There are likely many readers who braced themselves upon hearing that SSD prices are set to rise again. In the latter half of 2025, coordinated production cuts by NAND manufacturers were cited as the main culprit behind price hikes, and this was understood simply as a matter of tightened supply. However, the rebound in spot prices for 512-gigabit (Gb) TLC NAND dies that occurred from July to August 2026 cannot be explained by this understanding alone.
While the production shift itself is continuing, it isn't a new development. Manufacturers like Samsung and Kioxia have been converting production lines toward higher-value-added enterprise products over the past several months. But what drove the recent rebound was the technical factor of inventory adjustment, and behind it, a mid-term demand variable is gaining prominence: the new storage tier within NVIDIA's next-generation AI platform, "Vera Rubin."
512Gb Spot Price Bottoms Out and Reverses, From July's Low to $21 Range in August
According to TrendForce's weekly memory spot price series, 512Gb TLC NAND die prices continued a downward trend through June 2026 before sinking to $18.931 on July 20. This marked the bottom. The following week, on July 27, prices reversed to $19.250, a 1.69% increase week-over-week. This reversal continued, reaching $21.125 by August 9—a 4.97% increase week-over-week. This means prices recovered 11.6% in just three weeks from the July 20 low.
The pace of this increase differs in nature from the price hike phase seen in the latter half of 2025. Back then, it took several weeks from the announcement of coordinated production cuts by NAND manufacturers for the effect to be reflected in contract prices—a structure where supply reduction pushed prices up. This time, the spot price reversal is not accompanied by any production cut announcement.
Wccftech cites CMX demand as an explanation for the price rebound, but notes it remains only a "partial explanation," stopping short of identifying CMX as the primary cause. This is understandable, given that TrendForce itself explicitly stated regarding the July 27 rebound that it was driven by inventory replenishment from specific manufacturers rather than being supported by large-volume orders, and as of August 9 explained that consumption demand and trading volume remained sluggish. In other words, what directly explains the price movement over the past two weeks are technical factors like inventory and supply-demand dynamics, and CMX-derived demand itself has not been quantitatively confirmed. Still, the CMX demand variable cannot be ignored, because a structural tightness that cannot be fully explained by production shifts and inventory replenishment alone is progressing in parallel.
The timing of this tightening intensification from July to August falls just before the start of mass shipment for NVIDIA Vera Rubin (fall 2026, per NVIDIA's official announcement). It wouldn't be unnatural for GPU makers and server vendors to be stockpiling NAND ahead of mass production in July and August. One factor pushing up the scale of this procurement is the new storage tier known as CMX.
What Is CMX, and Why Does Vera Rubin Want This Much NAND?
CMX (Context Memory eXtension) is a new storage tier that NVIDIA revealed in technical detail in its March 16 official blog post, building on the inference-oriented new storage tier concept it announced in January 2026. It is designed to handle the KV cache (data that stores past inputs in memory) generated during large language model inference, and is classified into a new tier called "G3.5," positioned between HBM (High Bandwidth Memory) and general-purpose storage. NVIDIA's announcement shows a configuration where CMX connects to Spectrum-X Ethernet via a DPU (Data Processing Unit) called BlueField-4.
As inference conversations grow longer, the size of the KV cache balloons, eventually exceeding what GPU memory alone can handle. This cache bloat becomes especially pronounced in dialogues handling tens of thousands of tokens of long context, or in workflows where multiple agents operate in parallel. CMX is a mechanism that offloads this cache to flash storage, allowing long context to be handled without straining GPU-side memory.
Regarding specific capacity figures, NVIDIA CEO Jensen Huang, in his explanation at CES, presented a configuration where a single 2U rack-mounted CMX server is equipped with 600TB of TLC flash, with each BlueField-4 DPU managing 150TB. At the Pod level, combining multiple such units, the total flash capacity is said to reach 9,600TB (9.6 petabytes, an estimate reported by Korean media). NVIDIA's official blog itself only describes Pod-level capacity as "petabyte-class."
This figure operates on different premises than the roughly 1,152TB per NVL72 server estimate that we ourselves reported in January 2026. That earlier estimate was a rough calculation from Vera Rubin's early stage, whereas this CMX specification breaks things down more concretely into 2U server units, DPU units, and Pod units. While the capacity of a single CMX server remains at the terabyte scale (600TB), the Pod unit that bundles these together reaches the petabyte scale (9.6PB). The fact that the design jumps in order of magnitude at the level of the Pod—a shared infrastructure unit—rather than at the level of a single GPU server, is what connects to the purchasing power capable of moving the 512Gb die spot market.
The Trap of "512Gb" vs. "512GB": Why This Number Isn't the Capacity of a Retail SSD
The notation "512Gb" used repeatedly throughout this article is distinct from the "512GB" SSD consumers see on store shelves. It's confusing, but misunderstanding the unit difference risks misinterpreting the meaning of this price increase.
The "512Gb" TrendForce uses as an indicator represents the storage capacity of a single NAND flash "die," expressed in gigabits (Gb). A die refers to the smallest semiconductor chip unit cut from a silicon wafer during the NAND memory manufacturing process. Since a bit is one-eighth of a byte, a 512Gb die converts to 64GB of capacity. Meanwhile, the "512GB" SSD sold in stores is a module product that stacks eight or more of these dies together, with units expressed in bytes rather than bits. Even though both share the number "512," one is an indicator for a component—the die—and the other is the capacity of a finished product; they are numbers from entirely different layers.
The reason the industry deliberately tracks die-level spot prices as an indicator is simple: dies are an intermediate good before being incorporated into SSDs or enterprise storage, making them the place where price fluctuations appear earliest. Finished SSD prices incorporate assembly costs, distribution margins, and inventory absorption speed, so it takes time for movements in the raw material—the die—to be reflected. Conversely, the spot price of the 512Gb die functions as a leading indicator for how consumer SSD prices will move months later. This is why the current rebound carries significance.
NAND Manufacturers' Production Shift, and the Search for a Culprit Behind the Price Increase
NAND manufacturers have been advancing a production shift toward enterprise SSDs and QLC (a method recording 4 bits per cell) variants since the latter half of 2025. Samsung began mass production of its PCIe 6.0-compatible enterprise SSD "PM1763" in July 2026, and the Kioxia-SanDisk alliance began sample shipments of its next-generation 3D NAND "BiCS10" (1Tb TLC product) at its Kitakami plant's Fab2 on July 3, 2026. As for SK hynix, it has been reported that its 2026 NAND production capacity has already sold out. All of these represent moves to redirect production lines toward higher-margin AI and enterprise products.
Bearing the brunt of this shift are TLC and pSLC (pseudo-SLC) variants used in industrial and automotive equipment. According to one industry report, lead times (the period from order to delivery) for these variants can exceed 20 weeks depending on the case. If lines are preferentially allocated to high-value-added AI products, it's a natural consequence that supply of general-purpose variants tightens—this is exactly the structure that has continued since the latter half of 2025. The winners are NAND manufacturers like Samsung, SK hynix, Kioxia, and SanDisk, along with NVIDIA, while the losers are consumer SSD makers, DIY PC users, and industrial equipment manufacturers unable to predict delivery times.
The production shift itself is an existing factor continuing from 2025, and alone it cannot explain the price decline that persisted through June. Meanwhile, CMX, since its March 2026 announcement, sees actual procurement activity kick in just before Vera Rubin's mass shipment approaches in fall 2026—that is, from July to August—which nearly overlaps with the timing of this rebound.
Another clue is TrendForce's forecast that Q3 2026 NAND flash contract prices will rise 10-15% (quarter-over-quarter), with the pace of increase slowing compared to the first half of 2026. This asymmetry—where contract price growth is decelerating while only spot prices rebounded—cannot be fully explained by short-term factors like inventory replenishment alone. While TrendForce itself has not explicitly cited CMX demand as a reason, and no definitive conclusion can be drawn, the possibility cannot be ruled out that, amid a situation where the production shift is thinning the supply foundation, rising NAND demand ahead of Vera Rubin's mass production triggered a reversal in inventory.
An analyst going by the name Jukan posted on social media a view that NVIDIA's Rubin CMX would "literally absorb NAND supply like a sponge absorbing water" regarding NAND demand for CMX, assessing that the scale would be equivalent to Apple's entire annual NAND procurement volume on its own. According to Korean media reporting cited by Jukan, CMX-related NAND demand is expected to surge from 35 million TB in 2026 to over 100 million TB in 2027, though this figure is not Jukan's own estimate. As we have not been able to access the original Korean media report, this figure should be treated only as a reference point for grasping the scale. Nevertheless, given that the actual measured value of 512Gb die spot prices has already shown a rebound, the direction is not inconsistent with this assessment.
Spillover to Japan: When Will Consumer SSD Prices Move?
The spot price of 512Gb dies is an intermediate industrial good, not a figure Japanese consumers directly see. At the August 9 figure of $21.125, converted to yen (at approximately 1 dollar = 159 yen), that comes to about ¥3,359, while the July 20 low of $18.931 comes to about ¥3,010. A roughly ¥349 increase per die over three weeks may look small, but when this accumulates across dozens of units in enterprise SSDs, or in the petabyte-scale procurement of CMX servers for Vera Rubin, the impact becomes a non-negligible sum.
The timing of spillover to consumer SSD prices can be inferred from past price-hike phases. When coordinated production cuts by NAND manufacturers came to light in November 2025, we ourselves had reported at the time that market prices could rise by as much as 40-50%. About two weeks later, on December 1, 2025, mynavi.jp reported, citing TrendForce data, a contract price increase of over 60%, meaning it took less than a month for the production-cut information to be reflected in contract prices. From there, it took roughly two months from the production-cut announcement for consumer market supply concerns to be widely reported, culminating in our own January 2026 article that described it as a "giant tsunami."
This time, the sequence of events differs from that precedent. TrendForce had already forecast as of July 3 that Q3 2026 contract prices would rise 10-15% quarter-over-quarter, and this forecast preceded July 20, when spot prices bottomed out. The order this time is that contract price increase observations came first, with the spot price rebound catching up afterward to corroborate it. Applying the precedent's lead time—"less than a month from announcement to contract price reflection, followed by another 1-2 months before consumer supply concerns become visible"—suggests spillover to consumer SSD prices could surface between Q4 2026 and Q1 2027.
This spillover lag arrives at nearly the same answer even when calculated from a different starting point. TLC and pSLC variants for industrial and automotive use currently have lead times exceeding 20 weeks depending on the case, a delay directly attributable to AI-priority allocation being layered onto general-purpose variant delivery times. Adding 20 weeks (roughly 4.6 months) from mid-August lands in late December 2026 to January 2027, nearly overlapping with the "Q4 2026 to Q1 2027" range derived from the announcement-based approach. The fact that two different approaches—tracing the speed of reflection from announcement, and working backward from priority-allocation delivery delays—point to nearly the same period reinforces confidence in this spillover timeline. Domestic PC makers and SSD vendors should also be watched for moves to pass on rising procurement costs during this period.
What Will Be Confirmed Before Fall's Mass Production Start
NVIDIA has officially announced that Vera Rubin's mass shipment will begin in fall 2026. Whether this current rebound in 512Gb die prices represents a temporary pull-forward in procurement, or a structural demand that continues even after mass production begins, can only be answered once actual shipment volumes this fall and the accompanying pace of CMX server deployment become clear. BofA's observation regarding a "despec" for Vera Rubin Ultra's HBM4E capacity—reducing the amount mounted—is also worth noting. This concerns supply and implementation constraints on the HBM (High Bandwidth Memory) side, a separate memory category from NAND. It would be premature to conclude this represents a permanent shift toward lower-capacity SKUs, and some views suggest it's merely a temporary response to short-term constraints, but the fact that memory supply is beginning to spill over into GPU-side design decisions on both the NAND and HBM fronts is worth watching as a sign of broader memory tightness across AI infrastructure as a whole.
NAND manufacturers' production conversion plans are also not yet a settled factor. Samsung's PM1763 has only just begun mass production, and Kioxia and SanDisk's BiCS10 remains at the sample shipment stage. Specifications may still change once mass production begins, and how far supply volumes will actually build up requires observing production results over the coming quarters. Whether the 20-plus-week lead times seen in industrial and automotive applications will ease depends on how long AI-priority allocation continues.
If CMX establishes itself as a demand variable in the market, it would give NAND manufacturers further justification for investing in higher-value-added product lines. Whether this extends beyond simply repurposing existing production lines to investment decisions for new factories is also a point worth watching. The key markers for confirmation will be spot and contract price movements from this fall onward, and how far each company's production line conversion affects consumer supply—though the premises this time differ from the production adjustments of late 2025.
The production cuts in the latter half of 2025 were a deliberate supply adjustment on the production side, advanced alongside the parallel structural change of shifting lines from TLC to QLC. This time, on top of the continuing foundation of that shift in production mix, demand from a new procurement entity—Vera Rubin—is beginning to layer on top. What can be said from the data available at this point is that a new demand-side variable is beginning to be added to a price-hike structure that had previously been driven entirely by supply-side factors. The rebound in 512Gb dies itself is explained by technical factors like inventory and supply-demand dynamics, but whether this rebound becomes entrenched depends on how substantial this demand-side variable turns out to be.