Markets / Industries / Memory & Storage
Industry view · Memory & Storage
For decades memory was the commodity that got cheaper every year. The AI buildout inverted that: HBM now makes up 30-40% of an AI accelerator's manufacturing cost, DRAM contract prices rose roughly 90% quarter-on-quarter into early 2026, and a three-firm oligopoly is choosing margin over volume. The constraint on AI is no longer how many GPUs you can buy — it is how much memory you can put next to them.
01 · The thesis
The memory industry spent thirty years in brutal boom-bust cycles where any firm that under-built lost share. The AI era broke that reflex. Samsung, SK hynix and Micron — together roughly 70% of the DRAM market — have signalled they will not chase volume, diverting cleanroom capacity from commodity DRAM and NAND into HBM, where a single stack sells for $300-500 and demand is effectively pre-sold to Nvidia and the hyperscalers. The result is a structural shortage that TrendForce and others expect to persist past 2027.
The winners are not those with the best AI models but those who own the physics. HBM stacks 8-16 DRAM dies and sells at multiples of commodity bit pricing, so the same wafer earns far more as HBM than as a PC module. That single substitution is why SK hynix overtook Samsung in 2025 annual operating profit for the first time, why Micron's data-center revenue more than doubled, and why memory — the least glamorous layer — captured a disproportionate share of AI infrastructure economics.
The same fabs make commodity DRAM and HBM dies; reallocating wafers to HBM starves the rest of the market and is the root of the 2025-2026 shortage.
Advanced packaging and through-silicon-via stacking turn ordinary DRAM dies into HBM, the highest-value memory product and the hardest to yield.
Nvidia and AMD now gate their roadmaps on HBM qualification; HBM4 at 10+ Gb/s is a pass/fail for a GPU generation, not an afterthought.
CXL pooling and KV-cache servers add capacity beyond the GPU package to relieve inference bottlenecks that more GPUs cannot fix.
Enterprise SSDs and QLC NAND absorb training datasets and checkpoints; NAND pricing hit multi-year highs as AI data-center demand surged.
02 · Public players & exposure
We plot the listed players on two editorial axes — how exposed each is to AI disruption, against how ready its data, brand and position are to be the answer. The figures in the table are sourced; the placement is our read.
| Company | Stance | The sourced fact |
|---|---|---|
| SK hynix000660.KS | HBM kingpin | Held ~62% of HBM shipments in Q2 2025 and posted record 2025 results with revenue of 97.1 trillion won, overtaking Samsung in annual operating profit for the first time. |
| Samsung Electronics005930.KS | Catching up | Cleared Nvidia and AMD HBM4 qualification (11.7 Gb/s, above the 10 Gb/s bar) and regained the No. 1 DRAM share at 38.6% in Q4 2025. |
| MicronMU | US HBM share gainer | FY2025 revenue reached $37.4B vs $25.1B prior, with data-center revenue more than doubling and HBM reaching a multi-billion-dollar annual run rate. |
| NvidiaNVDA | Demand sink | HBM is now 30-40% of AI accelerator build cost; on a B200 the eight HBM3e stacks cost roughly $2,400, exceeding the logic die. |
| Kioxia285A.T | NAND revival | Market value topped 10 trillion yen in January 2026 on soaring NAND prices; plans US ADRs to ride AI storage demand. |
| SanDiskSNDK | Pure-play NAND | Spun out of Western Digital in February 2025; a smaller pure-play NAND maker ramping BiCS8 into AI data-center storage. |
| CXMTCXMT | China challenger | Revenue rose ~130% to about $8B in 2025 as capacity expanded from 100k to 290k wafers/month; global DRAM revenue share climbed from ~4% to ~8% in a year. |
| Astera LabsALAB | CXL connectivity | Powers Microsoft Azure's first CXL memory expansion; IPO'd in March 2024 near a $5.5B valuation on surging AI connectivity demand. |
| UnifabriXunifabrix | Memory-fabric startup | Building CXL Smart Memory Fabric pooling and pursuing early deployments across HPC and cloud environments. |
| Western DigitalWDC | Post-spin HDD | After spinning off SanDisk's NAND business in February 2025, WDC is concentrated on HDD, exposed to but not central to the HBM/DRAM supercycle. |
03 · The two clocks
Three structural clocks the industry is racing against
The capacity clock. HBM now consumes a meaningful and rising share of DRAM wafers, and analysts estimate AI data centers could absorb a large majority of high-end DRAM in 2026. Because suppliers chose not to expand aggressively, TrendForce and peers expect the shortage and elevated pricing to extend past 2027 — there is no fast switch back on.
The memory-wall clock. LLM inference with heavy KV-cache use can require tens to over a hundred GB per GPU, and the binding constraint is bandwidth and capacity, not compute. DRAM contract prices surged ~90% QoQ into Q1 2026 as all three makers reallocated to HBM, accelerating CXL pooling and KV-cache offload as the only way to scale inference economically.
The China clock. CXMT's DRAM capacity is climbing toward a low-double-digit share of global supply, with Alibaba, Tencent and Baidu as anchor customers. It first floods the domestic market — but as it climbs the bit-share curve toward HBM, it threatens the oligopoly's pricing power on commodity DRAM within a few years.
04 · Private flagships
The companies attacking this industry AI-first, with disclosed funding where available:
First mover on TSV-stacked HBM and Nvidia's primary supplier; completed HBM4 development in September 2025 and began Q4 shipments, holding the lead it took during the supercycle.
The only US-based maker in the DRAM/HBM triad; HBM3E 12-high ramped through FY2025 with HBM reaching a multi-billion-dollar run rate and data-center revenue doubling.
Supplies the retimers and CXL controllers that let hyperscalers pool memory beyond the GPU package; powers Microsoft Azure's first CXL memory expansion.
Japan's NAND champion, in a JV with SanDisk; reopened Fab2 at Kitakami to meet AI-driven NAND demand and plans US ADRs.
State-backed national champion scaling DRAM capacity faster than any incumbent, anchored by Alibaba, Tencent and Baidu, and beginning to climb toward higher-value products.
Builds Smart Memory Fabric pooling systems aimed squarely at the inference memory wall, targeting HPC and cloud environments.
05 · Signals
For the first time since 1992 Samsung lost the DRAM revenue lead, driven by HBM; Samsung reclaimed No. 1 by Q4 2025, underscoring how HBM now decides the ranking.
A pure-play NAND insurgent is created just as AI data-center demand pushes flash to multi-year-high pricing.
SK hynix completes HBM4 development in September; Samsung clears Nvidia/AMD tests at 11.7 Gb/s, narrowing the gap into 1Q26 contracts.
TrendForce raises its forecast to ~90-95% QoQ DRAM gains as makers divert capacity to AI server memory and CSPs lock 2-3 year supply contracts.
CXMT heads toward a larger share of global DRAM capacity, first absorbing domestic demand but setting up a medium-term challenge to oligopoly pricing power.
06 · The exposure read
AI rewards clean, structured advantage and punishes friction. The line runs through who owns the data, the brand and the customer — and who is merely a step the technology can route around.
Sources
TrendForce — Memory Wall / AI memory supercycle · TrendForce — 1Q26 DRAM up 81% QoQ on contract surge · TrendForce — 4Q25 DRAM revenue, Samsung regains No.1 · SK hynix — Record FY2025 financial results · CNBC — SK hynix doubles profit in 2025 on AI memory · Micron — FY2025 Q4 press release (8-K) · Futurum — Micron Q3 FY2025 HBM growth · TrendForce — Samsung HBM4 shipments to Nvidia/AMD · Counterpoint — Global DRAM and HBM market share · Silicon Analysts — NVIDIA B200 cost breakdown · Introl — The AI Memory Supercycle (HBM cost share) · TrendForce — Kingston warns NAND prices up 246% since early 2025 · SemiAnalysis — China's CXMT set to challenge DRAM · Seoul Economic Daily — Kioxia market cap tops 10 trillion yen