SSD procurement capacity for AI workloads will exceed 45 EB in 2024

Demand for AI systems and related servers has led to an increase in orders for enterprise-class SSDs in the last two quarters. According to TrendForce, manufacturers of solid-state drive components are setting up production processes in preparation for the mass production of next-generation drives that will hit the market in 2025.

An increase in orders for corporate SSDs from users of AI servers led to an increase in contract prices for this category of goods by more than 80% from the fourth quarter of 2023 to the third quarter of 2024. At the same time, SSDs play a key role in the development of AI, since only they are suitable for effective operation with models. In addition to actually storing model data, they are also needed to create checkpoints during training, so that in case of failures, you can quickly “roll back” the model and resume training.

Due to high write and read speeds, as well as increased reliability compared to HDDs, TLC drives with a capacity of 4–8 TB are usually chosen for training models. The effectiveness of RAGs and large language models (LLMs), especially for generating media content, depends on both the capacity and performance of drives, so TLC/QLC drives with a capacity of 16 TB or more are more preferable for inference.

Image source: TrendForce

Since the second quarter of 2024, the demand for SSDs for AI servers with a capacity of more than 16 TB has increased significantly. With the increasing availability of NVIDIA H100/H20/H200 accelerators, customers began to increase demand for 4 and 8 TB TLC SSDs. The agency believes that the purchasing capacity of SSDs in 2024 will exceed 45 EB, and in the next few years the demand for server SSDs will increase by more than 60% annually. In particular, SSDs for AI workloads will potentially consume up to 9% of all NAND memory in 2025, while in 2024 this figure is expected to be 5%.