"There is no AI without data." Dinesh Bahal, vice president of Micron, recently pointed out at the "MemoryS 2025 Summit" that storage technology has become the backbone of AI (artificial intelligence) infrastructure.

According to CFM flash memory market data, the global memory market size will reach a record high of $167 billion in 2024, including NAND Flash (data flash memory chips, a type of non-volatile memory) market size of $69.6 billion and DRAM (dynamic random access memory) market size of $97.3 billion. In terms of capacity, NAND Flash and DRAM bit capacity demand will increase by 12% and 15%, respectively, in 2025 compared to 2024.

From the perspective of downstream applications, the server market has become the core driving force for the development of storage demand.

Recently, SanDisk believes that the storage market is about to enter a state where demand is outstripping supply. According to the company's latest price increase letter, SanDisk will start implementing price increases on April 1, which will increase by more than 10%. In addition, a number of manufacturers have also issued price increase notices. This is seen as a key signal of a reversal in the industry cycle. So, will the storage market usher in an overall recovery?

Image source: Visual China-VCG41N2155090857

There is a price increase signal in the storage market

In recent years, the storage industry has been in a period of deep adjustment. From 2022 to 2023, memory chips will enter a downward cycle due to the decline in shipments of mobile phones and laptops. According to a Gartner report, the global memory market size will decline by 37% in 2023, making it the segment with the largest decline in the semiconductor market. At that time, the storage industry at home and abroad was under huge operating pressure, and leading companies such as Samsung, Hynix, and Micron all suffered significant losses.

Entering 2024, benefiting from factors such as AI-driven demand growth and the effect of original factory production reduction, storage prices continued to rise at the beginning of last year, but in the second half of the year, the recovery of consumption in the terminal market was less than expected, and storage prices began to fall, showing an obvious trend of high and low. However, judging from the 2024 performance or performance forecast released by a number of A-share storage industry chain companies recently, most companies achieved good performance last year.

Among them, Longsys (301308. SZ, share price of 107.09 yuan, market value of 44.547 billion yuan) performance forecast shows that in 2024, the company is expected to have an operating income of 16.5 billion yuan ~ 18 billion yuan, a year-on-year increase of 63% ~ 78%, and a year-on-year net profit turnaround; BIWIN Storage(688525. SH, share price 79.49 yuan, market value 34.279 billion yuan) performance express report shows that in 2024, the company will achieve operating income of 6.704 billion yuan, a year-on-year increase of 86.71%; The net profit attributable to the parent company was 176 million yuan. As an important partner of leading manufacturers such as Longsys and BIWIN Storage, Lianru Technology (688449. SH, share price 53.47 yuan, market value 24.596 billion yuan) released a performance express report showing that in 2024, the company will achieve a total operating income of 1.174 billion yuan, a year-on-year increase of 13.55%; The net profit attributable to the parent company was 118 million yuan, a year-on-year increase of 126.52%.

The performance of major international storage manufacturers is also accelerating. According to the data released by Micron for the first quarter of fiscal year 2025, the company achieved revenue of $8.71 billion, and the net profit attributable to the company's shareholders was $1.87 billion, a year-on-year turnaround.

The results confirm the recovery trend. Entering 2025, storage is starting to show signs of price increases. A week ago, NAND SanDisk sent a price increase letter to customers stating that SanDisk would increase prices by more than 10% starting April 1, an initiative that applies to all channel-oriented and consumer products.

Micron also said that it would raise the price of goods for channel dealers. In terms of the domestic market, the channel has recently reported that the retail brand "Zhizhi" of YMTC will also raise the delivery price for the channel from April, which may exceed 10%.

"There is already a clear signal in the spot market to stop falling." Tai Wei, general manager of CFM flash memory market, said. He expects that in the second quarter of this year, the prices of some NAND products will be the first to stabilize, and the third quarter will have the opportunity to usher in an overall recovery.

Huang Qiang, Vice President and General Manager of the Embedded Storage Division of Longsys, is also confident in the recovery of the storage industry this year: "The vigorous development of AI products on the consumer side and the prosperity of software applications will be a continuous process in the next few years, and at the same time, embedded storage products will enter a long-term prosperity. In this positive long cycle, short-term fluctuations in supply and demand are normal. In terms of the general direction, we are more optimistic about the growth of the relevant markets. ”

Be optimistic about traditional consumer apps

From the perspective of applications, the two major markets of mobile phones and PCs (personal computers) are well-known storage application markets, and the AI era has also put forward new storage requirements for these application fields.

For example, the rapid growth of the parameter size of current large language models (LLMs), such as GPT-3 and GPT-4, has exceeded the scope of what traditional storage systems can easily handle. The number of parameters of these models has shifted from billions to hundreds of billions or even trillions, resulting in a significant increase in the bandwidth and response time requirements of the storage system. As a result, the storage capacity of high-end smartphones is constantly increasing.

"Since last year, we have seen more and more large language models being used in smartphones, the largest large language model last year was about 1 million parameters, and this year we can see 3 million parameters [of the model] being applied in smartphones," said Oh Wenxu, executive vice president of Samsung Electronics' software development team. ”

In order to meet the growing demand for device-side AI and realize the device-side operation of large language models, some vendors are improving the speed of UFS (Universal Flash Storage) interfaces. Among them, Samsung launched the UFS4.1 solution, which supports smartphones to run million-level parameter AI models. In terms of the domestic market, both YMTC and Longsys released UFS 4.1 products at the summit.

In addition to mobile phones, the replacement cycle of AI PCs (artificial intelligence personal computers) is also bringing new impetus to the development of the storage industry. According to Canalys, in 2024, global AI PCs will account for 17% of total PC shipments. "As the end of support for Windows 10 approaches, it is becoming a major factor driving replacement, and many channel partners expect that the wave of replacements for most customers will explode in the second half of 2025, and the market penetration rate of AI PCs will continue to increase in 2025." Canalys analyst Kieren Jessop said. Micron predicts that "by 2025, 43% of PCs will have AI capabilities, and by 2028, this proportion will rise to 64%, and future AI PCs will require 80% more memory capacity than current PCs."

With the growing demand for large-capacity storage and high-performance storage in the downstream market, the adoption rate of SSDs (solid-state drives) in laptops and desktops will further increase. Upstream enterprises such as Lianru Technology have begun to benefit. Although the company's main business of data storage master control chip is concentrated in the field of solid state drive master control chip, it is currently extending to the field of embedded storage master control chip.

Jin Ye, deputy general manager of the storage division of Lianru Technology, introduced that the company's products have entered the front-end market of notebook computers in 2024. "It's a very stable market with high added value. At present, the company has completed the technical iteration of storage products from SATA (a serial hardware driver interface based on industry standards) interface to PCIe (a high-speed serial computer expansion bus standard) 3.0/4.0/5.0 standard; PCIe 3.0/4.0 products have been mass-produced, and PCIe 5.0 products are also in the introduction, testing, and mass production stages. This kind of product has extremely strict requirements for high-speed signal processing, power consumption management, stability, etc., with high technical threshold and very large investment. Jin Ye said.

In addition, there are some manufacturers who are actively embracing new consumer categories such as AI glasses. "AI wearables, including AI glasses, can be said to be the real first year this year, with rapid iteration of functions and experiences, and will continue to develop in the next few years. This will require storage to be smaller, thinner, lighter, lower power consumption, and faster. At present, we have 3/4x/5x product lines in this field, such as small-form eMMC (embedded multimedia memory) and ePOP (small-form or highly integrated embedded memory developed for size-constrained application scenarios), and some practical cases have been implemented. Huang Qiang said.

Demand for enterprise-class storage continues to be strong

At present, in order to improve the bandwidth in order to better exert computing power, mainstream AI training chips have adopted HBM (high-bandwidth memory, mainly to meet the needs of GPU chips) and computing power chips to package, so the growth of AI servers is obvious. Micron has calculated that the DRAM and NAND requirements of each AI server are 8 times and 3 times that of ordinary servers.

In addition, from ChatGPT to the Wensheng video model Sora to DeepSeek, which is popular all over the world, as well as the AI large models built by various companies, all need to use AI servers. This trend not only drives the demand for HBM, DDR5, SSD and other storage technologies, but also accelerates the pace of innovation in storage technology, leading them to lower latency, higher transmission speeds, lower power consumption, and lower costs.

For the first time in the company's history, Micron's data center accounted for more than 50% of its revenue. Micron President and CEO Sanjay Mehrotra said on the earnings call, "We are ahead of the curve in AI-driven storage demand, with data center demand well above year-over-year levels, providing a strong boost to the company. HBM will continue its strong growth momentum in the coming years, with a potential market size of more than $100 billion by 2030, which is a transformational trend, and we are confident in HBM's leadership in the industry and our roadmap for the next few years." ”

However, this feast is not available to all players. At present, the HBM market is dominated by Korean and American manufacturers (SK hynix, Samsung, Micron), and Chinese manufacturers are still in the catch-up stage. However, the AI wave has also spawned more opportunities, and some A-share manufacturers have also mentioned their own plans.

Among them, in recent years, in addition to making efforts in the high-end market, Longsys is also seeking opportunities for self-developed chips. In addition, the company's enterprise-class storage has also ushered in a period of growth.

With the popularization of AI technology, the enterprise storage market is undergoing unprecedented changes, and the doubling of model parameters and computing power requirements has also brought some challenges to storage. Yan Shuyin, Vice President and General Manager of Enterprise Storage Business Unit of Longsys, said: "The first is the performance bottleneck, AI inference requires higher storage bandwidth, lower latency, and higher energy efficiency ratio, which puts forward higher requirements for the design and manufacturing of storage products. The second is the optimization of the storage architecture, and how to achieve more efficient data processing capabilities in limited space and power consumption is the core problem we need to solve. In addition, AI inference scenarios require diversified storage requirements, and enterprises need to have rapid response and customization capabilities, which is also a major test for supply chain management and technology adaptation capabilities. ”

Disclaimer: The content and data in this article are for reference only and do not constitute investment advice. Do so at your own risk.

Cover image source: Visual China-VCG41N2155090857