Why have memory stocks become the first AI assets to be abandoned?

As of October 30, representative companies in the AI industry chain—including U.S. network equipment, power infrastructure, cloud providers, and optical communications firms—had fallen on average 22.0%, 29.4%, 33.9%, and 44.0%, respectively, from their intra-year highs.

By comparison, the 11 memory companies within the scope of the statistics saw an average drawdown of as much as 51.6%. In particular, the maximum declines of A-share companies Biwin Storage, Demingli, and Longsys exceeded 70%.

Even more surprising is that the performance of the memory companies remains astonishingly good.

In fiscal year 2026, Micron's net profit reached $84.969 billion, surging nearly ninefold year-on-year and hitting a record high. In the third quarter, the operating profits of SK Hynix's and Samsung's semiconductor businesses reached 83.33 trillion and 106.5 trillion won, respectively, also setting new records.

The performance of domestic module makers was even more exaggerated. In the first half of last year, Longsys, Biwin Storage, and Demingli together posted a loss of 329 million yuan, yet in the first half of this year they earned 23.76 billion yuan in one go.

In just one year, the companies went from collective losses to a surge in profit, yet their stock prices were halved one after another. Why has such an abnormal situation occurred? The answer lies precisely in the growth logic of this round of the memory super-cycle.

Over the past year, memory chip prices have risen all the way. Original manufacturers such as Samsung, SK Hynix, and Micron earned huge profits by raising prices, while downstream module makers, relying on low-priced inventory stockpiled in advance, enjoyed even greater profit elasticity.

It can be said that price increases were the biggest contributor to this round of profit explosion in the memory industry, and they are also the risk the market worries about most today.

On the one hand, the massive capital expenditures of cloud providers have begun to squeeze free cash flow, and future investment growth may gradually slow; on the other hand, Chinese memory manufacturers are accelerating capacity expansion, and the global market's supply landscape also faces new changes.

When the pace of memory price increases has already begun to narrow. For module makers that rely heavily on inventory price spreads, once the price-increase dividend weakens, the inventory leverage that previously amplified profits may in turn also amplify the pressure of falling profits.

The memory super-cycle may be far from over, but the capital market has already begun repricing: when the price-increase dividend fades, how much money can these companies actually still make?

The memory price-increase dividend begins to loosen

Over the past year, original memory manufacturers have made enormous profits, and the biggest contributor was price increases.

Based on a rough estimate from changes in shipment capacity and average selling prices of major manufacturers, price and product mix changes contributed about 81%–90% of the new revenue of Samsung, SK Hynix, and Micron.

Taking Micron as an example, in fiscal year 2026, Micron's revenue grew from $37.378 billion to $133.188 billion, up 256% year-on-year. But during the same period, DRAM (including HBM) and NAND shipment capacity each grew only about 25%.

In other words, Micron sold only a quarter more memory capacity, but revenue more than tripled. Roughly converted, its average selling price rose a full 2.85 times.

Now, the pace of price increases has begun to slow. According to data from TrendForce, the quarter-on-quarter increase in contract prices for ordinary DRAM has narrowed from 58%–63% in the second quarter to 13%–18% in the third quarter, and is expected to fall further to 10%–15% in the fourth quarter.

The NAND market has also shown signs of loosening. In October, contract prices for all three types of NAND wafers—MLC, TLC, and QLC—stopped rising and were flat quarter-on-quarter.

This change has already begun to show in the operating data of the original manufacturers.

In the fourth quarter of fiscal year 2026, SanDisk's revenue grew 51% quarter-on-quarter, of which about 67% of the incremental revenue came from price increases; by the first quarter of fiscal year 2027, revenue growth slowed to 23% quarter-on-quarter.

It should be emphasized that narrowing increases do not equal falling prices, but for memory companies that have already enjoyed huge price-increase dividends, this change is still worth watching.

More importantly, the two key conditions supporting price increases have also begun to change.

First is the demand side. Over the past two years, cloud providers have built AI data centers almost regardless of cost, and demand for hardware such as GPUs, memory chips, and servers has continued to explode.

But such massive investment has begun to affect the cash flow of tech giants.

In the third quarter of 2026, Google's free cash flow fell to negative $9.1 billion, negative for two consecutive quarters; Amazon's free cash flow over the past 12 months also turned from positive $14.8 billion in the same period last year to negative $42.4 billion.

Goldman Sachs expects the growth rate of cloud providers' capital expenditures to slow from 94% in 2026 to 50% in 2027, and further to 17% in 2028.

On the other hand, the supply side is also changing.

The most noteworthy is the capacity expansion of Chinese manufacturers. According to forecasts by semiconductor research firm SemiAnalysis, in 2026 ChangXin Memory will add about 85,000 wafers per month of DRAM capacity, equivalent to 81% of the combined new capacity of Samsung, SK Hynix, and Micron during the same period.

On October 10, ChangXin also announced key progress in a new-generation 4F² DRAM architecture. On the next trading day after the announcement, Samsung Electronics and SK Hynix fell 7.39% and 10.09%, respectively.

Why is the market so sensitive?

Because it is not the first time that Chinese manufacturing has changed the global market price system through technological breakthroughs and large-scale capacity expansion.

Photovoltaics is the most typical example.

According to International Energy Agency data, in 2010, China's photovoltaic module output accounted for about half of the global total; by 2024, that share had risen to 86.4%.

And along with the rapid expansion of Chinese manufacturing, global photovoltaic module prices also fell all the way.

Of course, DRAM and photovoltaics are not exactly the same. New capacity must go through process ramp-up, yield improvement, and customer certification, which does not mean it can immediately form effective supply.

But for overseas memory giants, the continued capacity expansion by Chinese manufacturers may indeed become an important variable affecting long-term pricing power.

On one side, the growth rate of capital expenditure by cloud vendors is facing a slowdown; on the other, Chinese manufacturers are accelerating the release of new capacity. The memory super cycle is not over yet, but the logic of continuous price increases based on tight supply and demand in the past is now facing new tests.

Module makers are the first to feel the pain

When the memory price dividend dissipates, downstream module makers are the first to feel the pressure.

Among the 11 memory companies surveyed, the three domestic module makers fell the hardest: Longsys, Biwin Storage, and Demingli.

In the third quarter of 2026, Longsys's revenue increased 3.2% quarter on quarter, but net profit attributable to shareholders fell 44.2%; Biwin Storage's revenue fell 29.9%, and net profit attributable to shareholders plunged 62.4%; although Demingli's revenue grew 24.2%, net profit rose only 2.2%.

The gross margins of the three companies also fell across the board. Longsys dropped from 61.3% to 44.7%, Biwin Storage from 56.0% to 42.8%, and Demingli from 38.8% to 32.9%.

Clearly the memory industry as a whole is still raising prices, so why is it becoming harder and harder for module makers to make money?

The answer lies in their special business model.

Unlike original manufacturers such as Samsung and SK Hynix, which possess wafer manufacturing capabilities, module makers need to purchase DRAM and NAND chips from upstream, and then process them into products such as SSDs and memory modules through controller chips, firmware development, packaging and testing, and other steps.

This business model has one characteristic: raw materials account for an extremely high proportion of costs.

Taking 2025 as an example, memory wafers accounted for 79.4% of Longsys's total procurement; direct materials accounted for 94.0% of Biwin Storage's production costs; and for Demingli's solid-state drive business, direct material costs accounted for as much as 96.2%.

This means that the profitability of module makers depends not only on their own processing and manufacturing capabilities, but also heavily on the procurement prices of upstream memory chips.

Over the past year, memory prices have continued to rise, and the low-priced chips purchased in advance by module makers can be sold at higher market prices.

Product prices have risen, but the costs carried forward are still the low prices from several months ago, and the difference naturally becomes profit. To put it plainly, module makers not only earn money from processing, but also earn a share from memory price increases.

But this dividend has a precondition: memory prices must continue to rise, and rise fast enough.

Now, the cost of newly purchased chips is constantly being pushed up, but the pace of memory price increases has begun to narrow. As the previous low-priced inventory is gradually consumed, profit margins are naturally squeezed.

Longsys also mentioned in its third-quarter report that, affected by the previously high price base, the price increases of DRAM and NAND have narrowed significantly, and the marginal contribution of price increases to the company's profitability is weakening.

Even more alarming is that the three module makers are still sitting on enormous inventories.

As of the end of September 2026, the combined inventory of Longsys, Biwin Storage, and Demingli reached about 81.5 billion yuan, an increase of 54.9 billion yuan from the end of 2025, a surge of 206%.

In the same period, the combined net cash flow from operating activities of the three companies was an outflow of about 26 billion yuan. That is to say, although book profits soared, these companies also invested a large amount of cash in purchasing and stockpiling chips.

Of course, an increase in inventory is not necessarily a bad thing. When orders are sufficient and prices continue to rise, stocking up in advance can lock in costs and create more profit.

But if prices stop rising in the future, or even turn downward, the situation will be completely different. Once market selling prices fall below procurement costs, module makers will not only lose the previous price-difference dividend, but may also face inventory write-down losses.

This is also why the stock price corrections of memory module makers are significantly larger than those of upstream original manufacturers.

Conclusion

More than two thousand years ago, a sentence in the Records of the Grand Historian: Biographies of the Money-Makers said: "When prices rise to the extreme, they turn cheap; when they fall to the extreme, they turn expensive."

Applied to today's memory industry, it is still worth pondering.

High prices bring high profits, high profits stimulate capital investment, and the ever-increasing capacity may in turn change the price trend.

This article comes from the WeChat public account "Silicon-Based Observation Pro", author: Silicon-Based Jun, published by 36Kr with authorization.