Consumer-grade memory prices are falling. How much longer can the storage supercycle last?

When trading stocks, just look at the Golden Qilin analysts’ research reports—authoritative, professional, timely, and comprehensive. Help you uncover high-potential themes and opportunities!

Recently, as Google’s TurboQuant algorithm continues to gain momentum, prices of storage products in the consumer-grade market have been fluctuating.

According to publicly available information, in the domestic market, the 7-day average transaction price of DDR4 on the second-hand platform Xianyu has fallen by 80 yuan compared with the previous week, to 360 yuan; the average transaction price of DDR5 has fallen by 20 yuan compared with the previous week, to 1,090 yuan. On the app DeDu, the price of 16G DDR5 memory sticks reached its peak from March 13 to 19 at about 1,674 yuan per stick, and has since dropped to around 1,399 yuan.

In the international market, over the past week, several U.S. retailers have cut prices on DDR5 memory across the board, with the maximum discount on a single kit reaching 100 dollars. Among them, price cuts for products related to Corsair have been especially pronounced. This round of price reductions is reflected on major e-commerce platforms such as Amazon U.S. and Newegg. Taking Corsair’s VENGEANCE series DDR5 memory with a 32GB capacity and a top 6,400MHz frequency as an example, its price is about 379.99 dollars, down significantly from a recent peak of 490 dollars.

Meanwhile, an SShanghai Securities News reporter noted that enterprise-grade memory prices have remained relatively steady. For example, on online trading platforms such as JD.com and Xianyu, the price of 64GB DDR5 memory has dipped slightly, but overall it is still above 12,000 yuan.

Why did consumer-grade memory prices see a sharp pullback in the short term? Why is there a price divergence between consumer-grade and enterprise-grade storage? And will the storage super cycle be sustained?

Multiple catalysts drive the price pullback

Through interviews and research, the reporter learned that technological shocks, supply-and-demand changes, and channel panic may be the main catalysts behind this round of price cuts.

First, the impact of the TurboQuant algorithm has caused market expectations to shift.

Last week, on its official website, Google demonstrated an AI memory compression technology called “TurboQuant,” claiming it can compress cache memory usage in large language model (LLM) inference to one-sixth, and achieves up to 8x performance acceleration on an NVIDIA H100 GPU. The market may interpret this as a potential decline in AI demand for memory.

Second, supply-and-demand dynamics have shifted, and the market supply of consumer-grade products has rebounded.

“Previously, the memory market experienced unprecedented supply-and-demand imbalance.” Bian Zhengrong, technical负责人 at Shanghai Mugo Network Technology Co., Ltd., a data center service provider, told the reporter of Shanghai Securities News, “From mid-2025 to early 2026, the three major international storage giants—Samsung, SK hynix, and Micron—will divert more than 80% of advanced production capacity to producing high-bandwidth memory (HBM) and high-end DDR5 to meet the explosive demand from AI servers. This directly leads to a steep shortage of consumer-grade DDR4/DDR5 memory supply, pushing prices higher along the entire curve.”

But as HBM production lines gradually stabilize, international manufacturers begin releasing some capacity into the consumer market. At the same time, domestic manufacturers such as ChangXin Memory are accelerating capacity releases, filling part of the supply gap. As a result, industry inventory levels have surged from the previous 3–4 week safety line to 6–7 weeks.

In addition, panic liquidation by channels is also a major factor driving the sharp pullback in consumer-grade storage prices.

In Bian Zhengrong’s view, as memory prices kept rising in the past, they attracted large numbers of channel dealers and speculators to stock up and drive speculation. “Now, those holding inventory are starting to pull back funds. At the time, memory sticks had become ‘black gold bars,’ with daily price increases of up to 50 yuan. But as the turning point in prices appears, inventory holders rush to sell to recoup cash, forming a vicious cycle of ‘the more it falls, the more it gets dumped, and the more it gets dumped, the more it falls.’ Even rare cases of price inversion emerged, where the dump selling price in the channel was lower than the brand’s factory ex-works price. This channel selling pressure peaked in March 2026, causing the maximum single-day decline of consumer-grade memory sticks to reach as much as 100 yuan.”

The storage super cycle may not be over yet

In the view of industry insiders, this round of short-term sharp price cuts in consumer-grade storage is not strongly representative. Growth in the storage market size is still mainly driven by enterprise-grade products, and the storage industry may still not have escaped the super cycle.

First, consumer-grade storage and enterprise-grade storage can be viewed as different products. There are significant differences between the two in technical complexity, capacity allocation, and profit margins.

As Bian Zhengrong explained, HBM uses a vertical stacking design, usually with 12–16 layers, and connects them through TSV (through-silicon vias). The manufacturing process complexity is far higher than that of traditional DRAM. SK hynix’s HBM3E yield has already reached 80%, but Samsung’s HBM3 yield is only 20%, and the application of hybrid bonding technology still faces challenges.

Plus, the gross margin of HBM products can reach more than 70%, far higher than the 20% of consumer-grade memory. This directly drives manufacturers’ capacity allocation strategies, creating a market landscape of “premium memory feeding on it, while consumer-grade memory rides a roller coaster.”

Second, enterprise-grade storage and consumer-grade storage have different application scenarios and demand elasticity.

A staff member at a storage distribution company told the reporter that demand for enterprise-grade storage is driven by infrastructure such as AI servers and data centers. The demand is rigid and not substitutable. The amount of DRAM consumed by a single AI server is 8–10 times that of a typical server, and HBM is mainly used in core stages such as training, so it cannot be fully replaced by compression technologies.

By contrast, demand for consumer-grade storage is driven by consumer electronics such as PCs and mobile phones, with very high elasticity and strong substitutability. When prices are too high, consumers can delay purchases or choose lower-spec models. Manufacturers can also control costs by reducing memory configurations (for example, from 16GB down to 8GB).

Finally, the market does not need to worry too much about the technological shock brought by TurboQuant.

Compressing KV Cache and optimizing long context is not a completely new technical approach. As early as April 2025, Google had already publicly released papers related to TurboQuant.

On similar technical lines, domestic efforts have already been in place. For example, Kimi Linear from Moonshot AI (MaaS?)—in handling long-context tasks—uses KV Cache in a way that can reduce by up to 75% compared with traditional full-attention models. DeepSeek V2’s proposed MLA method can also optimize KV Cache.

At the same time, the scope of TurboQuant’s current validation is relatively limited. Tian Feng, Dean of the Fast Thinking, Slow Thinking Research Institute and a special commentator, said that the technology has been validated only on open-source models such as Gemma and Mistral so far, and the adaptation results for Google’s core models such as Gemini have not yet been made public. The general applicability of the technology still needs observation.

“The Jevons Paradox” also shows that improvements in technical efficiency often reduce usage costs, thereby stimulating a larger overall demand. Just as lower costs for AI training and inference previously drove an explosive increase in demand for compute power.

From a supply-chain perspective, manufacturers’ production capacity is fully utilized in the short term. With server memory demand continuing to grow, in 2026, server DRAM demand is expected to rise 39% and HBM demand is expected to grow 58% year over year. According to publicly available information, Samsung plans to increase HBM4 capacity by 50% in 2026; SK hynix will reduce the share of DDR4 capacity to 20%; and Micron has already sold out all its HBM for 2026.

Author: Zheng Weihhan

		Sina Statement: This message is reposted from Sina’s partner media. Sina.com掲載 this article for the purpose of conveying more information, and does not mean that Sina agrees with its viewpoints or verifies the descriptions. The article content is for reference only and does not constitute investment advice. Investors act on this at their own risk.

A vast amount of information and precise analysis—right on the Sina Finance app

责任编辑：Lingchen