The Entire Micron Investment Thesis Comes Down to This One Number

But investors whitethorn beryllium focusing connected the incorrect number. Instead of tracking however galore AI chips companies sell, they should absorption connected however overmuch representation engineers battalion into each chip. That favoritism whitethorn look small. But it whitethorn correspond 1 of the astir overlooked trends successful the full AI concern story.

The 1 fig investors should ticker

Most investors ticker GPU sales. The logic seems straightforward. More AI spending leads to much GPUs, which successful crook make greater request for memory.

But different inclination has softly emerged. Each caller procreation of AI hardware requires importantly much representation than the erstwhile one.

Consider what has happenedover the past fewer years:

• Nvidia's H100 accelerator contained astir 80 gigabytes of high-bandwidth memory, oregon HBM.

• The H200 accrued that fig to astir 141 gigabytes.

• The latest Blackwell chips accrued representation capableness again to astir 192 gigabytes.

Here, the nonstop figures substance little than the inclination -- that each caller procreation requires much representation than the 1 earlier it. That inclination creates a 2nd maturation operator for Micron.

Why does AI support demanding much memory?

A utile mode to deliberation astir an AI strategy is to ideate a cook moving successful a engaged kitchen.

The GPU serves arsenic the chef. Memory serves arsenic the countertop. A larger countertop allows the cook to enactment with much ingredients simultaneously. A smaller countertop forces the cook to halt and reorganize.

AI systems look the aforesaid challenge. As AI models turn larger and much sophisticated, they request entree to much accusation simultaneously. Particularly, reasoning models present execute much computational steps earlier producing an answer. Moreover, applications present process text, images, audio, and video simultaneously.

Each of those trends increases representation requirements. As a result, representation present plays a larger relation successful wide AI show than it did conscionable a fewer years ago.