Infrastructure

HBM (High Bandwidth Memory)

A type of memory stacked directly next to an AI chip to feed it data fast enough to keep up with its compute — a key bottleneck in the AI buildout.

AI accelerators like GPUs can perform vastly more calculations per second than conventional memory can supply them with data — a gap known as the “memory wall.” HBM closes part of that gap by stacking multiple layers of memory chips vertically and connecting them to the processor with a very wide, very short interconnect, dramatically increasing bandwidth compared to standard DRAM.

HBM supply is dominated by a handful of makers — principally SK hynix, Samsung, and Micron — and demand from AI training and inference has turned it into one of the tightest-supplied, most strategically important components in the entire AI stack.

Related analysis