Author: hoidya | 09

1/ What exactly is the storage industry?

The storage industry is primarily composed of three core product categories: DRAM, NAND, and HBM. Together, they form the data storage system for all digital devices. Whether it's a mobile phone, a computer, or a data center, all must rely on this infrastructure to complete temporary data processing and long-term storage.

Functionally, DRAM is used for temporary data storage during operation, handling the high-speed read and write requirements of computing. NAND is used for long-term data storage, similar to the persistent memory layer of a device. HBM, on the other hand, is a new form that evolved in high-performance computing environments to solve the bandwidth bottleneck problem between GPUs and computing units.

From a system architecture perspective, storage is not a component independent of computing systems, but rather a fundamental dependency layer for all computing systems. Any computing task must first "read data," then perform "computation," and finally "write back the results." Therefore, storage is one of the fundamental constraints in the computing process, not an optional module.

Over the past two decades, demand in this industry has primarily come from three sources: consumer electronics (mobile phones and PCs), enterprise servers, and internet infrastructure. These demands are characterized by high fragmentation, delayed upgrade cycles, and limited scale at individual points of demand. Therefore, the market has long categorized it as a typical cyclical semiconductor industry.

2/ Why is storage long considered a cyclical industry?

The storage industry's long-term strong cyclicality stems from the asymmetry in its supply and demand structure. Demand is typically correlated with consumer electronics cycles and enterprise IT spending cycles, while supply is driven by wafer fab investments, exhibiting a significant time lag.

When demand rises, prices surge, prompting manufacturers to expand production. However, since capacity expansion typically takes 12 to 24 months, the new supply is often released in a concentrated manner after the demand inflection point, leading to a rapid price decline. This mechanism forms a typical boom-bust cycle.

This cyclical structure was particularly evident between 2010 and 2022. For example, the DRAM industry experienced a rapid decline from high profit margins to losses in multiple cycles, followed by a rebound after a new round of demand recovery. This volatility has led the market to view the memory industry as a cyclical asset class characterized by "high volatility and low predictability" in the long term.

At this stage, the industry's pricing mechanism is essentially inventory-driven. Prices rise when inventory decreases and fall when inventory accumulates, with demand itself playing more of a triggering variable than a structural one.

3/ What was the demand structure like before AI?

Before the advent of artificial intelligence, storage demand was primarily driven by consumer electronics and traditional internet infrastructure. Consumer electronics are characterized by long upgrade cycles and relatively predictable demand; for example, the replacement cycle for smartphones is typically two to three years. Servers and enterprise storage, on the other hand, rely more on the pace of IT capital expenditure and are also highly cyclical.

In this structure, storage, as a standardized product, is primarily priced based on supply and demand, rather than the long-term locked-in demand of a single large customer. Therefore, the market exhibits a high degree of spot market characteristics, and price signals can quickly reflect inventory changes and capacity adjustments.

In other words, prior to AI, the demand structure of the storage industry was fragmented and lacked long-term rigid constraints. This is the core basis for its cyclical characteristics.

4/ Why has AI fundamentally changed the structure of storage demand? (From cyclical goods to infrastructure)

In the past, storage demand was driven by consumer electronics (smartphones, PCs), which was essentially "delayed consumption." But AI brings a completely different demand function: it is a continuous computing system, and memory usage grows linearly or even superlinearly with model size.

Taking AI data centers as an example, the GPU is not the bottleneck in computation during training and inference, but rather the bottleneck in memory bandwidth. This directly drives HBM to become a rigid demand. Industry data shows that the demand for high-bandwidth memory in AI servers is growing at a rate far exceeding that of traditional DRAM, leading to a long-term lock-in of HBM production capacity, with some even seeing all pre-sales completed before 2026.

More importantly, there are changes on the supply side: because HBM's profit margin is significantly higher than that of traditional DRAM, manufacturers are proactively redistributing capacity, shifting wafer production from DDR4/DDR5 to HBM. This structural squeeze-out effect has led to a "non-demand-driven shortage" in both traditional DRAM and NAND.

Extreme signals have emerged in the market: spot prices for some DRAM and NAND have risen by 15-20% within the quarter, and "intraday price adjustments" have occurred.

5/ How was storage priced in the past?

Between 2010 and 2022, the pricing mechanism in the memory industry was highly typical, representing a standard semiconductor cycle model:

Prices are driven by inventory cycles, not by demand structure.

When inventory decreases → prices rise → manufacturers expand production → supply exceeds supply → prices collapse.

The core constraint of this mechanism is "the lag in capacity building (1-2 years) + the deferability of demand".

For example, in the previous cycle, the DRAM industry often experienced significant quarterly profit fluctuations, even turning from high gross profit to losses, and then quickly reversing.

However, this mechanism was disrupted in the AI ​​era because two variables changed simultaneously:

• First, demand has shifted from dispersed consumption to centralized procurement.
• Second, supply has shifted from "free market expansion" to "profit-first allocation (HBM priority)".

The result is that cyclical fluctuations still exist, but price elasticity is structurally compressed.

6/ What structural changes have occurred now?

The core change in the current (2024–2026) storage market is not price increases, but rather the shift in market structure from a "spot market" to a "contract allocation system".

First, there's the crowding-out effect of HBM. Because the profit per HBM wafer is significantly higher than that of DDR4/DDR5, Samsung, SK hynix, and Micron have all prioritized shifting their production capacity to HBM. Industry data shows that HBM is rapidly rising from a low single-digit percentage to a structural level of over 40% of DRAM revenue.

This structural adjustment led to two results:

• First, the supply of traditional DRAM is shrinking.
• Second, NAND has entered a state of passive shortage.

At the same time, the market has entered an extreme supply and demand state: DRAM industry revenue will grow by 17.1% year-on-year in Q2 2025, but the growth will not be driven by a surge in demand, but by a combination of rising prices and supply constraints.

Even more extreme signals come from the delivery side: industry lead time has extended from the normal 8–12 weeks to 39–52 weeks, and for some automotive-grade memory, it has even exceeded 70 weeks.

This signifies a key structural change: memory is no longer a "commodity that can be traded immediately," but rather a "rationed resource."

This will create a positive feedback loop:

Price increases → Manufacturers reduce spot supply → Buyers lock in orders in advance → Further reduce spot liquidity → Prices continue to rise.

7/ Who benefits from this structure?

The profit structure of the storage industry is undergoing a significant shift.

First layer: Supply side (Samsung / SK hynix / Micron)

These companies are transforming from "cyclical manufacturers" into "AI infrastructure providers." Among them, SK hynix's leading position in HBM has gradually made it a holder of structural pricing power, and its DRAM market share has increased to approximately 38%.

Second layer: Demand side (Microsoft / AWS / Google)

These companies lock in future supply through long-term contracts, which is essentially "time arbitrage": using current capital expenditures to lock in future AI computing power and memory costs.

The third tier: AI model companies (such as OpenAI)

They are caught between cash flow pressure and computing power demand, forming a closed loop through financing → capex → re-locking supply.

The key change is that pricing power is shifting from the "market" to the "contractual structure."

8/ Risks and Falsification Conditions

This round of "AI memory supercycle" has at least three clear falsification conditions:

First, if AI capex enters a contraction cycle (hyperscalers reduce investment intensity), the current demand structure will quickly become distorted, because memory demand is highly dependent on AI compute expansion.

Second, if the HBM technology path is replaced (e.g., by a new memory architecture or compute-memory fusion), the current HBM premium will be compressed, causing production capacity to flow back to DRAM/NAND.

Third, if the capacity expansion cycle accelerates again (Samsung/SK hynix re-enters aggressive capacity expansion), the current supply constraint will reverse into an oversupply cycle in 1–2 years.

In other words, this structure is predicated on the following:

The growth rate of AI demand > the rate of capacity expansion + the rate of technology substitution