With generative AI evolving towards agentic systems possessing long-term reasoning capabilities, AI Agents frequently need to access massive vector databases for retrieval-augmented generation (RAG) while executing tasks. The associated data volume is enormous and characterized by highly random access, which will significantly drive up demand for high-IOPS enterprise-grade SSDs.
On the supply side, investments by overseas original equipment manufacturers are no longer directed towards large-scale construction of new wafer fabs to increase wafer input. Instead, they are highly concentrated on HBM and migration to advanced process nodes, with NAND investment being relatively cautious.
Regarding prices, NAND Flash contract prices are expected to see a comprehensive and significant upward trend. The price increase in Q1 2026 could reach 55%-60% and is anticipated to continue until the end of 2026.
Applications like AI Agents are increasing NAND demand, with the NAND storage capacity corresponding to a single GPU continuously rising. AI is fundamentally reshaping the entire computing stack. As generative AI advances towards agentic systems with long-term reasoning capabilities, AI Agents require frequent access to large vector databases for RAG during task execution. This data is vast and requires highly random access, significantly boosting demand for high-IOPS enterprise SSDs.
Benefiting from an accelerated deployment pace of AI servers, NAND demand growth is projected to be 21% in 2026. GF Securities points out that NVIDIA has launched an AI inference context storage platform, showing a significant increase in eSSD capacity per GPU.
According to GF Securities' calculations, the eSSD demand for an H100 is 4TB, for B100/B200 it is 8TB, and optimistically, for Rubin it could be 24TB. Under different GPU generation and shipment assumptions, if VR200 shipments reach 14 million units, the corresponding NAND capacity demand would be approximately 336 exabytes. AI servers are becoming the downstream application with the fastest-growing NAND demand.
Supply constraints from reduced overseas OEM investment and upgrade yield losses are intensifying the NAND supply-demand tightness. On the supply side, capacity faces multiple constraints. Investments by overseas OEMs are not being used for large-scale new wafer fab construction to increase wafer input, but are heavily concentrated on HBM and advanced process migration, with NAND investment being cautious.
NAND Flash capital expenditure is forecast to grow slightly from $21.1 billion in 2025 to $22.2 billion in 2026, a year-on-year increase of only about 5%. Samsung is prioritizing meeting HBM4 production demands, while SK Hynix's M15X fab, expected to commence production in Q4 2025, focuses on HBM and advanced DRAM. Both companies are reducing or limiting NAND Flash capex, shifting investment priority to HBM and DRAM. Micron's capital spending is more focused on advanced DRAM, with investment in the NAND sector remaining at low levels.
Simultaneously, against the backdrop of ongoing QLC production line conversions and阶段性偏低 (stage-wise low) yields, there is some loss in effective NAND capacity, further exacerbating the tight supply-demand situation.
Price-wise, NAND Flash contract prices are expected to experience a comprehensive and significant upward trend. The price increase in Q1 2026 could reach 55%-60% and is likely to persist until the end of 2026. Full-year contract prices for 2026 are projected to rise by 105%-110%, driving the NAND industry's revenue growth rate up to approximately 112% year-on-year.
Risks include slower-than-expected development and demand in the AI industry; AI server shipments falling short of expectations; and slower-than-anticipated technological and product progress by domestic manufacturers.
Comments