AI Storage Demand Surges: NVIDIA CMX May Consume Over 100 Million TB of NAND Next Year, Samsung Has Started Supply

Demand for NAND flash memory in NVIDIA AI servers is entering a phase of explosive growth. Leveraging its capacity advantages, Samsung Electronics is rapidly converting this trend into a strategic opportunity.

According to a report by The Seoul Economic Daily on July 20, Samsung Electronics has initiated mass production of its 10th-generation V-NAND (V10) and officially begun supplying it to NVIDIA, while simultaneously allocating approximately 60% of its total V-NAND capacity to the production line for its 9th-generation product (V9) to meet the demand for NVIDIA's upcoming new storage device, "Context Memory Storage" (CMX). Industry estimates suggest that the NAND capacity required for CMX will surge from 35 million TB this year to over 100 million TB next year.

This wave of demand is having a profound impact on the supply landscape of the entire NAND market. NVIDIA's large-scale locking of NAND supplies means that the availability of memory chips for other enterprises and consumers will significantly narrow, putting upward pressure on market prices.

Meanwhile, Samsung Electronics Chairman Lee Jae-yong is expected to meet with NVIDIA CEO Jensen Huang in San Francisco in the near future. The two sides will engage in in-depth discussions on the supply of HBM and next-generation NAND products, as well as cooperation on building data centers within South Korea.

CMX Drives Surge in NAND Demand, NVIDIA Locks in Samsung Capacity

The emergence of NVIDIA's CMX stems from the rapid expansion of "Key-Value Cache" (KV Cache) data volumes during AI inference. Previously, GPU servers could process KV cache data internally, but as data scales continue to expand, independent external storage devices equipped with thousands of terabytes of solid-state drives (SSDs) have become a necessary configuration for AI servers.

According to wccftech, a single NVIDIA CMX unit is equipped with 576 SSDs, offering a total storage capacity of 9,600 TB. Industry estimates indicate that the NAND capacity required for CMX will be approximately 35 million TB this year and will exceed 100 million TB next year. Independent semiconductor industry analyst @jukan05 pointed out in a post on platform X that a scale of 100 million TB is roughly equivalent to adding a demand source the size of Apple Inc. to the NAND market, and its impact on the supply side should not be underestimated.

Samsung Electronics currently has a monthly V-NAND wafer production capacity exceeding 100,000 wafers. It is reported that NVIDIA requested an expansion of NAND capacity from Samsung just before the shipment of CMX, and Samsung is actively responding. Samsung's total NAND production for this year is expected to reach approximately 250 million TB. The company plans to leverage this globally leading capacity scale to establish its position as a core supplier for NVIDIA's CMX.

Stable V9 Production, Mass Production of V10, Trial Production of V11: Samsung Advances Three Generations Simultaneously

In terms of product lineup layout, Samsung is simultaneously advancing the production processes for three generations of V-NAND products.

Regarding V9, Samsung has increased its capacity share to approximately 60% of total V-NAND capacity, with yields stabilized above 80%, entering a mature mass production phase. A year ago, V8 still accounted for a significant portion of V-NAND output; today, its share has dropped to below 40%. The speed of this production line switch demonstrates Samsung's rapid response capability to AI server demand. Additionally, more than half of the NAND-dedicated cleanroom space at Samsung's Pyeongtaek P4 factory remains unused, providing ample room for subsequent capacity expansion.

Regarding V10, Samsung officially started mass production in the first half of this year. Currently, its output accounts for less than 5% of total V-NAND production, indicating it is in the ramp-up phase. V10 adopts a stacking architecture of approximately 400 layers, increasing storage density by over 50% compared to V9's 286 layers, allowing for higher-capacity SSDs to be installed within the same CMX module. Notably, V10 will be the first to commercially use molybdenum material on a large scale to replace traditional tungsten wiring, reducing wiring thickness by 30% to 40%. Samsung had previously applied this on a small scale in V9.

Regarding V11, Samsung initiated trial production early this year, targeting a stack layer count of 400 to 500 layers, an increase of approximately 100 layers over V10. Samsung plans to double the scale of trial production in the second half of the year to accumulate sufficient yield data and accelerate the establishment of a mass production system.

Supply Concentration Effects Emerge, Consumer Market Under Pressure

The deepening alliance between Samsung and NVIDIA in NAND, while boosting Samsung's profit expectations, has also exacerbated supply differentiation across the entire memory industry.

It is reported that Samsung's profits in 2026 are expected to exceed the sum of its cumulative earnings over the past 40 years, largely benefiting from the structural explosion in demand for AI server storage. However, the other side of this bonus is that as Samsung locks a large proportion of its capacity for major tech clients like NVIDIA, the supply of NAND to other enterprises and consumers will be noticeably squeezed, raising the risk of rising memory prices.

Previously, the AI boom first triggered a DRAM shortage and drove significant price increases; the NAND market is now facing a similar logic of supply tightening. As the shipment scale of NVIDIA's CMX continues to expand, this pressure is expected to intensify further.

In the long run, the accelerated R&D of V11 NAND may provide some relief for the consumer market—500-layer stacking technology could theoretically enable low-cost, large-scale production of high-capacity consumer SSDs, but this presupposes that Samsung is willing to allocate capacity for this purpose. Given the current direction of capacity allocation, this prospect is difficult to view optimistically.

Lee Jae-yong and Jensen Huang to Meet Soon, Cooperation Landscape Continues to Expand

The cooperation between Samsung and NVIDIA has extended from DRAM and HBM to foundry services, and now includes NAND flash memory within the alliance framework, with both the depth and breadth of collaboration continuously expanding.

It is reported that Lee Jae-yong will meet with NVIDIA CEO Jensen Huang in San Francisco in the near future. The topics of this meeting are expected to cover supply arrangements for HBM and next-generation NAND products, as well as cooperation plans for building data centers in locations such as Gwangju in Jeollanam-do, South Korea.

Industry insiders stated, "The alliance between Samsung and NVIDIA spans the entire line of memory and foundry businesses, and the cooperation discussions between the two leaders will further deepen." As NVIDIA's CMX shipments approach, the strategic significance of this meeting will become even more prominent.