$NVDA just tripled CPU memory capacity on its newest rack architecture — and it's about to reshape the supply chain.

Vera Rubin NVL72 ships with 1.5 TB of LPDDR5X per Grace CPU, up from 0.48 TB on the prior GB300 generation. That's a 170% jump per chip, and rack-wide memory balloons from 17 TB to 54 TB.

Here's where it gets interesting: memory now accounts for 62% of the total Superchip cost. HBM4 and LPDDR demand is spiking so fast that some SKUs may ship with trimmed configurations — 768 GB instead of the full load — just to keep production moving.

That supply squeeze is real. $MU and peers are staring down tighter allocations and longer lead times as $NVDA scales these next-gen AI racks. If you've been watching memory plays, this is the inflection point.

Capacity leaps like this don't happen quietly. They ripple through the entire supply chain — and the companies positioned to feed that demand are the ones to watch.