More than Price Increase: Storage Industry Restructuring Logic Behind AI Server's 8x DRAM Demand

Recently, the news of a 70% quarterly surge in server DRAM prices has plunged the electronic components industry into a dual emotion of "price frenzy" and "cost anxiety". Behind this wave of price hikes lies a core change: AI servers require up to eight times more DRAM (Dynamic Random Access Memory, commonly known as "memory") than regular servers. This surge in demand has disrupted the supply-demand balance in the storage market and is profoundly restructuring the global storage industry landscape.

To understand this transformation, two key points need to be clarified: the role of DRAM and the reasons for the high memory consumption of AI servers. Simply put, DRAM is like a "temporary workbench" for electronic devices - all running programs and processed data must be placed on this "workbench" to operate efficiently. A regular server only needs a few tens of GB of memory to meet the needs of office work, web browsing, etc., while an AI server training a large language model requires a single DRAM capacity of hundreds of GB or even TB, which is 8-10 times that of a regular server. More importantly, the global number of AI servers is skyrocketing from hundreds of thousands in 2023 to millions in 2025, and the total demand for DRAM is growing exponentially.

If the surge in AI servers is a "ferocious beast" on the demand side, then the capacity adjustment of storage giants is a "further blow" on the supply side, and both have jointly given rise to this supply-demand imbalance. From the demand side, AI's demand for DRAM shows a "rigid explosion" characteristic. AI performance relies on data feeding, and the more advanced the model, the higher the requirements for DRAM capacity and bandwidth. In addition to model training, the popularity of inference scenarios such as AI generation, intelligent customer service, and autonomous driving has further amplified the demand for DRAM, making it a core necessity in the AI era. From the supply side, the three giants Samsung, SK Hynix, and Micron monopolize over 95% of global DRAM production capacity. In order to seize the AI high-end storage dividend, they have shifted their production capacity from low gross margin ordinary DRAM to high value-added HBM (high bandwidth memory). HBM manufacturing consumes more wafer resources, directly reducing the supply of ordinary DRAM, and it takes 2-3 years for new wafer fabs to release production capacity, making it difficult to fill the short-term gap.

The dual effect of "surge in demand+contraction of supply" is reconstructing the supply and demand pattern of the storage market from four dimensions. One is the return of pricing power to the original storage factory, forming a seller's market. Due to their strong bargaining power downstream, Samsung and SK Hynix have not only raised prices by 70%, but also switched to quarterly signing to lock in dividends. Downstream companies are forced to compromise in order to ensure supply, and the bargaining power of original factory pricing has reached its peak in recent years. The second is the differentiation of market structure, with high-end and low-end storage experiencing two extremes of popularity. HBM has become a popular track due to the urgent need for AI, with SK Hynix having a market share of over 70%; Ordinary DRAM continues to experience shortages and price increases, and this differentiation has been transmitted to end devices, putting pressure on mid to low end consumer electronics due to chip shortages. Thirdly, domestic storage has ushered in a breakthrough window. Affected by geopolitical policies, international giants have limited production capacity in China, and Micron has withdrawn from some of its Chinese businesses to make room for domestic manufacturers. Changxin Storage has achieved mass production of HBM3 and Changjiang Storage has broken through high-end NAND technology, accelerating its entry into the global supply chain. The fourth is the redistribution of industrial chain value, with midstream sharing of dividends. The profits of storage original factories have skyrocketed, while midstream enterprises such as packaging and testing, equipment, etc. have also benefited. Companies such as Changdian Technology and Huahai Chengke have achieved profit growth through related businesses, and the profits of the industrial chain have tilted towards AI storage related links.

Looking ahead, the AI driven restructuring of the storage market will continue. TrendForce predicts that the growth rate of DRAM demand in 2026 will far exceed supply, and unit prices may rise by 58%. For the electronic components industry, this is both a challenge of rising costs and an opportunity for domestic substitution and capacity acquisition. Ultimately, this restructuring is an inevitable result of the large-scale application of AI, and the importance of storage as the "computing power base" is becoming increasingly prominent. For enterprises in the industrial chain, the ability to accurately grasp changes in supply and demand and stabilize the supply chain becomes the key to breakthrough.