Why are original storage manufacturers not rushing to expand production?

AI-driven storage demand has surged, with NAND prices soaring more than 20-fold, but both flash memory manufacturers and hard drive makers remain restrained in large-scale capacity expansion. Behind this is a supply discipline shaped by economic logic, technical barriers, and industry experience.

According to Wind Chasing Trading Desk, at a recent industry expert webinar hosted by Bernstein Research, the surge in NAND prices is reshaping storage architecture choices in data centers. Former Western Digital Executive Vice President and SanDisk Flash Business Head Robert Soderbery pointed out that the current per-GB price gap between NAND and HDD has widened to about 20 to 25 times, far exceeding the 2-3x critical point required for hyperscale cloud providers to switch total cost of ownership (TCO), making NAND substitution for HDD economically unattractive.

This expanding price gap is prompting AI data center operators to reassess storage architectures, and some demand may flow back from NAND to HDD. Meanwhile, after years of losses, NAND manufacturers generally remain cautious about capacity expansion, while the HDD oligopoly is proactively maintaining supply discipline. For investors, this means that supply constraints for both types of storage assets are unlikely to break in the short term, and a strong pricing environment may persist.

AI reshapes storage architecture, NAND demand focuses on latest process nodes

The storage architecture of data centers has evolved from a traditional “two-tier model” to an AI-driven “three-tier architecture.” Robert Soderbery explained that traditional data centers used small-capacity enterprise SSDs for compute tasks and nearline HDDs for bulk storage, with HDDs accounting for about 80% to 85% of total bit volume.

The rise of AI data centers has disrupted this pattern. Data preprocessing in AI workloads—including vectorization, embedding, and other pre-compute tasks—places extremely high demands on storage performance, which traditional nearline HDD cannot meet, thus sparking explosive demand for large-capacity enterprise SSDs. Capacity specs have jumped from 2–8TB to 32TB, 64TB, even 256TB, with NAND’s share in new AI storage deployments reaching as high as 60%-70%.

However, NAND demand is not evenly distributed across all capacity. Robert Soderbery noted that the high-performance, large-capacity NAND needed for AI workloads must use the latest process nodes (currently 7th to 9th generation), and due to years of pressure and insufficient capital expenditure, these latest nodes account for only about 30%-35% of industry total capacity. The high demand concentrated in this limited capacity is the core reason for the sharp rise in NAND prices this cycle.

Return on expansion is insufficient, NAND manufacturers choose to wait

Given surging prices, why are NAND makers not accelerating expansion? The answer is that investment returns remain unattractive.

Robert Soderbery explained that the NAND industry has suffered continuous losses over the past five to nine years, leaving manufacturers with deep cyclical scars and lacking confidence in demand sustainability. It usually takes about 15 months from capital expenditure to capacity delivery—about a year for equipment, then several months for commissioning and ramp-up. This means prices must continue to rise and remain high for several quarters before manufacturers gain enough confidence to initiate expansion, with new capacity then taking about five quarters to come online.

More importantly, even before this round of price hikes, NAND makers’ gross margins were only around 20%, a level inadequate to support large-scale capital investment. Robert Soderbery pointed out current expansion focuses on migrating existing capacity from older nodes to more advanced nodes to better serve AI market requirements, rather than simply increasing total capacity. The industry typically targets an annual bit growth rate of 20%-30%, much of which actually comes from QLC technology improving density rather than increased wafer input.

HDD oligopoly proactively maintains supply discipline

Unlike NAND, supply constraints for HDDs come more from proactive choice than from technical bottlenecks.

The current HDD market has only two major players—Western Digital (WDC) and Seagate—with Toshiba's competitiveness relatively limited. Robert Soderbery bluntly stated, "The first rule of an oligopoly is not to expand capacity before achieving target profit margins."

HDD manufacturers have strong incentives to maintain this discipline. Over the past decade, the HDD industry has long faced overcapacity and is now highly wary of repeating past mistakes. Additionally, HDD technology is inherently complex—including HAMR (Heat-Assisted Magnetic Recording), magnetic heads, mechanical structures, and materials science—making capacity expansion frictional and objectively supporting supply constraints. Robert Soderbery revealed that Western Digital has more PhDs working on HDDs than on NAND, showing its technical intensity.

This means HDD makers should be seen as oligopolists capable of maintaining reasonable long-term returns, not just cyclical plays.

The long-term threat of NAND replacing HDD is overestimated

The market has long worried that NAND will gradually eat into HDD market share, but this logic faces significant economic obstacles.

Robert Soderbery calculates that for NAND to replace HDD in terms of TCO, the price gap must narrow to within 2–3x. To cut costs to that level, the industry needs multiple process node upgrades, with each node requiring around $50 billion in capital expenditure. Effective replacement of HDDs would require about $150 billion cumulative investment—excluding capital required for serving the AI market.

Meanwhile, NAND's cost decline curve is flattening. NAND has shifted from 2D to 3D stacking, but the marginal cost benefits from higher stack layers are diminishing, slowing the cost decline. By contrast, HDD's HAMR technology is accelerating cost reduction, making the cost curves for the two more convergent. Robert Soderbery believes the real inflection point for substantial NAND threat to HDD replacement is "still quite far away."

Long-term contract transformation improves industry planning visibility

Another significant change is that long-term purchase agreements (LTAs) are shifting from historically "one-sided contracts" to truly bilateral constraints.

Robert Soderbery explained that past LTAs effectively only bound the supplier—suppliers had to reserve capacity, but customers could refuse delivery if not needed, and suppliers would bear penalty risks if unable to deliver. This asymmetry made it hard for manufacturers to carry out effective capital planning.

SanDisk recently disclosed it had signed five new long-term agreements involving about $42 billion in procurement obligations, with $11 billion in financial guarantees. Robert Soderbery was positive about this, believing such bilateral contracts can greatly boost manufacturers' investment confidence and help the industry overcome supply shortages. However, he also pointed out that if the market declines sharply, customers still have an incentive to renegotiate by paying penalties, so the actual binding force of the contracts depends on market conditions. Taking SanDisk as an example, the $11 billion guarantee accounts for 25% of its procurement obligations, meaning even if prices fall 50%, the effective price cut would then only shrink by about 25%.

HBF: A new frontier for NAND in AI computing

Beyond storage architecture evolution, High Bandwidth Flash (HBF) is a technological opportunity for NAND to penetrate AI compute layers.

HBF borrows packaging techniques from High Bandwidth Memory (HBM), using a stacked structure and wide bus interface designed for deployment alongside GPUs. Its core advantage exploits the unique nature of AI inference workloads—during model inference, large numbers of coefficients stream into GPUs, rather than requiring frequent read-write. While NAND has slower dynamic read/write speeds and higher latency, once streaming begins, its throughput is impressive, making it well-suited for this scenario. Combined with HBM-like bandwidth and NAND’s higher storage density and differentiated cost structure, HBF has potential competitiveness in AI inference scenarios.

Robert Soderbery believes HBF is "very promising," but it's still in early stages, and whether it can meet commercial requirements in power, performance, and reliability remains to be seen. He says that regardless of whether HBF can directly replace HBM, there is still much innovation potential for NAND in AI pipelines, and investors can treat it as a potential upside option.

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

The above content is from Wind Chasing Trading Desk.

For more in-depth interpretation, including real-time insights and frontline research, please join the [Wind Chasing Trading Desk Annual Membership]

Risk Warning and DisclaimerThe market involves risk, investment should be cautious. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of any individual user. Users should consider whether any opinions, views, or conclusions in this article fit their particular circumstances. You are solely responsible for any investment made on this basis. ```