Yesterday still the AI darling, today abandoned by the market.

Micron, down 7.7% in one day.

SK Hynix, plunged 8.32% at the open the next day.

What happened? Nvidia did one thing:

Rubin platform memory, cut from the original plan of 55TB to 28TB.
Almost halved the configuration. Memory modules cut from 192GB to 96GB.

The reason given was “justified”: supply chain tightness, prioritizing delivery.

But do you believe it?

On the same day, Jensen Huang endorsed Marvell, and AI interconnect chips surged 32% overnight.

On one side, storage is bleeding; on the other, interconnects are celebrating.

The AI hardware supply chain has, for the first time, visibly split.

This is not a coincidence. It’s a structural signal.

“Nvidia isn’t cutting memory; it’s cutting the valuation logic of storage vendors.”

The market’s previous AI hardware pricing was based on a solid premise: big models = large GPU memory, memory is never enough, and storage supercycles never end.

So Micron and Hynix rose for a year, their P/E ratios soaring to the sky.

But now Nvidia tells you: memory can be cut, performance can be compromised, but interconnects must be strengthened.

Why? Because the Scaling Law has hit a bottleneck. Single cards can no longer stack memory; costs are too high, yields too low. Nvidia’s solution: connect multiple low-end chips with interconnects, rather than stacking a super-luxury single machine.

What does this mean?

The demand structure for AI compute power is shifting from “stacking memory” to “interconnecting.”

Who benefits? Companies like Marvell that make interconnect chips.

Who gets hurt? Storage vendors.

“Before, those who had memory ruled the world; now, those who master interconnects will.”

“You thought the supply chain’s temporary reduction was just that—temporary. Actually, Nvidia is quietly changing the game.”

Is this divergence a short-term mispricing, or an early signal of the restructuring of AI compute demand?

Here’s the answer: it’s an early signal, and it’s just beginning.

Dylan Patel of SemiAnalysis later said, “This isn’t a disaster,” and their report wasn’t meant to scare the market. But listen—doesn’t this sound like reassurance before a plunge?

If this reduced configuration during Rubin mass production is confirmed, then the timeline for the AI storage supercycle will have to be redrawn. Micron and Hynix’s current prices have not yet priced in this risk.

Of course, if supply chain issues are resolved before mass production and memory is added back, today’s plunge could be a golden buying opportunity.

But the question is: do you dare to bet Nvidia will add it back?

I advise you not to bet. Because Nvidia never makes cuts without reason. When it cuts memory, there’s only one explanation: the next bottleneck in large model training isn’t GPU memory, but inter-chip interconnects.

“When industry leaders start cutting, don’t rush to buy the dip of what’s being cut.” #分享美股交易赢英伟达股票 #预测NBA总冠军赢20,000U $NVDA $MU $BTC