AI inference is in strong demand, NAND is "more stable with a longer cycle"! JPMorgan: eSSD is the main player this round, and Kioxia is the top pick.

AI inference is in strong demand, NAND is "more stable with a longer cycle"! JPMorgan: eSSD is the main player this round, and Kioxia is the top pick.

Stop fixating on GPUs and HBM high-bandwidth memory. The AI inference wave is freeing NAND flash from its fate as a "highly cyclical commodity" and turning it into a high-growth AI infrastructure asset.

According to Windchaser Trading Desk, on January 23rd, JPMorgan's Asia-Pacific tech research team released an in-depth report "Semiconductors: NAND—A Longer, Stronger Upcycle," announcing that the NAND sector has entered a brand-new supercycle driven by AI inference. Unlike previous cycles driven by smartphone and PC shipments, the core driving force of this cycle is enterprise SSDs (eSSD).

JPMorgan believes that as AI workloads shift from training to inference, and with supply bottlenecks in HDD (hard disk drives) for nearline storage, the NAND market is experiencing unprecedented structural growth. Investors have severely underestimated NAND’s strategic importance in the age of AI inference.

Farewell to the “cyclical curse”: 10% growth is history, brace for a 30% TAM surge

For a long time, NAND has been seen as a classic highly cyclical commodity: technological progress brings cost reductions, manufacturers rush to expand production, and this leads to price collapses.

But now, this logic has been broken.

JPMorgan’s review of the past 25 years found that over the last 20, 15, or 5 years, NAND's TAM (total available market) compound annual growth rate (CAGR) has always hovered between 7%-12%.

However, forecasts show that over the next three years (2025-2027), this figure will leap to 34%.

This leap in growth isn’t just from blind shipment stacking, but a rare “volume-plus-price” resonance. The report notes that the previous expansion of the NAND market mainly relied on high growth in Bit Shipment to offset plunging ASPs (average sales prices). In this cycle, ASP will become a positive force for growth.

JPMorgan predicts that in 2026, NAND’s blended ASP will rise sharply by 40% year-on-year. More crucially, this increase won’t be fleeting, as prices are expected to only slightly decline by 2% in 2027, staying high.

This change in fundamentals directly leads to a revaluation. The report notes that the market is pricing storage stocks by shedding their purely “cyclical” nature and assigning an “AI growth” attribute instead.

Since January 2025, the global storage sector’s market cap has soared 242%, but JPMorgan believes that considering NAND’s irreplaceability in AI inference, this revaluation is far from over.

The “invisible cornerstone” of AI inference—Why eSSD?

There’s a common misconception in the market: AI only benefits DRAM (especially HBM), and NAND is just a supporting act. JPMorgan spends much of the report correcting this view: in the AI inference stage, the importance of eSSD (enterprise solid-state drives) is no less than that of HBM.

The report clarifies the essential difference between training and inference:

  • Training phase: The focus is on compute power and bandwidth. Massive data needs parallel throughput; HBM is the undisputed lead, while NAND mainly stores cold data, its value limited.
  • Inference phase: The focus shifts to latency and context. Real-time user queries to large models require rapid retrieval and token generation from enormous parameters.

Secondly, key technology breakthrough—KV Cache Offloading.

As the context window for AI models keeps getting longer, GPU VRAM (HBM) capacity quickly gets squeezed. To resolve this bottleneck, KV Cache Offloading technology was introduced: intermediate state data that had to be crammed into expensive HBM is now offloaded to external storage.

This creates very high requirements for external storage: it must be fast and must be large.

JPMorgan specifically mentions Nvidia's ICMS (Inference Context Memory Storage) platform based on BlueField-4 DPU, unveiled at CES 2026. This platform allows GPUs to bypass CPUs and access SSDs with ultra-low latency, so eSSD is no longer just a "hard drive" but becomes "secondary memory" in the AI compute system.

Demand feedback is direct and explosive: the report says eSSD bit shipments grew a staggering 86% year-on-year in 2024, a level not seen since 2012.

Looking ahead, JPMorgan forecasts that the storage capacity required per AI server will reach over 70TB, more than 3x that of general servers (~20TB). By 2027, eSSD will account for 48% of global NAND bit demand, overtaking smartphones (30%) and PCs (22%) to become the largest demand pillar in the NAND industry.

The perfect storm—HDD shortage and QLC’s rise

The eSSD boom isn’t just because of incremental AI demand, but also forced replacement in the existing market. The report reveals a market-ignored supply chain crisis: a collapse in hard disk drive (HDD) supply.

  • Two desperate years for HDDs

Due to several years of storage industry downturn, HDD makers (like Seagate, Western Digital) slashed capital spending. When AI-driven data storage demand suddenly soared, the HDD industry found itself unable to expand production quickly.

JPMorgan cites TrendForce and Gartner data: today's large-capacity nearline HDDs have delivery cycles up to two years, with severe supply shortages.

  • QLC’s cost-performance inflection point

Facing HDD shortages, data center clients have no choice but to turn to NAND flash. This gives QLC (quad-level cell flash) a perfect opportunity to rise.

Although SSD prices (even QLC) remain several times higher than HDD (currently about 6-8x), in the race to build AI infrastructure, "availability" and "performance" take priority over "cheapness." Furthermore, SSD’s advantages in energy efficiency and space usage align better with stringent power and cooling requirements in AI data centers.

The report specifically notes that SSD’s penetration in "business-critical" storage is still just 19%, so future space for HDD replacement is enormous. JPMorgan calculates that every 1 percentage point increase in SSD penetration brings about $2 billion in additional revenue for the NAND sector.

"Tacit agreement" on supply: Why aren’t manufacturers expanding wildly?

Demand is booming, prices are alluring—historically, manufacturers should be buying equipment like crazy and building new factories. But JPMorgan observes just the opposite: supply-side restraint is unprecedented.

  • Capital expenditure intensity at record lows

Data shows: over the next three years, NAND capital expenditure (Capex) as a proportion of sales will drop to 15%-16%. By comparison, it was 30%-50% most years over the past decade and peaked at 68% in 2018.

JPMorgan analysts view this “low investment” as structural rather than temporary.


Caption: NAND sector capital expenditure intensity (Capex/Sales) is expected to stay low for a long time—this is the biggest difference in this cycle, constraining future capacity release.

  • Physical constraints: limits of stacking

Why not grab the profits? Because they’re out of reach. The report explores technical bottlenecks in depth:

  • Etch difficulty rises exponentially: When NAND stack heights breach 300 or 400 layers, drilling uniform holes (Channel Holes) into a 30μm stack becomes extremely difficult. High aspect ratio etching causes hole deformation, leading to cell failure.
  • Wafer warpage: Hundreds of stacked layers produce huge mechanical stresses, leading to wafer bends of several hundred microns—a disaster in manufacturing.
  • Hybrid bonding’s high threshold: To address these issues, manufacturers must adopt hybrid bonding, separately manufacturing logic chips and storage arrays before bonding. This requires ultra-high-precision equipment and logic chip yield drag down total output.

Thus, NAND expansion now is not “buy equipment” but “painful tech migration.” JPMorgan predicts that in 2026, global NAND wafer output will only grow by 3% year-over-year: supply growth is almost entirely reliant on tech upgrades (more layers), not new factories expanding capacity. With bit demand surging 21%, this structural supply constraint means persistent undersupply all year, supporting ASP’s 40% surge.

Who are the biggest winners?

Based on the above, JPMorgan conducts a detailed competitive analysis of global leading storage chip manufacturers and gives a clear investment ranking.

Asia’s top pick: Kioxia 

Kioxia is JPMorgan’s top pick for this round in the NAND cycle.

  • Unique technology (CBA architecture): The report highly praises Kioxia’s exclusive CBA (CMOS Bonded Array) architecture. Unlike competitors, Kioxia manufactures the peripheral logic (CMOS) and storage arrays on separate wafers, then connects them through hybrid bonding. Logic circuits can use advanced processes (like 28nm/14nm), enabling high I/O speeds, perfectly matching AI inference’s need for maximal data throughput.
  • Server share surges: With BiCS 8 mass production, Kioxia’s server business revenue share will soar from 20% in 2023 to 61% in 2027. This structural customer shift brings huge profit flexibility.
  • Purity and valuation: As a pure NAND maker, Kioxia directly benefits from NAND price increases, and versus the already-risen SK hynix, still has attractive valuation.

The long-term king: SK hynix

JPMorgan is "long-term bullish" on SK hynix, mainly thanks to its acquired subsidiary Solidigm.

  • Absolute QLC dominance: By acquiring Intel’s NAND business (now Solidigm), SK hynix inherits floating gate technology’s QLC advantages. Solidigm has almost no rivals in the 30TB/60TB ultra-large eSSD market, with strong pricing power.
  • Dual engine: SK hynix is the only supplier leading both in AI training (No. 1 HBM market share) and AI inference (QLC eSSD leadership).

Near-term catch-up: Samsung Electronics

On memory giant Samsung, JPMorgan’s thesis is "short-term catch up."

  • Challenger’s flexibility: Samsung started QLC and eSSD a bit behind SK hynix, but its vast production capacity matters. The report notes Samsung is ramping up V9 QLC mass production, with a chance to recapture market share by 2026.
  • Valuation rebound: Given Samsung’s stock lagged SK hynix and Micron, while its server business revenue share is projected to climb from 29% in 2023 to 66% by 2027, there’s good value for near-term trades.

Differentiated competition: Micron Technology

Micron is taking a differentiated approach. The report says its 6500 ION eSSD series uses 232-layer TLC tech to attack QLC market leaders, positioning as "TLC performance at QLC price." JPMorgan believes Micron will continue to benefit from strong demand at US data centers, maintaining an "overweight" rating.


Caption: Since January 2025, despite overall gains in memory equities, pure DRAM and pure NAND stocks have begun to diverge in performance. JPMorgan believes pure NAND plays (like Kioxia) will show stronger elasticity in the coming cycle.

Risks and opportunities coexist

Even with the major upcycle established, JPMorgan notes possible risks at the report’s end.

The key concern comes from weak demand for consumer electronics (smartphones, PCs). Growth in smartphone and PC shipments remains sluggish. If NAND prices rise too quickly and cause BOM (bill-of-materials) costs to jump, terminal vendors may downgrade specs (e.g., cut 512GB back to 256GB), or further dampen consumer upgrade motivation.

The report shows NAND cost as a share of notebook ASP is projected to rebound from the low of 4.3% to above 10%, hitting PC maker margins hard.

Nevertheless, JPMorgan’s conclusion is firm: This is a “supercycle” jointly created by a structurally explosive AI inference demand and structurally constrained supply. NAND is no longer subordinate to DRAM, but is the indispensable “hot data” reservoir in AI infrastructure. For investors, looking for manufacturers with high eSSD revenue share and leading technical approaches will be key to catching this wave.

 

~~~~~~~~~~~~~~~~~~~~~~~~

The wonderful content above comes from Windchaser Trading Desk.

For more detailed interpretations, including live commentary and frontline research, please join [Windchaser Trading Desk▪Annual Membership]

Risk Warning and DisclaimerThe market involves risks, and investments need to be made cautiously. This article does not constitute individual investment advice, nor does it consider specific users’ investment goals, financial status, or needs. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular situation. Investment decisions made on this basis are at your own risk.