NVIDIA's HBM Cut Shakes Global Memory Stocks, Except for Yangtze Memory

HBM has always been Huang’s biggest headache. To secure HBM supply, in June this year, Huang traveled across Asia with his team.

However, Huang Renxun suddenly changed his stance; on August 7, he significantly reduced the memory configuration of NVIDIA’s next-generation flagship GPU, Rubin Ultra.

According to leaks, over the past several weeks, NVIDIA internally tested at least three low-memory variants. Ultimately, they reduced the HBM configuration for the mainstream model from the originally planned 12-Hi (~384 GB) to 8-Hi (192 GB), nearly halving the memory per card. The GPU chiplet count was also reduced from 4-die to 2-die, and power consumption dropped from 2,300 W to 1,800 W.

Upon the announcement, SK Hynix's U.S. ADR fell 4.97% on the same day, and its Korean shares dropped over 10% the following day.

It’s bad enough that sellers of HBM are struggling, but the key issue is that SanDisk, which sells NAND, is also suffering.

Two days before the news about NVIDIA was released, SanDisk issued its strongest earnings report ever.

In the fourth quarter of fiscal year 2026, the company's revenue reached $8.97 billion, surging 372% year-over-year and increasing 51% quarter-over-quarter, exceeding market expectations by approximately 7%; non-GAAP earnings per share were $39.25, surpassing analyst estimates by 14%.

In fact, due to rising AI demand for hardware, the strong performance of memory stocks is no longer new.

SanDisk reported a quarterly net profit of $6.9 billion and an annual net profit of $11.4 billion, representing an astonishing 797% year-over-year increase. The company also announced a $14 billion share repurchase program, with gross margins remaining at the historical high of 83% to 85%.

Looking solely at these numbers, this is undoubtedly SanDisk's most impressive performance since its founding. However, the market's response was a post-market drop of over 8%, ending with a final decline of about 5%.

Even before the earnings report, in July, SanDisk's stock had already dropped approximately 40% for the month, with a maximum drawdown of over 55% from its June all-time high.

Record profits, plummeting stock prices—this seemingly contradictory scenario is currently unfolding across the global memory industry.

From Samsung to SK Hynix, from Micron to Kioxia, nearly all major memory chip manufacturers are experiencing the same script.

Profit hit a record high, yet the stock continued to decline. SanDisk’s latest earnings report was just another repeat performance. But its value lies in the fact that, as a pure NAND player, SanDisk’s performance and guidance provide a clear mirror for observing the current position of the memory industry cycle.

SanDisk provided revenue guidance for the first quarter of fiscal year 2027 during its earnings call. The company expects Q1 revenue to range between $10.3 billion and $10.8 billion, with a midpoint of approximately $10.55 billion, while the consensus market expectation is between $10.8 billion and $11.16 billion.

The guidance's midpoint was about 2% to 5.5% lower than market expectations—a modest margin, but one that carries weight given current market sentiment. Click here Missed opportunities are infinitely amplified.

SanDisk's Q4 revenue increased by 51% quarter-over-quarter, but according to the midpoint of Q1 guidance, the sequential growth rate is expected to drop sharply to approximately 17.6%. Meanwhile, SanDisk disclosed that about two-thirds of the Q4 sequential revenue growth came from price increases, with only one-third attributable to higher shipment volumes. This means that once price increases stall, the growth rate will decline even more rapidly.

So even though SanDisk was earning more and more, its growth rate suddenly slowed down. The stock market trades on expectations—investors buy your stock not based on how much you’re earning now, but on how fast you’re expected to grow in the future. When the growth rate slows, expectations change, and the stock price falls.

On July 29, SK Hynix released its second-quarter 2026 financial results, reporting revenue of KRW 7.932 trillion, a 257% year-over-year increase; operating profit of KRW 6.054 trillion, a 557% year-over-year surge; an operating profit margin of 76.3%; and a gross margin of 83%.

Undoubtedly, these figures have set new historical records for the storage industry, even surpassing NVIDIA’s 75% gross margin. As a hardware company, SK Hynix’s gross margin has now reached parity with software giants like Meta.

Additionally, SK Hynix achieved an operating profit of 60 trillion won in a single quarter, surpassing the full-year profit total of 47.2 trillion won for 2025.

On the day the earnings report was released, Hynix's stock price plunged more than 19% in a single day, marking its largest single-day drop in history, with a cumulative decline of 25% for the week, erasing approximately 30.8 trillion Korean won, equivalent to about RMB 1.45 trillion.

The reason is again "below expectations." The market had forecast Hynix's Q2 revenue at 85 trillion KRW, but the actual figure was only 79.3 trillion KRW, a shortfall of approximately 5.7 trillion KRW. This was due to delays in HBM4 shipments, a weaker product mix, and lower-than-expected proportion of high-end products. Under the spotlight, shortcomings are magnified, creating various justifications for selling.

Samsung Electronics experienced a similar situation. In its second-quarter earnings report released at the end of July, Samsung reported an operating profit of 89.5 trillion Korean won, a year-over-year increase of 1,814%; net profit reached 71.6 trillion Korean won, up 1,299% year over year.

The numbers are also historic, but after the earnings report, Samsung's stock continued to decline, falling approximately 17% for the month of July alone, with a market capitalization loss of about 24.48 trillion Korean won. Since its June high, the combined market capitalization loss of Samsung and SK Hynix has exceeded 4 trillion RMB.

The market prices not current profits, but changes in profits. When a company’s profits rise from $100 million to $1 billion—a 900% increase—its stock price surges; when profits climb from $1 billion to $6 billion—a 500% increase—the stock may still rise; but if the market expects the next quarter’s profit growth rate to decline from 500% to 200%, even if absolute profits hit a new high, the stock price will begin to fall.

The peak of cyclical stocks is never when profits are highest, but rather when growth begins to slow.

This is precisely what is meant by the "cliff edge": earnings are still growing, but the growth rate suddenly plummets, causing an immediate reversal in market expectations for the future.

SanDisk's Q1 guidance is the first clear signal of a cliff on the NAND track. The growth rate dropped sharply from 51% quarter-over-quarter in Q4 to just 17% in Q1—a decline so steep that the market believes the NAND price increase cycle has come to an end.

As mentioned earlier, two-thirds of SanDisk's Q4 growth came from price increases, meaning that pricing contributed the majority of the growth.

But prices can't rise forever—today, various hardware products in the market have already seen dramatic price increases due to AI-driven demand for production capacity. If prices continue to rise, downstream customers' purchasing power will hit its limit, prompting an increasing number of alternative solutions to emerge, because even the wealthiest households are running out of grain.

Although Samsung and SK Hynix's HBM businesses are still growing, the growth rate has begun to slow, and the delay in HBM4 shipments has cast a shadow over the growth momentum of high-end products.

TrendForce forecasts that in the third quarter of 2026, DRAM contract prices will still rise sequentially by 13% to 18%, while NAND flash contract prices will increase sequentially by 10% to 15%. However, the NAND increase is significantly lower than the previous optimistic expectation of 21%.

TrendForce stated that NAND Flash supply is expected to become more relaxed in the second half of 2027, potentially putting downward pressure on prices. DRAM conditions are slightly better due to ongoing capacity constraints from HBM demand and strong demand for AI servers.

However, TrendForce also believes this improvement won’t last long, as markets have their limits—DRAM is not a consumable, and there’s no reason for the market to sustain continuous demand for it.

If SanDisk's guidance cliff signaled the peak of NAND cycles, then NVIDIA's reduction of memory configuration in Rubin Ultra is undoubtedly a devastating blow to the memory market.

TrendForce subsequently stated that NVIDIA is concurrently evaluating multiple configurations, including HBM4e 8Hi, HBM4 12Hi, and HBM4 8Hi, with the final specifications to be confirmed only after validation in the second half of 2026.

As soon as this news broke, Samsung and SK Hynix were left stunned.

HBM is the highest-priced and most profitable category of memory chips, and it has been the primary driver of profit growth for Korean manufacturers during this super cycle.

SK Hynix currently holds approximately 70% of the HBM market, while Samsung holds about 20%.

Previously, the market's high valuation of Hynix was entirely based on HBM, as the HBM capacity for AI chips has increased with each generation, continuously raising the value per card, and HBM's growth curve has remained steeply upward.

According to SemiAnalysis, driven by ongoing HBM price increases, the BOM cost of the Rubin Ultra cabinet has risen from $6.6 million to $8 million, with HBM alone accounting for approximately 28%.

For HBM manufacturers, there is still a significant difference in revenue and profit between selling eight chips and selling sixteen chips.

NVIDIA's move has essentially shattered the HBM market.

404K Research estimates that Rubin Ultra accounts for approximately 20% of global HBM demand in 2027; if its memory is reduced from 384GB to 192GB, it would equate to a 10% decrease in global HBM sales.

But the core issue isn't here. Originally, Rubin Ultra was planned with a 12-layer HBM4e stack, but the mainstream version has now been reduced to an 8-layer HBM4. The unit price of HBM4e is 20% to 30% higher than HBM4, and a 12-layer stack requires 50% more chips than an 8-layer stack.

When combined, Semi Analysis states that the value of HBM per GPU has decreased by more than 50%, reducing HBM expenditure per rack from $2.24 million to approximately $900,000.

HBM4e is SK Hynix's most profitable product line due to its low yield for 12-layer stacking and limited competition, enabling extremely strong pricing power.

If NVIDIA's flagship models switch to HBM4 8-Hi, not only will the unit cost decrease, but Samsung and Micron's competitiveness will also improve relatively.

On July 10, during the day SK Hynix went public on Nasdaq, CEO Kwon Noh-jung stated in an interview that 2027 will be the most severe supply shortage in the history of the memory industry, with supply-demand imbalance persisting beyond 2030; customer demand far exceeds production capacity, even if the company pushes its expansion efforts to the maximum.

The market has generally assumed that HBM prices will continue rising. However, it now seems time to hit the brakes.

When discussing global memory stocks, there is one variable that cannot be avoided: domestic substitution.

When SK Hynix and Samsung began their steep decline in July, rumors circulated that domestic storage manufacturers would break the dominance of SK Hynix and Samsung.

ChangXin Technology listed on the STAR Market on July 27, surging 465.82% on its first day, surpassing a market capitalization of RMB 3.28 trillion in less than 24 hours.

To understand the real impact of domestic substitution on the global memory landscape, we first need to understand where CXMT stands.

According to Counterpoint Research, CXMT's market share in the global DRAM market increased from 3% in Q1 2025 to 8% in Q1 2026, making it the fourth-largest DRAM supplier worldwide.

Although this demonstrates that domestic storage has achieved a breakthrough from 0 to 1, it still has a long way to go before challenging the dominance of the three major players.

Technically, CXMT has now achieved mass production of DDR5 and LPDDR5, with its proprietary 4F process surpassing the 17-nanometer node, successfully avoiding the need for EUV lithography machines.

This technical level is roughly equivalent to that of Korean manufacturers two years ago and is already competitive in the mainstream consumer and server markets; however, CXMT still has a clear gap in the HBM field.

HBM is particularly specialized, requiring more advanced manufacturing processes and more complex packaging technologies; CXMT currently does not have the capability for large-scale mass production of HBM.

In terms of production capacity, CXMT is expanding rapidly. By the end of 2026, the company’s monthly wafer start capacity is expected to exceed 300,000 wafers.

According to a mid-July report from Xinhua News Agency, both CXMT and YMTC are building new factories, with overall production capacity expected to more than double after 2027.

Currently, CXMT operates three 12-inch wafer fabs in Hefei and Beijing, with a monthly capacity of approximately 280,000 to 300,000 wafers, expected to reach 300,000 to 350,000 wafers by the end of 2026.

Apple is undoubtedly a major client for the global memory industry, and now, in an effort to reduce the manufacturing cost of the next-generation iPhone, Apple has proactively engaged with CXMT to discuss procurement partnerships for mobile DRAM (primarily LPDDR5X).

Apple's intention was to secure lower prices from ChangXin compared to Samsung and SK Hynix, while also using ChangXin as leverage to pressure Korean companies into lowering their prices. This is a strategy Apple has used for over a decade: introducing new suppliers to pressure existing ones into offering better terms.

However, CXMT refused the price reduction request, clearly stating that its DRAM pricing for equivalent specifications would not be lower than that of Samsung and SK Hynix, with some high-end models even higher. In fact, the vast majority of CXMT’s production capacity has already been secured by major domestic hardware manufacturers, while cloud providers such as Tencent and ByteDance are also competing to secure supplies, leaving CXMT’s monthly output of 300,000 wafers nearly fully utilized.

This article is from the WeChat public account "Letter AI," authored by Miao Zheng.