recently,Dell Technology Group CEO Michael Dell said at a public event that total memory demand in the AI accelerator field is expected to surge 625 times in 2028 compared with 2023.The supply shortage problem in the global AI memory market is difficult to alleviate in the short term, and the imbalance between supply and demand will continue to exist.
This 625-fold demand growth is formed by the superposition of two core factors. The H100 accelerator launched by NVIDIA in 2023 is equipped with 80GB of memory.By 2028, the memory capacity of a single AI accelerator will increase by 2TB, achieving a 25-fold increase.
At the same time, the deployment scale of AI accelerators in global data centers will also expand 25 times in the next five years, and the double increase will drive the overall memory demand to rise exponentially.
The current shortcomings of the AI memory supply side have emerged. The global storage industry will be at a cycle low in 2023. The three major memory original manufacturers have suspended production expansion plans due to performance losses, which directly leads to insufficient subsequent production capacity reserves.
It takes about 4 years to build a new DRAM wafer fab from planning to mass production, and leading memory companies are still cautious about expanding production at this stage. The speed of front-end process capacity improvement is limited, making it difficult to match the growth pace of AI memory demand.
In addition, the growth driver of AI memory demand continues to be strong, and most of the top 25 countries in the world's GDP are promoting the construction of sovereign AI.
Enterprises continue to increase investment in AI infrastructure to improve production efficiency, and AI-related capital expenditures by very large enterprises continue to grow.
Michael Dell said that although memory prices may rise or procurement delays may occur, the layout of AI infrastructure for enterprises and cloud service providers is imperative, and the industry's rigid demand for AI memory will further exacerbate the contradiction between supply and demand.
In addition, it is generally believed in the industry that the supply shortage of AI memory will continue until 2028, and there is no effective capacity replenishment solution in the short term that can completely solve this problem.