AI development is shifting from model training to inference-centric Agentic AI, creating a massive surge in memory demand. TrendForce now projects the global memory market will hit $1.28 trillion by 2027. This shift happens because AI agents use iterative cycles and larger context windows, which forces a reliance on high-performance DRAM and HBM to manage KV cache efficiency.

Server architectures are changing to support these workloads, increasing the number of CPUs needed for scheduling and preprocessing. While traditional GPU ratios were 1:8, NVIDIA is using a 1:2 configuration in its NVL72 rack. This shift increases server DRAM requirements and gives suppliers more leverage in contract pricing, especially as HBM production eats into conventional DRAM wafer capacity.

NAND flash is also seeing a spike as cloud service providers ramp up infrastructure spending. Because HBM is too expensive and HDDs are too slow for real-time workloads, high-performance SSDs like SCM and SLC are filling the gap. TrendForce revised its 2026 NAND forecast to $270.6 billion as enterprise token consumption rises fourfold.

These structural shifts mean memory prices will likely rise through 2027. The combination of increased CPU deployment and the adoption of specialized SSDs creates a supply deficit that conventional hardware cannot easily resolve.