focused mainly on GPU supply and compute expansion.
- New evidence
- Larger models, longer contexts and more inference concurrency make data movement a determinant of accelerator utilization.
- Updated view
- HBM, server DRAM and enterprise SSDs deserve separate, recurring coverage as distinct value layers.
China's capability is uneven—not absent
- Prior thesis
- China's memory industry was often summarized as fully import-dependent.
- New evidence
- CXMT has built scale in commodity and server DRAM, while YMTC NAND can serve SSD and data-center use cases; leading-edge HBM still trails in process, packaging and qualification.
- Updated view
- Commodity DRAM and NAND are in scale-up mode, while HBM remains the critical gap.
03 · EVIDENCE & ANALYSIS
Evidence and analysis
01｜Why has memory become an AI bottleneck?
GPUs perform the computation, but only when data arrives on time. If memory bandwidth falls short, expensive accelerators wait. Effective AI compute is therefore determined jointly by GPUs, HBM, server DRAM, enterprise SSDs, networking and power.
02｜Why is global DRAM still highly concentrated?
Global DRAM revenue reached about $154.73 billion in 2Q26. Samsung, SK hynix and Micron held 39.4%, 24.9% and 23.3%, respectively—about 87.6% combined. Capital intensity, process technology, yields and customer qualification create formidable barriers.[1]
03｜Why is HBM the scarcest layer?
HBM requires stacked dies, through-silicon vias, advanced packaging, thermal control and joint qualification with accelerator platforms. HBM3e remains the 2026 workhorse as HBM4 enters production and qualification. Micron says HBM4 is shipping in volume for a lead platform and HBM4E is targeted for 2027 production.[7]
04｜Why does conventional server DRAM tighten too?
HBM consumes more wafer input and advanced manufacturing resources, while suppliers prioritize AI servers. TrendForce estimates the 2026 DRAM sufficiency ratio at roughly negative 1% to negative 2% and expects the gap to widen in 2027.[3]
05｜What does NAND have to do with AI?
Models, training data, checkpoints, vector databases and overflow caches cannot all remain in expensive working memory. Enterprise SSDs provide the capacity layer. Revenue across the five largest NAND brands rose 77% quarter over quarter in 2Q26, led by AI servers and high-capacity enterprise SSDs.[2]
06｜Why will 2027 diverge?
DRAM and HBM remain constrained by AI demand, allocation and long expansion cycles. NAND may move toward looser supply in 2H27 as new capacity and technology migrations add output. Enterprise SSD demand can remain structurally firm while consumer NAND returns to price competition.[3]
07｜How far has China progressed?
CXMT's fifth-generation DRAM platform has entered mass production, including two 24-gigabit LPDDR5X products. The company says gross dies per wafer improve by at least 50%, but gross die count is not yield and does not prove cost parity. YMTC's 3D NAND can serve phones, SSDs and data centers; high-layer yields, enterprise qualification and controller ecosystems remain key tests.[4][6]
08｜How should seven companies be grouped?
SK hynix and Micron Technology offer the most direct HBM and server-DRAM exposure. Samsung Electronics combines DRAM, HBM and NAND at scale but is more diversified. Kioxia Holdings and SanDisk are better viewed through enterprise SSD and NAND cycles. CXMT and privately held YMTC should be assessed as localization leaders without allowing industrial progress to substitute for valuation discipline.
09｜Current ACIS research priority
Structural growth: HBM and server DRAM. Structural growth plus cycle: enterprise SSDs. Traditional cycle: consumer NAND. Long-duration localization: CXMT and YMTC. At the company level, SK hynix and Micron lead research priority; Samsung Electronics is a full-stack recovery watch; Kioxia Holdings and SanDisk are tactical; CXMT needs a tighter link among technology, earnings and valuation, while YMTC remains an industry and potential-IPO watch because it is not publicly listed.
10｜What would change the view?
Key reversal conditions include materially looser HBM supply in 2027, large hyperscaler AI-capex cuts, Chinese HBM winning large-scale mainstream accelerator qualification, NAND capacity outrunning enterprise demand, or persistently high memory prices damaging smartphone and PC volumes.
11｜Companies and ticker codes
Samsung Electronics | KRX: 005930; SK hynix | KRX: 000660; Micron Technology | Nasdaq: MU; Kioxia Holdings | TSE: 285A; SanDisk | Nasdaq: SNDK; CXMT | SSE: 688825; YMTC | Private company.
04 · INVESTMENT IMPLICATIONS
Industry and asset implications
SK hynix and Micron
Most direct AI-memory exposure; watch HBM4 qualification, long-term agreements, margins and capex discipline.
Samsung Electronics
Broadest full-stack capability; sustained HBM4 share recovery could add both technical and cyclical upside, but it is not a pure AI-memory asset.
Kioxia Holdings and SanDisk
Closer to enterprise SSD and NAND-cycle exposure; looser 2027 supply is the central risk.
CXMT and YMTC
Localization remains valid, but yields, server and enterprise qualification, cash flow and public-market valuation require separate proof.
Research priority reflects industry position and evidence quality—not a valuation-blind buy list.
05 · VALIDATION & RISKS
What to verify next
HBM4
Samsung, SK hynix and Micron win broader qualification and ship on schedule into mainstream GPUs or custom accelerators.
Failure signal: Qualification, thermal or yield issues depress share and margins.
2027 contract pricing
Long-term agreements continue to lock volume with scarcity premiums.
Failure signal: Supply eases and contract prices decline sequentially.
Enterprise SSD
High-capacity QLC orders, cloud procurement and enterprise revenue keep expanding.
Failure signal: New NAND supply grows faster than AI-storage demand.
China catch-up
CXMT advances server DRAM and HBM qualification while YMTC expands enterprise customers and shipments.
Failure signal: Yield, equipment restrictions or valuation excess weaken commercial conversion.
What would change our view?
Memory remains highly cyclical. High prices invite capacity and can also reduce server, smartphone and PC demand. Customer concentration, platform qualification, yields, export controls and excessive capex can produce a correct industry thesis but poor security returns.
06 · FAQ
Key questions
How are HBM and DRAM related?
HBM is a premium stacked form of DRAM with wider data paths and advanced packaging near the GPU.
Can Chinese memory be used in AI systems?
Yes. Some commodity and server DRAM, NAND and SSD products can already serve AI data centers. Leading-edge HBM remains behind.
Why is NAND not a pure AI growth exposure?
Enterprise SSDs benefit from AI data demand, but NAND supply expands more readily and consumer demand still follows smartphone and PC cycles.
Does the strongest industry trend automatically identify the best stock?
No. Valuation, earnings conversion, capex, cycle position and what is already priced in still determine returns.
07 · TERMS & SOURCES
Terms, sources and related research
Key terms
- High-bandwidth memory (HBM)
- Stacked DRAM placed close to a compute chip to deliver data at very high speed.
- Dynamic random-access memory (DRAM)
- Fast volatile working memory used while servers run tasks.
- NAND flash
- Non-volatile memory that retains data without power and is widely used in SSDs.
- Enterprise SSD
- Data-center storage designed for capacity, endurance, reliability and sustained throughput.
- Through-silicon via (TSV)
- A vertical electrical connection through silicon used to link stacked dies.
- Memory wall
- A bottleneck created when compute capability grows faster than data-transfer performance.
[1] TrendForce｜2026年第二季度动态随机存取存储器市场 ↗
[2] TrendForce｜2026年第二季度闪存市场 ↗
[3] TrendForce｜2027年存储市场分化展望 ↗
[5] Reuters｜长鑫如何成为中国动态随机存取存储器龙头 ↗
Evidence comes primarily from TrendForce, Reuters, company releases and exchange filings. A valid industry thesis, corporate operating benefit and an attractive security price are separate questions. This is industry research, not personalized securities advice.