JP Data’s demand model sizes the server market from OEM, ODM, and hyperscaler-direct channels, then splits accelerated versus general-purpose systems. CPU demand is rolled from the bill of materials into x86 and Arm. Accelerator demand separates GPU from ASIC / TPU. RAM demand splits server DRAM into HBM and DDR. SSD demand is the enterprise / server slice of Others — including OEM, ODM, and hyperscaler-direct (MAG / MANGOS / neoclouds in the HS slice), not a full NAND TAM. Together these views show where AI infrastructure dollars are going through 2025. Continue
The supply model tracks who can fill that demand. CPU supply is estimated Intel and AMD x86 revenue plus Arm server CPU vendors. Accelerator supply covers NVIDIA GPU, AMD Instinct, and captive or merchant ASIC (TPU, Trainium, Maia, MTIA, Gaudi). RAM supply is SK hynix, Samsung, and Micron server DRAM — HBM plus server DDR, excluding consumer memory. SSD supply is enterprise SSD from Samsung, SK hynix, Kioxia, Sandisk, Micron, and Solidigm; client NAND is excluded. Continue
Signals turn demand and supply into market tightness and spending indicators. Coverage is supply divided by demand for CPU, accelerators, RAM, and enterprise SSD. Lead time is modeled channel weeks for CPU, GPU, and HBM against research survey bands. Hyperscaler spend follows MAG, MANGOS, and neocloud CapEx. Use these series to see when the stack is tightening or easing. Continue
The Forecast page on this public cut shows history through 2025. The full model extends further. Continue