High-Capacity SSDs Will Enable AI Workloads But Also Drive HDD Demand

At the recent FMS conference in Santa Clara, almost all of the SSD companies were introducing high-capacity SSDs, many over 200TB, with promises for large form factor SSD with 1PB capacities in the future. These SSDs leverage higher logical density, four bit per cell, or QLC flash memory and lots of chips to achieve these capacities.

The SanDisk keynote differentiated a couple of different uses for SSDs to support AI workloads. One type are fast eSSDs to support high bandwidth DRAM memory, HBM. The other type are high-capacity storage eSSDs for a higher performance data lake than HDDs can offer. These two types of SSD are shown below.

The slide below shows Sandisk’s high capacity eSSD. It is a QLC BiCS8 NAND flash U.2 and EDSFF form factor SSDn that is expected to have capacities up to 256TB by 2026.

The Sandisk Keynote showed a path to a 512TB version by 2027 and 1PB product sometime in the future. The composite image below shows the Kioxia, Micron and Samsung announcements of their high capacity QLC SSDs. All of the SSD companies are exploring product for near storage AI applications. Many of the higher capacity products are using the E3.S form factor, which can hold more NAND flash chips to enable higher capacities.

The Silicon Motion keynote gave an illustration of a traditional representation of the memory and storage hierarchy showing trends for NAND flash supporting GPUs directly, like the HBF announcements by SK hynix and Sandisk at the FMS. It also shows an ultra-high-capacity SSD layer to support warm storage for AI applications. Silicon Motion supplies controllers for SSDs.

Higher capacity SSDs can take less rack space than HDDs and they offer higher performance than HDDs offer. This can be an advantage for AI training and inference with RAG, but flash memory is currently about 6X more expensive per storage capacity than HDDs and is expected to remain so, for some time into the future, as shown in the image below from the WDC investor day last February. Seagate shows similar trends. For instance, by 2026 44TB HDDs should be in production, a 38% increase from the largest HDDs available today. This is because the expected storage capacity growth in HDDs has accelerated with the introduction of HAMR HDDs to roughly match the growth in SSD capacities.

As a consequence, we consider these larger SSDs will be used for data lakes directly feeding the memory attached to GPUs for AI workflows. However, HDDs will continue to provide lower cost storage for longer term data retention and so these higher capacity SSDs will result in greater growth of HDDs as well.

Coughlin Associates has updated its projections for storage capacity shipped for HDDs, SSDs and magnetic tape, shown below. This new projection increases our expectations for growth of SSD storage from prior versions out to 2030 with some reduction in HDD capacity shipments as a consequence.

The Coughlin Associates projection for HDD storage capacity prices out to 2030 is shown below.

If we assume that SSDs remain at 6X the cost per storage capacity by 2030 and the HDD price per GB of $0.0051 in 2026, the NAND flash price would be about $0.031 per GB. With the projections for shipping capacity of SSDs and HDDs of about 3.0ZB and 10.7ZB, projected revenue for SSDs and HDDs in 2030 is $93B and $55B respectively. The rising storage boat, driven by AI, is expected to result in significant revenue growth for HDDs as well as SSDs.

FMS 2025 shows growth of high-capacity SSDs, up to 1PB as well as high-capacity HDDs to support the growth of AI workflows.