Memory chips, which account for about one-third of integrated circuits, are considered a barometer of the semiconductor market. Under the dual influences of AI catalysis and the sluggish consumer electronics market, the memory market has recently seen a divergence.

In a recent interview with First Financial Daily, Duan Xiting, Senior Vice President of the Terminal and Automotive Storage Business Group at the leading NAND Flash controller chip manufacturer Silicon Motion Technology, said that due to the demand driven by AI, the demand for data centers in the first half of the year was very hot. On one hand, there is a need for HBM (High Bandwidth Memory) to quickly deliver data to processors for computation, and on the other hand, there is a need for large-capacity SSDs (Solid State Drives) to store a large amount of data.

HBM is made up of multiple DRAM chips (Dynamic Random-Access Memory) stacked together and is almost a must for high-performance AI chips. It was the first to be catalyzed by AI, while SSDs are made from NAND Flash chips (a type of non-volatile storage medium, referred to as flash memory below), and both types of chips occupy the main memory chip market. However, under the visible hot demand for data centers, the market demand is actually unevenly distributed. Duan Xiting told the reporter that the price of original factory flash memory chips continued to rise in the first half of the year, while the price of consumer electronics terminals showed a disconnect. Behind this, the demand in the consumer terminal retail market is very bleak, and it relies on the demand from data centers to support the price increase of original factory flash memory chips. In this situation, how much the flash memory chips will rise in the second half of the year still needs to be observed.

Advertisement

At the two ends of the cold and hot, on one side, DRAM manufacturers are actively expanding the production of HBM, and NAND Flash manufacturers are promoting price increases, while on the other side, the "cold wave" in consumer electronics has not completely receded. As the market demand is in a period of transition, some profound changes brought about by AI are also taking place.

The memory market is unevenly cold and hot.

In terms of the market, the demand for HBM is still on the rise. According to a recent report by SEMI (International Semiconductor Industry Association), sales from the wafer factory equipment department are expected to grow by 2.8% in 2024, reaching 98 billion US dollars, higher than previously predicted, due to a large amount of investment in DRAM and HBM in the industry.

Several major DRAM manufacturers are still competing to expand production. The SK Group recently stated that its semiconductor subsidiary SK Hynix plans to invest 103 trillion won by 2028, with about 80% of the amount to be used to invest in HBM. There are also reports that Micron is building an HBM test production line and a mass production line in the United States, planning to increase the market share of HBM to about 20% by 2025.

The reporter learned that the expansion of production has gradually eased the tight supply of HBM. "The iteration speed of HBM technology is very fast, and the second half of the year is mainly focused on the demand for HBM3e. Overall, HBM is still in a state of supply not meeting demand. However, in the Nvidia H100 chip equipped with HBM3, a trend of cycle convergence to a quarter has been observed," TrendForce analyst Wang Yuqi told the reporter.

However, apart from HBM, there is a divergence in the storage market. Due to the need to store a large amount of data in AI training and other links, the demand for flash products in data centers has also increased, but the other major application field of flash, consumer electronics, is still sluggish. According to IDC data, global PC shipments grew by 3% year-on-year to 64.9 million units in the second quarter of this year, and global smartphone shipments grew by 7.8% year-on-year to 289 million units in the first quarter, but both shipments are not as high as the same period in 2022.

The demand is reflected in the price of flash memory chips. "The price of flash memory chips rose very quickly in the first half of the year, and the products needed by data centers were in short supply. But currently, the retail market for consumer electronics terminals in the second half of the year is still not good enough. Whether the flash memory factory will continue to raise prices depends on how much momentum the data center will have to purchase storage products in the second half of the year. I think the price increase of flash memory chips in the second half of the year will converge, but it is still uncertain how much it will converge," Duan Xiting told the reporter.Looking at the spot market prices of flash memory chips, TrendForce analyst Gao Guofeng told reporters that after the flash memory manufacturers suffered significant losses last year, their primary goal for this year is to achieve or increase profits, especially as there has been no substantial expansion of production capacity this year, leading to a continuous rise in flash memory prices. However, due to poor demand in the spot market, spot prices have not risen as actively as the original factory prices.

TrendForce Consulting predicts that in the third quarter, Enterprise (corporate) SSDs will benefit from the expanded use of AI, but the demand for consumer electronics continues to be sluggish. It is forecasted that the average price increase of NAND Flash will narrow from 15% to 20% in the second quarter to 5% to 10% in the third quarter.

Although not directly producing storage particles, the business of many domestic companies is also related to storage systems, storage modules, and SSDs. After experiencing a market recovery in the first half of the year and facing the uncertain demand in the second half, the performance and stock prices of flash memory-related manufacturers have shown some reactions.

In terms of performance, Tongyou Technology and Baiwei Storage have recently announced that their net profits for the first half of this year have turned from loss to profit, with expected profits of 0.4 billion to 0.52 billion yuan and 2.8 billion to 3.3 billion yuan, respectively. Jiangbo Long also announced a profit turnaround in the first half of the year, with profits of 5.2 billion to 6.1 billion yuan. Looking at the stock prices, Tongyou Technology, Langke Technology, Jiangbo Long, and Baiwei Storage all had a wave of stock price increases at the beginning of this year, with the highest intra-year stock prices recorded in March, March, April, and early July, respectively. However, after the peak, the stock prices continued to fluctuate or decline.

AI has changed the industry competition.

Data centers and consumer electronics terminals are both facing energy issues brought about by the consumption of AI computing power. Consumer electronics terminals that use batteries and cannot use auxiliary methods such as liquid cooling like data centers have more stringent requirements for the power consumption of storage products. Duan Xiting told reporters that the "wind" of AI needs to "blow" from data centers to consumer terminal devices, facing three major challenges, including controlling terminal energy consumption, reducing large language models, and reducing costs.

The competition among storage manufacturers has shifted from faster data transfer speeds and larger storage capacities to more energy-saving. The reporter recently learned from Samsung Electronics that Samsung has completed the fastest LPDDR5X DRAM verification on MediaTek's next-generation Tianzu mobile platform, and the product highlights the adaptation of energy consumption reduction to mobile devices with AI functions. Flash memory and SSD manufacturers like Kioxia previously told reporters that while promoting large-capacity storage, they will also focus on low power consumption and energy-saving. The 2TB four-level cell (QTC) storage device that Kioxia began sampling this month also emphasizes power performance and adaptation to AI scenarios.

The requirement for energy-saving has, to a certain extent, become a competition for high-process chips for flash memory product main control chip manufacturers. Unlike storage particles that improve storage density through 3D stacking and do not emphasize advanced processes, there is also a main control chip on flash memory products that needs to rely on process progress to improve performance. The reporter learned that in response to the high-performance and low-power requirements of AI scenarios, this main control chip has also begun to move towards high-process chips.

"In the past, we mainly evaluated the quality of the main control chip by speed, but in the AI era, we look at the data transfer speed per watt. This sets a higher threshold for main control chip manufacturers. Our latest product has already used TSMC's 6nm EUV process. As we enter the 6nm, 5nm, and 3nm processes, the investment for manufacturers will be very huge, and the industry competition threshold will be higher." Duan Xiting told reporters that NVIDIA's AI chip iteration speed is basically one per year, and the iteration speed of flash memory main control chips is similar, which requires main control chip manufacturers to consider whether they have enough financial resources to continue investing in research and development.

"To meet the energy-saving and heat dissipation requirements of AI servers, more and more manufacturers are beginning to use 6~12nm processes to develop SSD main control. This not only greatly increases the threshold of development funds (it is estimated that the funds required for developing 6/7nm processes are about 20 to 30 million US dollars), but also makes the research technology more complex. These factors will widen the competitive situation among main control manufacturers." Gao Guofeng also told reporters.Whether it's servers or consumer electronics, the use of faster HBM, larger-capacity flash memory products, and advanced process flash memory controllers to adapt to AI workloads may drive up costs. Market research firm TechInsights recently reported that AI will push up the average price of chips.

Although there is uncertainty about whether flash memory chips will continue to rise in price in the second half of the year, TSMC, the foundry leader, seems more certain to raise prices, benefiting from the strong demand for advanced process chips driven by AI. In June this year, TSMC was reported to have raised prices for some advanced process chips. Industry insiders told reporters that TSMC's price increase is based on its pricing power in the field of advanced processes. In the more competitive mature process field, it is not very daring to raise prices significantly, and there is currently no sign of stopping the price increase for advanced processes.