As of Q2 2024, the storage industry is still in an upward cycle, and the entire industry chain is full of vitality. Under the influence of multiple factors, how will the price of storage chips evolve in the second half of the year, and where will the future of storage technology point to?

In the past three quarters, the performance of the storage industry has lived up to expectations. Since Q3 last year, the decline in storage chip prices has narrowed, and the contract prices of some storage products have risen in Q4. Storage products have fully recovered in Q1 2024. As of Q2 2024, the storage industry is still in an upward cycle, and the entire industry chain is full of vitality.

Storage chips are very cyclical, and technological innovation and product iteration, capacity expansion and contraction, and changes in downstream application demand will all lead to chip price fluctuations. In order to seize market opportunities, storage manufacturers need to make targeted layouts. Under the influence of multiple factors, how will the price of storage chips evolve, and where will the future of storage technology point to?

01 Price trend of memory chips in the second half of 2024

Due to the continuous decline in memory chip prices, several major memory manufacturers suffered collective losses in 2023, with losses estimated to reach a record high of US$5 billion. According to a Gartner report, the global memory market size will drop by 37% in 2023, becoming the largest decline in the semiconductor market.

When there is an oversupply in the memory market and inventory backlogs in the upstream, memory manufacturers can regulate inventory levels by reducing prices and production. Production reduction and price reduction are a combination of punches. Whenever the sales price is close to the production cost, the manufacturer will increase production reduction until the price of memory chips rises.

Since Q4 2022, major storage manufacturers such as Kioxia, Micron Technology, SK Hynix, and Western Digital have successively started to reduce production and adjust supply. 

Then, Samsung also joined the ranks of production cuts in the first half of 2023. By September last year, its NAND Flash production was reduced to 50% of the total capacity, which was mainly concentrated in products below 128-layer stacking.

With the joint efforts of storage manufacturers, the contract prices of NAND Flash (flash memory) and DRAM (memory) stopped falling in Q3 and Q4 last year, respectively. By Q1 this year, all storage products have recovered, with DRAM contract prices rising by 20% and NAND rising by 23%-28%. Market research agency TrendForce previously predicted that DRAM would rise slightly by 3%-8% in Q2 this year and NAND would rise by 13%-18%, but revised the Q2 data in early May, with DRAM expected to rise by 13%-18% and NAND expected to rise by 15%-20%.

The agency also pointed out that the "Hualien Earthquake" on April 3 this year was a turning point. 

Before the earthquake, the upward momentum of DRAM and NAND was slightly insufficient, and their spot prices and trading volumes had continued to decline. But after the earthquake, PC manufacturers began to accept the sharp increase in DRAM and NAND contract prices. Until the end of April this year, a new round of contract price negotiations were completed one after another, and their price increases exceeded expectations.

What will be the price trend of memory chips in the second half of 2024? Because flash memory manufacturers rapidly reduced production in the year before last and last year, Duan Xiting, senior vice president of the CAS business group of Silicon Motion Technology, predicted in an interview with "International Electronic Business" in late October last year that flash memory prices may be very high in the second half of 2024. However, he never thought that this trend would begin to emerge in the first half of this year. In late March this year, he added to this prediction that the growth of flash memory will be slightly flat in the second half of this year, and the overall growth may continue until Q1-Q2 next year. There are many variables that affect memory, and it is expected that the growth may slow down in the second half of this year, mainly depending on whether the speed of AI moving to the terminal is fast enough. According to his point of view, high-performance SSDs can reduce the amount of memory used, but this will not happen immediately in the second half of this year, so the demand for memory is still relatively large, and the price increase trend of memory depends entirely on whether AI has achieved great success on mobile phones and PCs.

02 Storage demand "rejuvenates" and drives a new cycle

In addition to the measures to adjust inventory on the supply side, the "rejuvenation" on the demand side is also driving the arrival of a new cycle. I still remember that the demand for online learning/office in the early days of the COVID-19 pandemic boosted the sales of mobile Internet devices, and now these devices have gradually entered the replacement cycle.

In Q4 2023, global PC and smartphone sales ended the downward trend. Canalys data shows that in Q4 2023, PC shipments increased by 3% year-on-year, and smartphone shipments increased by 8% year-on-year; in Q1 2024, total shipments of desktops and laptops increased by 3.2%, and smartphone shipments increased by 11% year-on-year. The increase in sales of smartphones and personal computers has driven the demand for storage to a certain extent.

The popularity of HBM (high bandwidth memory) is also a positive factor. Because HBM supports high-speed data transmission between GPU and CPU, its application scenarios are concentrated in the GPU video memory of high-performance servers, and a small part is used in CPU memory chips. In the first half of last year, HBM soared 500% on the back of GPU. However, due to the high technical threshold of HBM, the shipment volume is not large so far. From the perspective of HBM competition, SK Hynix and Samsung have obvious advantages - SK Hynix accounts for about 50% of the market share, Samsung accounts for about 40% of the market share, and Micron accounts for less than 10%.

More than a year after AI technology was implemented in the data center field, it began to sink into edge and terminal devices. Qualcomm released the third-generation Snapdragon 8 platform in October last year, which supports models with 10 billion parameters on the terminal side. MediaTek released the Dimensity 9300 in November last year, which supports AI large models with 1 billion to 33 billion parameters. These two chips are aimed at the end-side reasoning of the 7 billion parameter large language model, and both can achieve 20 tokens per second (the smallest unit of text processing in the natural language model), which can realize AI instant simultaneous translation, AI note organization, AI smart calls, image AIGC elimination, flash cutout and other functions.

Today, several smartphone manufacturers have released AI phones, including the vivo X100 series, OPPO Find X7 series, Honor Magic 6 series, Samsung Galaxy S24 series, Xiaomi 14 Ultra series, Meizu 21 PRO series, OnePlus Ace 3V series, Realme 12 Pro series, Huawei Pura70 series, etc. These AI phones have higher requirements for storage to meet the needs of high data transfer rates and low latency.

Tai Wei, general manager of China's flash memory market, pointed out that the average capacity of mobile phones in 2024 will exceed 200GB, and memory will also evolve towards higher-performance LPDDR5x. This year, the average capacity of mobile phone DRAM is above 7GB. Observing the storage capacity of existing AI mobile phones in the market, the memory of the vivo X90 series is 8GB and 12GB, and the flash memory is up to 512GB. The memory of the X100 series is increased to 12GB and 16GB, and the flash memory is up to 1TB; the memory of the OPPO Find X7 series and the OPPO Find X6 series are both 12GB and 16GB, but the flash memory of the former is up to 1TB, and the flash memory of the latter is up to 512GB; the memory of the Samsung Galaxy S23 series is 8GB, and the memory of the Samsung Galaxy S24 series is 8GB and 12GB. Overall, the storage capacity of AI mobile phones is increasing, whether it is memory or flash memory.

AI is also a new engine in the personal PC market. AI PCs are embedded with AI chips and support localized AI models. They require faster data transmission speeds, larger storage capacity and bandwidth. 32G LPDDR5x and 1TB PCIe 4.0 SSDs will usher in better applications.

At present, manufacturers such as Intel, Microsoft, Huawei, Lenovo, Dell, and Asus have released AI PC products. However, analysts in the AI industry believe that the AI PC products of some manufacturers on the market have gimmick attributes, because the AI function of PC is still in its early stages, "a hot and disruptive application is needed to greatly promote the sales of AI PCs."

The above AI-related application scenarios can all drive more storage demand, and they are the focus of storage manufacturers. For example, considering the demand for large language models in chips, Samsung is developing a new product using UFS 4.0 technology, the number of channels will be increased from 2 to 4, the sequential read speed will reach 8GB/s, and the new product is expected to be mass-produced in 2025. In addition, JEDEC (Solid State Technology Association) will finalize the LPDDR6 memory specifications in Q3 this year, which will improve the performance of low-power devices. Samsung and SK Hynix are expected to be the first to mass-produce LPDDR6 products.

03 Future Development Direction of Storage Technology

According to the forecast of China's flash memory market, the storage industry will benefit from the recovery of the industry, the promotion of advanced technology and emerging market applications, and the storage market size is expected to increase by at least 42% in 2024. In terms of total storage output, NAND Flash will exceed 800 billion GB equivalent, an increase of 20% over last year; DRAM will reach 237 billion GB equivalent, an expected increase of 15%.

As the storage industry ushered in a turning point in Q3 last year, the profit margins of storage manufacturers have gradually improved, and SK Hynix has already returned to profitability in Q4 last year. From the quarterly reports of listed companies in Q1 2024, it can be seen that the performance of many storage companies has increased significantly, and the profit margins of enterprises have also been correspondingly improved. At the same time, storage manufacturers are also focusing on high-profit product lines and actively promoting the development of new technologies. This article focuses on two types of products: DRAM and NAND Flash.

(1) NAND upgrade direction: expanding capacity and reducing costs

NAND Flash is mainly developing in two directions: expanding capacity and reducing costs. The market's pursuit of flash memory capacity can be divided into vertical expansion, logical expansion, and horizontal expansion.

Vertical expansion is reflected in the storage manufacturer's efforts to produce products with higher stacking, and everyone is moving towards 300-layer NAND. Representative manufacturers include Samsung, Micron, Western Digital, Kioxia, Yangtze Memory, SK Hynix, etc.

The evolution from SLC, MLC, TLC, QLC to PLC is logical expansion, and the data stored in each unit is getting larger and larger so that the SSD can save more data. Generative AI has driven a surge in server storage demand, and the commercialization of QLC flash memory has accelerated in 2024. For example, Samsung's 128TB prototype SSD is based on QLC.

Horizontal expansion adds more storage capacity at each layer. This increases the density of the pores in each layer to achieve a larger unit bit capacity, which is a direction to pay attention to when exploring new technologies and new processes in the future.

The increase in the number of flash memory stacking layers also promotes hybrid bonding technology to become the mainstream architecture. Hybrid bonding is mainly used in two ways: the first is wafer to wafer, which is used for CIS and NAND; the other is die to wafer hybrid bonding, which is very meaningful for HBM. The specific operation is to optimize the distribution relationship between the peripheral logic circuit and the storage unit, such as from CNA and CUA to CBA (peripheral circuit directly bonded to the storage array). There may be multi-layer bonding scenarios in the future.

(2) The sixth generation 1c process DRAM is about to go into mass production

Since DRAM entered the 10nm process era, it has evolved from 1x, 1y, 1z, 1α to 1β (fifth generation). Samsung, SK Hynix, and Micron completed the research and development of 1β last year and are now mass-producing 1β DRAM products. In addition, Samsung's sixth-generation 1c process DRAM, which is highly anticipated, will also be mass-produced within 1 to 2 years.

The rapid growth of AI servers in 2023 will also drive the increase in demand for HBM and DDR5. In order to meet market demand, major original manufacturers are accelerating the launch of advanced DRAM products. With the launch of mainstream server CPUs that support DDR5 second-generation memory products, 5,600 MT/S rates will enter the mainstream market in the second half of this year.

AI servers have also driven the demand for 128GB/256GB memory modules, and these high-capacity modules are limited by the TSV capacity of packaging and testing plants. Fortunately, storage original manufacturers will launch 32Gb single Die this year, so that 128GB memory modules do not need to use TSV technology, which will promote 128GB products into the mainstream server market. In addition, CXL entered the practical stage and HBM3e started mass production in the first half of this year, which will also promote a major upgrade of server memory.

04 Summary

In addition to AI-related applications, automobiles are also the main application driving storage demand. With the commercialization of L3 and above autonomous vehicles, the performance and capacity requirements for automotive storage are also greater, and the storage capacity of a single vehicle will soon enter the TB era.

The AI market will also sink into automobiles, and more stringent requirements will be placed on automotive storage. Smartphones, personal computers, servers, electric vehicles. As long as it is an application scenario that AI can reach, more new forms of storage products will be seen.

View more at EASELINK