Why Storage Is Always in Short Supply: The Narrative and Investment Logic of AI Storage

1. Why AI Discussions Always Circle Back to Storage

Over the past two years, market discussions on AI infrastructure have first focused on GPUs, then on power, followed by optical communication, liquid cooling, transformers, and data center REITs.

But as we continue to dissect the system internally, we inevitably return to the same fundamental question:

Models can grow larger, inference can increase, Agents can become more complex, but all these capabilities must ultimately reside within a real physical system: data must be stored, moved, read, and written back.

This is why "storage" has recently become a core narrative in AI investment.

The storage discussed here is not limited to hard drives in the narrow sense, nor just memory modules in consumer electronics. It refers to the entire AI storage hierarchy:

Tier	Corresponding Products	Role in AI
Hottest Tier	HBM	Placed next to GPUs, handles model execution, hot KV cache, high-bandwidth data exchange
Hot Tier	DDR5 / LPDDR / SOCAMM / CXL Memory	Handles CPU-side scheduling, long-context extension, multi-Agent orchestration
Warm Tier	Enterprise SSD / High IOPS NAND	Handles persistent KV cache, vector databases, RAG retrieval, training data reads
Cold Tier	HDD / Object Storage	Stores training corpora, logs, archives, multimodal data lakes

The essence of AI is not performing a single matrix multiplication. When truly deployed in applications, it becomes a continuous token factory: every user request, every round of tool calls, every retrieval, every intermediate state of an Agent generates new read/write demands.

Therefore, compute is the engine, power is the fuel, optical communication is the road, while storage simultaneously serves as the fuel tank, fuel line, warehouse, and workbench. Without it, even the most expensive GPUs can only idle.

---

2. The "Multiplier Formula": Why Storage Demand Gets Amplified Many Times Over

We can understand AI storage demand with a simplified formula:

AI Storage Demand ≈
Model Weights
+ KV Cache
+ Concurrent Users
+ Context Length
+ Agent Steps
+ RAG / Vector DB Retrieval Data
+ Training & Log Retention
- Algorithm Compression Efficiency

If expressed as a multiplicative model, it roughly looks like:

Total Storage Pressure = Base Model Scale × Context Multiplier × Concurrency Multiplier × Agent Steps × Retrieval Multiplier ÷ Compression Efficiency

This is the "multiplier theory" you might recall: a single variable might seem like just a few times growth, but when multiple variables multiply, the result becomes staggering.

For example:

• Models evolving from 70B to hundreds of B / MoE / multi-model collaboration not only increase weights but also complicate scheduling and caching.
• Context length expanding from 8K, 32K to 128K, 1M tokens leads to linear growth in KV cache.
• Users shifting from a few developers to enterprise employees, customer service, code Agents, and search portals increases inference concurrency.
• A single request evolving from "one question, one answer" to multi-step Agent execution, potentially calling search, code, databases, browsers, and toolchains.
• RAG is not just reading a piece of text but accessing vector databases, document libraries, historical logs, images, videos, and code repositories.

KV cache can be calculated more intuitively:

KV Cache Bytes
= 2 × Layers × KV Heads × Head_Dim × Bytes_Per_Element × Context_Tokens × Batch/Concurrency

A rough calculation for a 70B-scale model using GQA, assuming 80 layers, 8 KV heads, head_dim 128, BF16:

KV Cache per Token
= 2 × 80 × 8 × 128 × 2 bytes
≈ 0.31 MB

One 128K-context Session
≈ 40 GB KV Cache

1000 Concurrent Long-Context Sessions
≈ 40 TB Hot KV Cache

And this is just the hot cache, not including model weights, activations, retrieval libraries, logs, training data, and redundant backups.

Therefore, AI's demand for storage is not a linear "more users, more hard drives" relationship, but simultaneous amplification across multiple dimensions. Micron's materials at COMPUTEX 2026 also offered a directional assessment: AI context length is rising rapidly, and server memory capacity is also increasing significantly. In other words, the system bottleneck is expanding from "can it compute?" to "can it remember, read fast enough, and move fast enough."

---

3. Why Compression Technology Isn't Necessarily Bearish for Storage

The market often panics due to innovations in certain technical paths, such as KV cache compression, sparse attention, MLA, and quantization techniques: if efficiency improves, won't we need as much storage?

In the short term, compression certainly reduces the storage cost per token.

But in the long term, this is closer to Jevons paradox:

When the efficiency of using a resource improves, the end result isn't necessarily a decrease in total demand; instead, it may open up new use cases, leading to continued increase in total consumption.

After cars became more fuel-efficient, people didn't necessarily drive less; they might drive farther. AI is similar:

• If KV cache compression improves 4x, product managers might immediately increase context length by 4x.
• If inference costs drop, enterprises will hand more workflows over to Agents.
• If model response becomes faster, users will increase call frequency.
• Previously "uneconomical" analyses of long documents, long videos, code repositories, and enterprise knowledge bases become viable scenarios.

Thus, efficiency gains truly change "unit cost" and "application boundaries," not simply eliminate demand. Storage demand will only be truly suppressed if the rate of algorithmic efficiency improvement consistently outpaces the multiplicative growth of usage volume, context length, concurrency, and Agent steps.

This is also the most critical judgment in storage investment:

If: Usage Growth × Context Growth × Agent Steps Growth > Algorithmic Compression Efficiency Improvement
Then: Storage Demand Continues to Expand

If: Algorithmic Efficiency Suddenly Leaps Ahead Significantly, While AI Application Monetization Slows
Then: Storage Returns to Cyclical Stock Logic

---

4. How This Storage Cycle Differs from Past PC / Mobile Phone Cycles

Traditional storage is a classic cyclical stock.

Past demand primarily came from PC, mobile phone, and server replacement cycles. The industry logic was simple:

1. Prices rise.
2. Manufacturers expand production.
3. Supply exceeds demand.
4. Prices fall.
5. Manufacturers cut production.
6. The next cycle begins again.

This logic hasn't disappeared, but AI has changed three variables.

1. Demand Shifts from Consumer Electronics to Infrastructure Investment

PC and mobile phone demand is influenced by replacement cycles, with limited elasticity. But AI data centers are more like energy, power grid, and cloud infrastructure construction; cloud providers lock in capacity years in advance.

SK hynix's Q1 2026 financial report shows that AI infrastructure demand is driving sales of high-value-added products like HBM, high-capacity server DRAM, and eSSD. The company's quarterly revenue reached 52.6 trillion KRW, operating profit 37.6 trillion KRW, with an operating margin of 72%. This is not the profit margin seen in traditional consumer electronics upcycles.

2. HBM is Not Ordinary DRAM

The difficulty with HBM isn't just stacking more DRAM dies; it involves TSV, advanced packaging, yield, customer qualification, and GPU platform co-design.

Ordinary DRAM can be closer to a commodity, but HBM is more like a key component bound to the AI accelerator platform. Customers can't just buy a batch from the spot market and use it; they need to pre-qualify, lock in capacity, and sign long-term agreements.

More importantly, HBM consumes advanced DRAM wafer and packaging capacity. When high-end HBM consumes capacity, ordinary DDR5 and server DRAM also tighten. This is the "spillover from high-end shortage to low-end" transmission.

3. NAND is Also Revitalized by AI

The biggest issue with NAND in the past was its tendency towards oversupply, rapid technological capacity expansion, and high price elasticity. But AI inference and Agent systems have brought new eSSD demand:

• Vector databases.
• RAG retrieval.
• Persistent KV cache.
• Multimodal data lakes.
• High-frequency random access.
• AI PC / On-vehicle local caching.

TrendForce's January 2026 report mentioned that AI Agents, RAG, and high random access patterns are increasing enterprise SSD demand and projected significant revenue growth for both DRAM and NAND markets in 2026, with contract price increases potentially continuing into 2027.

Therefore, this cycle isn't about a single product price hike; it's about HBM, server DRAM, SOCAMM, eSSD, and HDD data lakes all entering the AI storage hierarchy together.

---

5. The Current Storage Narrative: From "Selling Chips" to "Selling AI Memory Capability"

The core change in the storage industry's valuation isn't because prices have risen, but because the market is starting to believe:

Storage is no longer just a cyclical product; it's a strategic asset within AI infrastructure.

Micron's Q2 FY2026 revenue reached $23.86 billion, far exceeding the $8.05 billion from the same period last year, with a GAAP gross margin of 74.4% and an operating margin of 67.6%. In its announcement, the company directly attributed this performance to strong demand, tight supply, and the enhanced strategic value of storage in the AI era.

At GTC 2026, Micron also announced that its HBM4 36GB 12H has entered volume production and shipment, targeting the NVIDIA Vera Rubin platform. Simultaneously, it launched products like PCIe Gen6 data center SSDs and SOCAMM2. This indicates that storage vendors are no longer just selling individual chips but are building complete memory hierarchies around AI data centers.

The complete narrative chain for this cycle is:

AI application growth
→ Token production growth
→ Growth in context and Agent steps
→ Increased demand for KV cache / RAG / data lakes
→ Simultaneous tightness in HBM, DRAM, eSSD, HDD
→ Storage vendors sign long-term agreements, raise prices, improve product mix
→ Non-linear improvement in margins and cash flow
→ Market begins discussing revaluation from cyclical to growth stocks

This is also why, after this round of storage stock gains, the market debate isn't about "how much money they'll make this year," but rather "how many years these profits can be sustained."

---

6. Cyclical or Growth Stock: The Correct Answer Might Be "Cyclical Growth Stock"

Simply classifying storage as a cyclical stock might cause one to miss the structural revaluation brought by this wave of AI infrastructure.

But directly treating storage as a software-like growth stock is also risky.

A more reasonable definition is:

Storage is a cyclical growth stock whose boom cycle is extended by AI, with an elevated profit center, but which still cannot escape the capital expenditure cycle.

Why Bulls Are Willing to Assign Higher Valuations

Bulls are betting on three things:

First, the HBM supply bottleneck cannot be resolved in a year.
Wafers, advanced packaging, yield rates, customer qualification, and platform adaptation all have physical cycles.

Second, as AI inference moves from training to Agents, storage demand doesn't decrease; it diffuses.
Training emphasizes HBM, inference emphasizes KV cache, Agents emphasize long context and retrieval, and enterprise applications emphasize eSSD and data lakes.

Third, long-term agreements and prepayments have changed cyclical volatility.
If major customers lock in capacity in advance, or even secure supply through strategic agreements, prepayments, and minimum price clauses, the revenue visibility for storage vendors will be higher than in traditional cycles.

Why Bears Still Cannot Be Ignored

The bears also have valid points.

First, in the history of the storage industry, someone has always said "this time is different."
But whenever supply increases, demand slows, and inventory piles up, prices still fall.

Second, algorithmic efficiency is a real variable.
MLA, sparsification, KV cache compression, MoE, and on-device small models could collectively reduce per-unit storage demand, altering the demand slope.

Third, cloud vendor CAPEX is not unlimited.
If AI application monetization fails to keep pace with capital expenditures, cloud vendors may slow server procurement in 2027-2028.

Fourth, competition will rebalance.
SK hynix, Micron, and Samsung are all expanding HBM. Samsung catching up on yield, Micron ramping HBM4 volume, and Chinese vendors advancing in mid-to-low-end DRAM/NAND will all change price expectations.

Therefore, the difficulty in storage investing is not judging "whether it's good or bad," but judging:

How many years can this round of excess profits be sustained?
For how many years is the market currently pricing?
Which arrives first: supply release or demand slowdown?

---

7. Investment Map: Different Companies Capture Different Profit Pools

The storage chain cannot be viewed as a monolith. Different companies capture different profit pools.

1. Global Memory Makers: SK hynix / Micron / Samsung

These are the core suppliers of HBM and high-end DRAM.

• SK hynix: Strongest first-mover advantage in HBM; its margins have already proven the extreme elasticity of AI storage.
• Micron: The most direct expression of a memory maker in the US stock market; HBM4, data center SSDs, and SOCAMM2 all reinforce the AI storage platform narrative.
• Samsung: Once HBM yield and customer qualification catch up, it could lead to supply rebalancing and potentially capture a larger market share.

From an investment perspective, this group represents the "memory makers' pricing power" thesis.

2. NAND / SSD / HDD: AI Data Lakes and Warm/Cold Tiers

If HBM is the hot tier, eSSD and HDD are the warm and cold tiers.

AI Agents, RAG, multimodal training, and log retention will continue to consume enterprise SSDs and high-capacity hard drives. The logic here is more cyclical than HBM's, but if the AI storage hierarchy truly solidifies, NAND and HDD will re-enter the AI infrastructure narrative from being "forgotten cyclical products."

3. A-Share Chain: Technology Royalties, Module Elasticity, Distribution Inventory

A-share storage stocks can be divided along a spectrum:

Type	Representative	Characteristics
Technology Royalty Type	Montage Technology	DDR5 interfaces, server memory ecosystem, relatively low cyclical exposure
Module & Brand Type	Longsys, BIWIN Storage	Benefit from price increases and inventory, but also bear inventory reversal risk
Distribution Elasticity Type	Shannon Core Technology	High elasticity during upturns, greatest risk during downturns
Domestic Substitution Foundation	CXMT, YMTC, etc.	Still gaps in high-end HBM; may capture structural spillover in mid-to-low end

For this group, one cannot just look at "storage price increases." One must examine whether they capture technology royalties, inventory gains, brand/channel advantages, or domestic substitution.

---

8. Future Space: Not Just HBM, But a Complete AI Storage System

The future potential of storage cannot be calculated using only the HBM TAM (Total Addressable Market).

A more complete AI storage TAM should include:

AI Storage TAM
= HBM
+ Server DRAM
+ SOCAMM / CXL Memory
+ Enterprise SSD
+ High-Capacity HDD / Object Storage
+ Storage Controllers / Interface Chips
+ Data Center Storage System Software & Cache Layers

HBM is the brightest layer because it's the scarcest, has the highest gross margins, and is most tightly bound to GPUs. But in the long run, AI inference and Agent proliferation will diffuse demand across the entire storage pyramid.

The future can be divided into three stages:

2025-2026: HBM & DDR5 Supply Tightness

Mainly trading on HBM, server DRAM, DDR5, and eSSD price increases. Memory makers' margins recover rapidly, and the market begins shifting valuation from cyclical to growth stock metrics.

2027: Long-Term Agreement Fulfillment & Tightest Supply-Demand Window

This year will depend on two variables:

• Whether long-term agreement prices with major customers can be maintained.
• Whether the new HBM capacity and yield from Samsung / Micron / SK hynix can be released quickly.

If demand remains strong, 2027 could be one of the most profitable years. If supply catches up or CAPEX slows, stock prices may anticipate a reversal early.

Post-2028: The Real Watershed

By 2028, the market will ask again:

• Is HBM still in short supply?
• Have AI Agents generated substantial real revenue?
• Have eSSD / ICMS / long context become new growth curves?
• Has supply from China and Samsung depressed prices?
• Has capital expenditure by storage vendors led to a new round of oversupply?

If the answers lean bullish, storage will continue to be priced as an AI infrastructure growth stock.
If the answers lean bearish, it will revert to being a traditional cyclical stock.

---

9. The 10 Most Important Metrics to Track

For storage investing, one cannot just look at stock prices and spot prices. What truly matters are these metrics:

1. Cloud Vendor CAPEX Guidance: Whether Microsoft, Google, Amazon, Meta, Oracle continue to raise AI infrastructure budgets.
2. AI Revenue Realization: Whether cloud AI revenue, model API revenue, enterprise Agent payments keep pace with capital expenditures.
3. HBM Specification Upgrades: Capacity, bandwidth, power consumption, and customer qualification for HBM3E, HBM4, HBM4E.
4. HBM Long-Term Agreements: Existence of prepayments, minimum prices, volume lock-in clauses.
5. Samsung Yield Catch-up: Once caught up, the supply landscape will change.
6. Micron HBM4 Volume Ramp: Determines whether MU can maintain its growth stock narrative.
7. SK hynix Gross Margin and Cash Flow: Judges the sustainability of current excess profits.
8. DDR5 / eSSD Contract Prices: Reflect real enterprise customer procurement better than scattered spot prices.
9. Inventory Changes: Once module maker and distributor inventory gets out of control, a cycle reversal can happen quickly.
10. Algorithmic Efficiency Breakthroughs: If KV cache compression, MLA, sparsification enter production at scale, they will change the demand slope.

---

10. Conclusion: Why "Always Talk About Storage"

Because every upgrade in AI capability pushes more pressure onto the storage system.

Larger models need more HBM.
Longer context needs more KV cache.
More Agent steps need more intermediate state.
More widespread RAG needs more enterprise SSDs.
More multimodal data needs more data lakes.
More users need higher concurrency and larger memory pools.
Improved efficiency, in turn, opens up more use cases.

This is the core narrative of storage:

It is not a peripheral component in the AI industry chain, but one of the most fundamental, most easily underestimated, and most prone to cyclical explosive bottlenecks after AI transitions from demos to real productivity.

But in investing, you cannot rely on narratives alone.

The best opportunities in storage often appear when the market still treats it as a cyclical stock, but its fundamentals have already begun to show characteristics of a growth stock. The most dangerous phase is when everyone starts treating it as a growth stock, while supply, inventory, algorithm efficiency, and CAPEX have quietly begun to shift.

Therefore, the final judgment of this article is:

Storage is not a permanent growth stock, but AI has endowed it with a longer, higher, and more complex cycle than in the past. The real money to be made is not just from the phrase "storage price increases," but from understanding how much longer this cycle has been extended by AI and how many years of pricing the market is currently factoring in.

Reference Sources

1. SK hynix News Center — Official source for the Q1 2026 financial data (HBM, high-capacity server DRAM, eSSD performance) cited in the article.
2. Micron Newsroom — Official release channel for the FY2026 Q2 results, HBM4 mass production shipments, and product updates like SOCAMM2 mentioned in the article.
3. TrendForce News Center — Publisher of the January 2026 enterprise SSD / DRAM / NAND market research reports cited in the article.