CPUs, Memory & Storage
Everything in an AI server that isn’t the accelerator itself.
What this layer does
A GPU is useless without memory next to it, a host CPU to feed it, fast storage to stream training data, and on-board voltage regulation to convert 48V down to the sub-1V the silicon actually wants. The sub-categories here are all individually large markets — HBM alone is a ~$40–60B run-rate business with structural shortage — and they ride the same AI capex cycle as the GPU itself, but with different cyclicality and competitive structures.
Sub-categories
The CPU sitting next to the GPUs. Lower attach value per box than in the pre-AI world, but still material at scale.
Stacked DRAM bonded directly to the GPU package. The other supply bottleneck (alongside CoWoS) gating GPU shipments. Effectively a three-supplier market.
The conventional memory feeding the host CPU. Capacity per server keeps rising with model context windows and KV cache.
Hot training data and checkpoint storage. AI is driving QLC/PLC density and large-capacity enterprise SSDs.
The boring beneficiary — cold training data lakes. Three-vendor oligopoly.
Compute Express Link — sharing memory across servers. Astera Labs is the public name; deployment timing still uncertain.
Converting 48V/12V down to ~0.7V at thousands of amps per package. Dollar content per GPU rising sharply.
The interposers and boards under the silicon. Tight capacity for ABF substrates feeds CoWoS bottleneck.