Memory is no longer a “side-component” narrative — it has become a profit centre in the AI spend stack.

HBM is now the true choke point.

Not GPUs. Not wafer supply.

Whoever owns the HBM / NAND / DDR5 pricing power owns the incremental dollar of AI compute.

And that is why storage stocks are behaving like GPU cyclicals.

Why this rotation is logical

For LLM training and inference, the cost of feeding the accelerators — low-latency, high-bandwidth, high-capacity memory — is the bottleneck.

Nvidia’s own COGS leverage depends on partners’ ability to ramp HBM.

Hyperscalers cannot expand AI clusters without memory.

So investors are finally pricing memory at the top of the value chain — where it always belonged in AI infra, but was not valued as such last year.

On the Sandisk / $30+ EPS scenario

Morgan Stanley’s $30+ cycle EPS implies two aggressive premises:

1. structural HBM scarcity persists into 2026 (no oversupply)

2. pricing discipline is maintained (no Korean or Japan price war)

If both stick — Sandisk is a margin-monster.

This is why the bull targets north of US$250 – US$300 do not look like fantasy in this regime — because this is not the 2017 NAND cycle. AI shifts the memory curve upward and flattens the downside.

Forward watch

The key question now:

> Can memory makers avoid the old mistake — chasing share instead of price?

If they keep supply disciplined, storage remains the second purest AI leverage trade after Nvidia.

If they break discipline — this becomes another 18-month boom–bust.

That — more than TAM headlines — is what determines whether these stocks merely pop or actually compound beyond 2025.