Storage prices plummet: a correction signal or short-term fluctuation?

Recently, the topic “Memory stick prices see a steep drop” has climbed to the top of the trending search list. Reports say that mainstream 16GB DDR5 (the latest product based on the DRAM industry’s mature process technology) memory sticks have fallen from a high of 1,000 yuan in December 2025 to around 700 yuan. The 32GB kits have even shrunk by 27% within a month, dropping from 3,000 yuan to 2,200 yuan.

From the channel side, some loosening in pricing is reflected. On the secondhand e-commerce platform Xianyu, the成交均价 (transaction average price) of DDR4 over the past 7 days fell by 80 yuan compared with the previous week, to 360 yuan; the transaction average price of DDR5 also declined slightly by 20 yuan to 1,090 yuan. For mainstream e-commerce platforms, on JD.com, Kingston 16GB DDR4 memory is priced at 849 yuan and up, while ADATA 16GB DDR5 is 1,400 yuan; US-based Corsair’s 16GB DDR5 is 1,429 yuan. In mid-March this year, 16GB DDR5 memory sticks once surged to 1,674 yuan. “The market started dropping in price last week, but actually it didn’t fall that much,” a seller in North Huaqiang, Shenzhen (000062), said.

Although prices have pulled back, based on historical prices, the current level is still high. According to data from TrendForce, from the third quarter of 2025 to now, the spot prices of DRAM (dynamic random-access memory) and NAND (flash memory) have cumulatively risen by more than 300%. In September 2025, the price of 16GB DDR5 memory sticks was only around 300 yuan. In early November 2025, the market price of 16GB DDR4 memory sticks in North Huaqiang was around 400 yuan; at that time, many merchants, because prices were rising too fast, were “afraid of height” regarding the outlook and generally didn’t dare to stock up in large quantities.

“Still too early for normal prices,” many users commented on social platforms.

In the view of multiple industry insiders, this round of price fluctuations is more of a phased adjustment. In fact, the current supply-and-demand relationship for memory is still extremely tight. At the MemoryS2026 summit held recently, CFM Flash Market General Manager Tai Wei said that in 2026, no mainstream AI (artificial intelligence) storage product globally can achieve a complete balance of supply and demand. The focus of the storage industry has already shifted from “who is cheaper” to “who can get the goods.” To sum it up in one sentence: how hot AI compute is, how severely scarce storage is.

Why did prices pull back?

After months of rising memory prices, they finally saw a decline, which has triggered heated discussion from the outside. But more attention is on this: in this round of price pullback, is it a trend reversal, or just a phased fluctuation?

TrendForce analyst Wang Yuqi told reporters that the spot market has indeed weakened recently, but contract prices for 2026’s second quarter are still expected to rise, and this uptrend is expected to continue through the end of the year. Therefore, this round of price movement is expected to be a short-term pullback.

“Right now, what’s falling is the spot market. Spot-market consumers (for high prices) have limited ability to accept them, so a pullback is very normal. We’re bullish through 2027—what we’re referring to is the contract market,” another storage-industry analyst said.

Executives of listed storage companies believe that this decline in storage prices is more a business behavior by some merchants. “From a supply-and-demand perspective, the industry hasn’t seen a clear turning point yet—at least in the second quarter this year, the tight situation of supply and demand hasn’t eased. However, with prices already at a high level, demand-side will face some suppression, and the time and number of periods of sideways movement will relatively increase.”

In fact, judging from the current supply-and-demand situation, it is still extremely tight. Tai Wei believes that the capacity expansion cycle in the storage industry lasts 18 to 24 months, and the earliest new capacity release will only happen in 2027. “However, we believe even if new capacity is added, it still can’t fully meet market demand. The problem of storage supply shortage will be difficult to resolve in the short term.”

This round of the upcycle started in the second half of 2025 and accelerated notably in the fourth quarter of 2025. A report released by market research firm Counterpoint shows that as of the first quarter of 2026, memory prices rose 80%~90% quarter over quarter, reaching an unprecedented record high. Taking server-grade memory as an example, the contract price of 64GB RDIMM (dual inline memory module with registers) has jumped from 450 dollars in the fourth quarter of 2025 to above 900 dollars in the first quarter of 2026; the institution believes the second quarter is likely to break the 1,000-dollar threshold.

During this upcycle, prices have been transmitted from the upstream to the channels, modules, and complete system ends. In some products’ spot markets, some price increases have already reached 10 times or even dozens of times; the industry’s focus has also shifted from “who is cheaper” to “who can get the goods.”

Also worth noting is that this round of price pullback is mainly concentrated in the consumer market, while enterprise-level and AI server-related storage demand remains strong. At the MemoryS2026 summit, multiple storage practitioners spoke intensively. They believe that as artificial intelligence accelerates into real-world deployment, storage demand is still being pushed to new highs.

Among them, Tai Wei pointed out that from model training, to inference, to multimodal applications, storage bandwidth and capacity requirements are all being raised. “HBM (high bandwidth memory) is shifting from a niche product to a key resource in the AI era. Large-capacity DDR5 has gone from an optional configuration to a standard configuration for AI servers. Enterprise SSDs (solid-state drives) are no longer only a carrier of capacity; they’ve become a key element for breaking performance bottlenecks in the compute architecture. This year, the growth rate of demand for server memory will exceed 40%, and its share in overall storage application will exceed 50%.”

At the same time, as AI applications shift from model training to more frequent actual use, companies’ requirements for data read speed and responsiveness have increased clearly. “High-performance storage is no longer a nice-to-have; it is a core foundation that determines the efficiency and scale of system decision-making,” said Zhang Shiwan, executive vice president of Samsung Electronics. Samsung plans to release EDSFF drives with a thickness of only 1T from 2026 to 2027. This方案 can increase the total capacity and bandwidth per single rack multiple times, maximizing improvements in space-operation efficiency.

What will happen next?

Market demand remains strong, but prices have pulled back—this phenomenon has drawn market attention to new variables. Some analysts believe that breakthroughs in the AI algorithm field recently have, to a certain extent, affected market expectations. On March 26, Google released a new compression algorithm called “TurboQuant,” which can significantly compress the key-value cache memory used when running large language models without losing model accuracy.

This technology once sparked market concerns about a decline in storage demand. However, industry views on this are relatively rational.

Wang Yuqi believes that TurboQuant compresses KV Cache (key-value cache mechanism), which reduces memory usage without affecting model accuracy. Essentially, this kind of technology improves resource utilization efficiency rather than weakening overall demand. “As inference costs fall, AI application scenarios and model scale typically expand in parallel, so in the long run it still helps drive growth in overall storage demand. There has indeed been short-term volatility in the market recently; as AI and data center construction continue to advance, the trend of storage prices still has support. Short-term fluctuations won’t change the industry’s long-term growth trend.”

Industry insiders analyzed for reporters that in the long term, this kind of technology is beneficial: it helps optimize AI storage architecture, alleviates the previous situation where supply and demand imbalance became too severe, and thus promotes healthier development for the industry.

For example, to cope with the cost pressure on downstream users caused by rising storage prices, Jiangbo Long self-developed HLC (High Level Cache, advanced cache technology). With it, users’ experience won’t be affected even while reducing DRAM usage. On the embedded side, Jiangbo Long and Unisoc jointly developed it. Based on tests using Unisoc chip platform, after pairing 4GB DDR with the HLC technology, the start-up response time for 20 Apps is 851ms (milliseconds), close to the level of normal configurations for 6GB/8GB DDR—reducing terminal DRAM capacity requirements and optimizing BOM (bill of materials) costs.

At the same time, the industry insiders emphasized that HLC is not a “short-term solution” for the storage price hike cycle. “Regardless of whether changes in storage cost are obvious, at least from a performance perspective, HLC can improve storage efficiency. It’s a design approach combining hardware and software, which helps AI’s long-term development on the edge. In the future, when the industry truly enters the inference stage, cost will definitely become a key consideration.” It is reported that the HLC technology can be applied to different scenarios such as phones, computers, and tablets, and it has already entered the product promotion stage.

Other storage vendors are also reducing reliance on high-end memory through technological innovation. For example, Kioxia (Toshiba) has launched AiSAQ (scalable vector search technology). By storing data in SSDs, it reduces the need for DRAM; meanwhile, it can also reduce latency and improve performance. For large-scale data, AI systems can process large datasets without having to purchase expensive DRAM, because the workload resides in large-capacity SSDs. “For end-to-end data ingestion, it can accelerate by 7.8 times; for search and the number of query items, it can significantly improve.”

Overall, the current storage market shows the characteristics of “price volatility coexisting with demand growth.” On the one hand, under pressure at high price levels, the spot market sees phased pullbacks; on the other hand, the medium- to long-term demand driven by AI is still being released continuously.

(Editor-in-charge: Zhang Yang HN080)

     【Disclaimer】This article only represents the author’s personal views and is not related to Hexun. The Hexun website maintains neutrality regarding the statements and judgment opinions made in the text, and provides no express or implied guarantees regarding the accuracy, reliability, or completeness of the content included. Readers are requested to use this information only for reference and bear all responsibility themselves. Email: news_center@staff.hexun.com