Nvidia's upcoming Rubin platform introduces CMX, a new context memory storage tier that slots between GPU HBM and bulk storage, using BlueField-4 DPUs, Spectrum-X Ethernet, and racks of NVMe SSDs to hold expansive KV cache for agentic AI workloads.[[1]](https://www.pcgamer.com/hardware/nvidia-rubin-will-soak-up-nand-supply-like-a-sponge-absorbs-water-says-one-analyst-thanks-to-the-ai-servers-new-way-of-using-ssds/)

One configuration deploys 576 SSDs for 9,600 TB per CMX-equipped setup, and industry estimates project NAND demand for these systems rising from roughly 35 million TB this year to more than 100 million TB in 2027.[[2]](https://x.com/jukan05/status/2079129709921878438)

Analyst commentary frames the surge as equivalent to layering another major consumer of NAND onto the market, with suppliers like Samsung positioning to capture the volume while consumer SSD availability faces familiar allocation pressure.[[3]](https://finance.yahoo.com/news/nvidia-blow-ssd-market-rubin-152739474.html)

The figures appear in recent coverage of Nvidia's storage architecture announcements and align with broader projections from Citi and related supply-chain reporting on Rubin server ramps.[[4]](https://www.tweaktown.com/news/109677/oh-no-nvidias-next-gen-vera-rubin-ai-systems-to-eat-up-millions-of-terabytes-of-ssds/index.html)