Nvidia’s latest move on Rubin Ultra has rattled the memory trade. According to the source article, the company sharply reduced the memory specification for the next-generation flagship GPU on Aug. 7, cutting the mainstream HBM setup from a planned 12-Hi stack at roughly 384GB to an 8-Hi, 192GB version. That leaves memory per card close to half the earlier plan. The GPU package was also trimmed from 4-die to 2-die, while power was reduced from 2300W to 1800W.
The article says Nvidia internally tested at least three lower-memory variants over recent weeks. Once the report surfaced, SK hynix’s U.S.-listed ADR fell 4.97% that day, and its South Korean shares dropped more than 10% in the following session.
That reversal stands out because HBM had previously been framed as one of Jensen Huang’s biggest operational headaches. The source text says Huang traveled across Asia with his team in June to secure HBM supply. By early August, though, Nvidia was no longer leaning only on supply expansion. It was also reconsidering how much memory its next flagship product actually needed.
A common pattern across memory names
The article uses Sandisk as a clean read-through for the broader cycle, especially on the NAND side. Two days before the Nvidia report, Sandisk released what the piece described as the strongest earnings report in its history. For fiscal 2026 fourth quarter, revenue reached $8.97 billion, up 372% from a year earlier and 51% from the prior quarter, around 7% ahead of market expectations. Non-GAAP EPS came in at $39.25, beating analyst estimates by 14%.
Quarterly net profit reached $6.9 billion. Full-year net profit was $11.4 billion, up 797% year over year. Sandisk also announced a $14 billion share buyback plan, while gross margin stayed in the 83% to 85% range, a record level for the company.
Yet the stock reaction moved in the other direction. The source says Sandisk fell more than 8% in after-hours trading at one point and closed down about 5%. Before earnings, the stock had already dropped around 40% in July alone and had fallen more than 55% from its June record high at the trough.
The issue, in the article’s telling, is not the absolute level of profit but the speed of growth. On its earnings call, Sandisk guided fiscal 2027 first-quarter revenue to $10.3 billion to $10.8 billion, with a midpoint of about $10.55 billion. The market had been looking for roughly $10.8 billion to $11.16 billion. That left the midpoint around 2% to 5.5% below expectations.
The change in quarterly momentum is what mattered. Sandisk’s fourth-quarter revenue rose 51% sequentially, but using the midpoint of first-quarter guidance, sequential growth would slow to about 17.6%. The company also disclosed that roughly two-thirds of fourth-quarter sequential revenue growth came from pricing, while only one-third came from higher shipments. In other words, if price increases stop doing most of the work, growth can cool much faster than the headline profit figure suggests.
The article argues that this is exactly how cycle names begin to lose support in the market. Stocks are not being priced on how much money they made in the just-finished quarter. They are being repriced on what the next leg of growth may look like.
SK hynix and Samsung Electronics have run into the same problem. On July 29, SK hynix reported fiscal 2026 second-quarter revenue of KRW 79.32 trillion, up 257% year over year, and operating profit of KRW 60.54 trillion, up 557%. Operating margin hit 76.3%, and gross margin reached 83%. The source text says those figures set new records for the memory industry and notes that SK hynix generated more operating profit in a single quarter than the KRW 47.2 trillion it earned in all of 2025.
Even so, the stock dropped sharply. According to the article, SK hynix fell more than 19% intraday on the day of the earnings release, marking its biggest single-day decline on record. By the week cited in the report, the stock was down 25% cumulatively, wiping out around KRW 308 trillion in market value, or roughly RMB 1.45 trillion.
The article attributes that selloff to results falling short of what the market had already built in. Revenue expectations had been around KRW 85 trillion, versus actual revenue of KRW 79.3 trillion. The reported gap was linked to delayed HBM4 shipments, a weaker product mix, and a lower-than-expected share of high-end products.
Samsung’s setup looked similar. The article says Samsung’s late-July second-quarter report showed operating profit of KRW 89.5 trillion, up 1814% year over year, and net profit of KRW 71.6 trillion, up 1299%. Despite those historic figures, Samsung shares kept falling, down about 17% in July alone, with around KRW 244.8 trillion in market value erased. From the June highs referenced in the story, Samsung and SK hynix together had lost more than RMB 4 trillion in market value.
The “guidance cliff” in NAND and pressure on HBM
The report frames this as a classic cycle-stock setup: the top is not usually marked by peak profits, but by the point where growth starts to slow. Sandisk’s first-quarter guidance is presented as one of the clearest signals yet that the NAND pricing cycle may be nearing its peak. Fourth-quarter sequential growth was 51%; first-quarter guidance implies about 17% at the midpoint. That deceleration is substantial, and much of the preceding quarter’s growth came from price rather than volume.
TrendForce, as cited in the piece, expects DRAM contract prices in the third quarter of 2026 to rise another 13% to 18% sequentially, while NAND flash contract prices are expected to increase 10% to 15%. But the NAND gain is already well below an earlier optimistic view of 21%. TrendForce also said NAND Flash supply should loosen in the second half of 2027, which could put prices under correction pressure.
DRAM looks somewhat better in that outlook because HBM continues to crowd out capacity and AI server demand remains firm. Still, the article says TrendForce does not expect that condition to hold indefinitely.
Rubin Ultra changes the HBM math
If Sandisk’s guidance is the article’s signal for a NAND peak, the Rubin Ultra memory reduction is portrayed as a direct hit to HBM expectations. TrendForce later said Nvidia is evaluating several parallel options, including HBM4e 8hi, HBM4 12hi, and HBM4 8hi, and that the final specification will not be decided until validation in the second half of 2026.
That matters because HBM is the highest-priced and highest-margin category in memory, and the article describes it as the main engine behind the earnings surge at the Korean suppliers. SK hynix holds about 70% of the HBM market by the numbers cited in the report, while Samsung holds about 20%.
The valuation case for SK hynix had been built around a simple line of thinking: each AI GPU generation uses more HBM, the memory value per card keeps rising, and the HBM growth curve stays steep. SemiAnalysis, according to the source, estimated that with continued HBM price increases, the bill of materials for a Rubin Ultra rack had already climbed from $6.6 million to $8 million, with HBM alone accounting for about 28%.
404K Research estimated Rubin Ultra would make up around 20% of global HBM demand in 2027. If the memory capacity per unit falls from 384GB to 192GB, the article says that would be equivalent to roughly 10% less global HBM sold.
The mix shift may matter even more than the unit count. Rubin Ultra had originally been planned around 12-layer HBM4e, but the mainstream version under discussion now points to 8-layer HBM4. The source says a single HBM4e die is priced 20% to 30% above HBM4, and a 12-layer stack uses 50% more chips than an 8-layer stack. SemiAnalysis therefore concluded that the HBM value per GPU would fall by well over 50%, and HBM spending per rack would shrink from $2.24 million to about $900,000.
In the article’s account, HBM4e is one of SK hynix’s most profitable product lines because yields on its 12-layer stack are low, competition is limited, and pricing power is strong. If Nvidia’s mainstream products shift toward HBM4 8-Hi, the average selling price falls, and Samsung and Micron would become relatively more competitive.
The source also recalls remarks made on July 10, when SK hynix CEO Kwak Noh-Jung said in an interview that 2027 would see the worst supply shortage in memory industry history, with demand outstripping capacity even after aggressive expansion, and the imbalance lasting beyond 2030. The market had largely assumed HBM pricing would keep moving in one direction. Rubin Ultra’s reported redesign has forced a rethink.
China’s domestic substitution story has limits
No discussion of global memory names is complete without domestic substitution in China, and the article gives that topic a full section. When SK hynix and Samsung began sliding in July, one line of speculation was that Chinese memory suppliers could eventually break the Korean duopoly in key segments.
CXMT listed on Shanghai’s STAR Market on July 27. The stock rose 465.82% on its first day, and its market capitalization surpassed RMB 3.28 trillion in less than 24 hours, according to the article.
Counterpoint Research data cited in the piece shows CXMT’s share of the global DRAM market rising from 3% in the first quarter of 2025 to 8% in the first quarter of 2026, making it the world’s fourth-largest DRAM supplier. The report says that is enough to confirm a meaningful breakthrough from zero to one for China’s domestic memory industry, but not enough to challenge the top three suppliers yet.
On technology, the article says CXMT has achieved scaled production of DDR5 and LPDDR5. Its in-house 4F process has reached 17nm without relying on EUV lithography. The piece describes that level as roughly where the Korean suppliers were about two years earlier. It is competitive in mainstream consumer and server DRAM, but HBM remains a different story.
HBM requires more advanced process technology and more complex packaging, and the source says CXMT does not yet have the ability to mass-produce HBM at scale. Capacity expansion, though, is moving quickly. By the end of 2026, monthly wafer starts are expected to exceed 300,000. A Xinhua report from mid-July, as cited in the article, said both CXMT and YMTC are building new fabs, and their combined capacity after 2027 is expected to be more than double current levels.
The source adds that CXMT currently operates three 12-inch wafer fabs in Hefei and Beijing, with monthly capacity of about 280,000 to 300,000 wafers, and that figure is expected to reach 300,000 to 350,000 by the end of 2026.
Apple also appears in the report as a major prospective customer in the broader memory market. The article says Apple has been in talks with CXMT over mobile DRAM procurement, mainly LPDDR5X, as it looks to lower the manufacturing cost of the next iPhone generation. The strategy described is familiar: bring in a new supplier to gain leverage over incumbent vendors such as Samsung and SK hynix.
But the article says CXMT rejected requests for lower pricing and made clear that its DRAM quotes for comparable products would not come in below Samsung or SK hynix, with some higher-end models priced above them. Most of CXMT’s capacity, according to the piece, has already been locked up by major domestic hardware companies. It also says cloud players including Tencent and ByteDance are competing for supply, leaving CXMT’s roughly 300,000 wafers per month essentially fully loaded.
The original piece was carried by MarsBit and attributed to the WeChat public account “Zimu AI,” written by Miao Zheng.