Next-generation NVIDIA AI systems could consume millions of TBs of SSDs and worsen storage shortages

The next hardware crisis may not be in graphics cards, but in storage.

The new architecture of artificial intelligence from NVIDIA, codenamed Vera Rubin, brings a revolutionary technology called ICMS (Inference Memory Context Storage)whose projected demand is so aggressive that it could trigger a new SSD shortage crisis global, affecting prices and availability of components for all consumers.

According to a recent Citi report, released after CES 2026, implementing Vera Rubin systems will require colossal volumes of NAND memory. It is estimated that each NVL72 rack equipped with this technology will need approximately 1.152 TB de SSDs dedicated only to support AI cache operations.

If projections of shipping 100,000 units in 2027 come true, NVIDIA alone could absorb about 9.3% of all global supply the NAND.

What is ICMS and why does it consume so much SSD?

The big bottleneck for so-called “Agentic AI”, systems that act autonomously and maintain context for long periods, is the storage of data history, known as KV Cache. Currently, this data resides in the GPUs’ HBM memory, which is extremely fast, but expensive and has limited capacity.

To solve this, NVIDIA introduced ICMS (not the one you’re thinking of), a new storage layer connected via BlueField-4 DPUs. The solution works as a compromise between RAM and traditional storage, allowing AI to access terabytes of context almost instantly.

The problem is scale: to maintain the necessary speed, the system requires banks of high-performance SSDs in massive quantities, creating a new source of pressure on memory chip manufacturers.

Numbers that scare the market

The accounts presented by analysts show a scenario of exponential growth in demand for storage:

• 2026: with 30,000 Vera Rubin units anticipated, demand would be 34.6 million TB (2.8% of the global total).
• 2027: with 100,000 units, consumption jumps to 115.2 million TB.

The volume represents a supply shock that the NAND memory industry did not appear to have fully anticipated.

With Data Centers already vying for dSilicon wafers to manufacture HBM and DRAM, the entry of this new requirement for server SSDs could create a cascade effect, also making domestic NVMe SSDs used in PC Games e consoles.

Also read:

Strong impact on the end consumer

Historically, when the enterprise server market takes a disproportionate share of component production, the hardware consumption suffers from price adjustments. That’s what happened during the mining boom and the DRAM crises we’re seeing.

If NVIDIA truly captures nearly 10% of the world’s NAND production just for its top-of-the-line AI systems, SSD makers like Samsung, SK Hynix e Micron they will have to decide between fulfilling billion-dollar AI contracts or maintaining the flow to retail. Let’s face it, we already know where this is going.

For those planning to build a PC or expand console storage in the coming years, the scenario suggests that current prices may not be maintained for long or that they will fluctuate drastically, turning high-capacity SSDs into a luxury item again.

Font(s): Jukan (X)