The recent storage market has been difficult to quote
Old Feng, a storage distributor in southern China, exclaimed to the media, “Storage chips are priced almost at a rate of one per minute. Some distributors only provide quotes to the public at 2 pm every day, afraid that the prices quoted in the morning will be lost by the afternoon
This phenomenon of “multi value in one day” is a microcosm of the severe fluctuations in the global storage market before and after National Day.
In early September, SanDisk announced a 10% increase in the price of NAND Flash; Micron closely followed suit, suspending quotes for DDR and LPDDR series, and stating that future price increases may reach 30%; Module manufacturers Weigang and Shiquan also announced a rare suspension of quotations – this is the first time since 2017 that storage module manufacturers have collectively “stopped reporting”.
Stimulated by intense price signals, downstream manufacturers and buyers actively inquired about orders and stocked up before the holiday, further boosting market heat.
At the same time, distributors have become the “amplifiers” of this market trend, with some channel partners choosing to temporarily suspend shipments or even “raise prices during holidays”. Lao Feng revealed, “Some distributors simply let their employees rest, pay their wages as usual, and wait until the price further increases before selling the goods
Chen Dong, a senior executive in the storage industry, told Leifeng.com: “This is a typical ‘buy up, not buy down’ behavior. In addition to normal demand, the purchasing side will also reserve a large amount of strategic inventory, and the demand is expected to be further released on a large scale in Q4 of this year and the first half of next year. ”
This expected demand forms a self reinforcing dynamic cycle:
Demand side: Expected price increase → Strategic stocking → Consumption of channel inventory
Supply side: Anticipating strong demand → Holding back on sales, controlling quantity and raising prices → exacerbating supply tension
Market results: worsening supply-demand imbalance → accelerated price increase → strengthened expectation of price increase
In this cycle, prices are constantly raised and supply is further compressed.
This price game appears to be a tug of war between price increases and decreases, but in essence, it is a phase of oscillation under the restructuring of the AI industry chain. The price increase is just a result, but the cause and effect behind it are far more complex than the price curve. Add the author of this article on WeChat EATINGNTAE to delve into the game and opportunities in the storage market.
AI drives storage prices to rise, with DDR4, which is about to be discontinued, experiencing the most significant price increase
The industry generally believes that the current rise in storage is triggered by the expansion of AI computing power.
Compared to traditional servers, the demand for storage in AI servers is growing exponentially – the DRAM usage of a single AI server is about 8 times that of a traditional server, and the NAND usage is about 3 times; The wafer consumption required to produce HBM of the same capacity is more than three times that of standard DRAM, which means that fewer HBM storage chips can be delivered under the same production capacity, and AI servers are the dominant force in the HBM market.
At the same time, the explosion of AI infrastructure construction is driving up storage demand globally.
The United States currently has 5462 data centers, with an investment of over $40 billion in data centers under construction, an increase of nearly 400% since 2022. Driven by AI demand, the expansion pace of the storage market is increasingly synchronized with the construction of computing infrastructure.
The global storage market is expected to exceed $200 billion by 2025, with AI computing power contributing over 40% of the incremental demand. ”Fan Fan, who works in the chip industry, told Leifeng.com.
However, compared to the rapid growth in demand, the actual adjustment pace on the supply side still lags behind.
The three major original manufacturers (Micron, Samsung, SK Hynix) have continued to promote their production line adjustment plans this year: on the one hand, reducing the production of traditional products such as DDR4 to free up production capacity to meet the growing demand for DDR5 and HBM; On the other hand, the newly built and renovated HBM production lines are still in the climbing stage, and it is difficult to form effective supply in the short term.
The result is a ‘capacity gap period’ between the old and new generations of products.
This structural shift has created a temporary supply vacuum, further exacerbating the imbalance between market supply and demand.
It is interesting to note that DDR5 or HBM did not experience the most significant increase, but rather DDR4, which is about to retire.
Taking SK Hynix DDR4 64GB 3200 memory modules as an example, the price in September was about 1500 yuan, and now the price has fluctuated between 3500-3800 yuan, with a price increase of 1.5 times. ”Storage distributor Lao Feng told Leifeng.com.
At one point, the price of DDR4 even exceeded that of DDR5: on one hand, it was due to the “scarcity premium” caused by the exit cycle, and on the other hand, it was due to the structural price increase driven by AI. The purchasing side was caught in a dual dilemma of “price increase if replaced, price increase if not replaced”. Add the author of this article on WeChat EATINGNTAE to analyze the real squeeze of this round of price increases on the survival space of small and medium-sized AI enterprises.
It is reported that Samsung and SK Hynix are considering extending the production of DDR4 until 2026. However, Chen Dong, a senior executive in the storage industry, believes that this possibility is unlikely.
The industry always pursues technological innovation, and production lines will naturally migrate to higher generations. Moreover, from a profit perspective, stock shortages are not a bad thing for original factories. High prices and good profits, expanding production will actually weaken profit margins
Moreover, from the perspective of market structure, the continuous rise in storage prices undoubtedly benefits the storage factories that control the core production capacity the most.
The National Day holiday coincides with the transition between the third and fourth quarters, “Lao Feng analyzed.” It is not ruled out that the original factory may actively control sales and shipping pace after achieving quarterly performance targets to maintain high prices and ensure profit margins
He believes that this “quantity control price” strategy is not uncommon in the storage industry, especially during periods of tight supply and demand and strong expectations of price increases, and is more likely to be adopted by upstream manufacturers.
At the same time, macro level uncertainty is also affecting market sentiment.
The upgraded “rare earth dispute” in October has become a new potential variable. Although rare earths are not directly used in the transistors or storage units of memory chips, they are indispensable materials in key equipment such as lithography machines and etching machines required for chip manufacturing. If there is a fluctuation in rare earth supply, equipment costs and delivery pace will be affected, which will be transmitted to the production plan and capacity release of wafer fabs.
However, according to the latest news, US Treasury Secretary Besson stated on October 27, 2025 that China and the United States have reached a preliminary consensus, and China plans to postpone the implementation of the latest rare earth export control measures, which were originally scheduled to take effect on December 1, for one year.
This news eased market tension in the short term, but also exposed a fact: the sensitivity of the market to geopolitical risks has been significantly amplified, and any new policy signals may trigger expected changes, thereby affecting the behavior of both buyers and sellers.
In the context of sustained strong demand for AI and incomplete release of original factory production capacity, the market has entered a typical “seller cycle”. Before entering the next round of price negotiations, manufacturers also have to recalculate – to what extent will the increased storage prices affect the cost structure of AI servers?
A price increase of tens of thousands of yuan, AI server customers will receive full payment according to orders
The configuration is locked, and we don’t have any room for adjustment. The upstream only dared to raise prices recklessly because they saw that this was a necessity
Zhang Zhan, from an AI server manufacturer in southern China, described the current predicament of Leifeng.com in this way.
Unlike general servers, the hardware configuration of AI servers is no longer a variable that can be flexibly chosen, but a rigid requirement directly determined by algorithms.
Taking an AI model with 175 billion parameters as an example, simply loading all parameters into the GPU for inference requires hundreds of GB of high-speed memory (HBM); If the memory capacity does not meet the standard, the model cannot run at all.
Similarly, in the process of AI training, massive amounts of data need to flow at extremely high speeds between GPU, memory, and storage. If the bandwidth of HBM is insufficient or the IOPS (read/write per second) of SSD does not meet the standard, the training cycle may be extended from a few days to several weeks or even months, which is almost unacceptable for commercial deployment.
In other words, reducing storage configuration is equivalent to weakening the effective computing power output of the system – not only will efficiency significantly decrease, but the model may also fail to run.
When the price of storage chips continues to rise, the dilemma faced by AI server manufacturers is not “whether to buy”, but “how to buy”. The hardware configuration of AI servers cannot be compromised, and costs can only be forced to rise.
Taking an AI training server equipped with 8 NVIDIA A800 GPUs as an example (2023 data), the overall cost is about 900000 yuan, of which the GPU part accounts for 84%, which is the absolute cost core. The storage part (including memory, SSD, and RAID cards) costs about 95000 yuan, accounting for about 10%.
According to DIGITIMES, Samsung Electronics recently offered a contract price for the fourth quarter, with enterprise grade solid-state drives (SSDs) increasing by 15-35%, and server storage RDIMM prices rising by 40% -50%.
Although A800 has become a product of the past, the memory used by the former is not exactly the same as the increased memory used by the latter. However, based on the above data, it can still be estimated that this storage price increase is expected to lead to an overall cost increase of about 3% to 5%, and the actual cost increase will vary depending on the specific configuration.
This is already a clear pressure for OEM manufacturers. ”Chen Dong, a senior executive in the storage industry, pointed out that “OEM manufacturers have very low profit margins and cannot digest upstream price increases alone. They can only pass on costs to customers. ”
Zhang Zhan, an AI server manufacturer in southern China, said that this price increase is also reflected intuitively in terminal products: “Taking the 2 million yuan level AI server we are selling as an example, the price difference between Dell and AMD heads can be tens of thousands, so the fluctuation range of tens of thousands of storage chips is basically within the bargaining range of the whole machine. ”
Although cost pressures are gradually transmitted along the industry chain, in the AI server market, price increases have not changed the decision-making logic of buyers.
Chen Dong added, “AI investment emphasizes timeliness. Customers cannot wait for the price to be suitable before purchasing, and the iteration speed of AI is too fast. People would rather buy at a high price than accept deployment delays caused by price reasons. ”
In other words, price increases will compress profits but not weaken demand.
However, frequent price fluctuations have indeed disrupted the pricing mechanism of the industry.
Chen Dong pointed out, “In the past, storage contract prices were usually signed quarterly or annually, providing customers with relatively stable price space; now, contract cycles have been compressed to monthly or even batch quotes, and spot market fluctuations are more frequent, with bargaining power clearly concentrated upstream
This uncertainty is also transmitted to the whole machine quotation end. Li Li from a cloud computing manufacturer revealed, “The validity period of server quotations is very short now. Previously, it could last for one month, but now it is often only seven days or even shorter, just to cope with fluctuations in accessories
But he further pointed out that price is not a purchasing barrier, shortage is the biggest risk. Even if the price increases, as long as we can get the goods, everyone is willing to pay this premium
Since October, the three major DRAM manufacturers have successively suspended contract quotations for the DDR5 series, leading to serious supply disruptions in the DRAM supply chain. Downstream suppliers can only passively turn to the spot market, further reducing the certainty of supply. Add the author of this article on WeChat EATINGNTAE to explore the balance between cost control and supply chain security.
In this pattern, price fluctuations have become the norm, which means that the increase in storage prices only brings about the compression of profit margins and the lengthening of delivery pace, rather than the cooling of demand. The AI server market is still expanding, and manufacturers’ attitude of “accepting orders at high prices” has also supported the upstream upward trend to a certain extent.
Two key signals of a turning point in the storage market: capital and production capacity
The industry generally expects that this storage rally will continue until at least the end of 2026.
According to SK Hynix, it plans to further expand production capacity next year, and all orders for DRAM and NAND products in 2026 have been fully booked. In the short term, the space for price decline is still limited.
However, the market is not without potential turning points.
Chen Dong, a senior executive in the storage industry, told Leifeng.com that to determine when the storage market will reach a turning point, two key signals can be focused on – capital heat and capacity release.
Firstly, we need to see if North American AI capital investment is slowing down; Secondly, it depends on the supply situation of the original factory’s production capacity. The current market demand is still determined and strong, while original factories are generally in a state of production reduction. Only when manufacturers start to expand production on a large scale or when new process production lines are in place, and the demand side does not grow synchronously, can prices reverse. ”
In other words, the AI wave has ignited a “price cycle” for storage, but it is not just inventory or market sentiment that determines when it will cool down, but also the pace of capital and technology.
Despite the recent price competition initiated by Samsung to seize the market in the HBM field, many industry insiders believe that HBM belongs to a specific high-end market, and its price war has a very limited impact on DRAM and NAND.
In other words, the local competition of HBM has not changed the pattern of the entire storage market being dominated by the word ‘rise’.
And this AI driven “storage price surge” is not only a shock to the supply chain, but also the real cost of the computing age——
When algorithms determine hardware, hardware is also redefining the price of AI.
The author of this article will continue to pay attention to the cost and supply impact of the storage price increase on downstream products. Now, the focus is on studying the transmission mechanism of end-to-end products. Friends in related fields are welcome to add WeChat EATINGNTAE to exchange market observations and response strategies.
暂无评论内容