Key points
- The AI trade has rotated by sector almost every week: software, then photonics, then memory.
- I think custom AI chips and the connectivity names are the next leg, led by Marvell (MRVL) and Broadcom (AVGO).
- The fuel is a wall of earnings from Aug 26 to Sep 3, starting with NVDA and MRVL.
I've been staring at the same setup for three weeks, and it finally clicked. AI no longer trades as a single block. One corner gets the attention, runs hard for roughly five days, and then gives way to another.
Watch the tape with me. Three weeks ago it was software, and Microsoft (MSFT) ran 22% the week it reported earnings. The next week the money sprinted into photonics, the optical-networking crowd, and Coherent (COHR) went up 44% in five sessions, an optics sweep we wrote up at the time. Last week it was memory and storage, where Sandisk (SNDK) jumped 35% and SK Hynix (SKHY), the Korean memory giant, ran 21% on its US shares, part of the memory squeeze we've been tracking. Same trade, different jersey.
| Week | Sector that led | Standouts |
|---|---|---|
| Jul 27-31 | Software | MSFT +22% |
| Aug 3-7 | Photonics / optics | COHR +44%, AAOI +44% |
| Aug 10-14 | Memory / storage | SNDK +35%, SKHY +21%, STX +20% |
| Aug 10-14 | AI clouds / data centers | SHAZ +54%, NBIS +48%, CRWV +16%, IREN +7% |
And Nvidia (NVDA)? It just ground higher through the whole thing, like it wasn't even part of the rotation.
The handoff was not perfectly neat, either. AI cloud and data-center stocks ran even harder than memory last week. SharonAI (SHAZ) surged 54%, and Nebius (NBIS) added 48%. CoreWeave (CRWV) joined the move, as did former bitcoin miners that now host AI workloads, including Iris Energy (IREN).
Why does this keep happening? The money stays inside AI but refuses to chase whatever already ran. It parks in the group that hasn't had its turn yet, and that group is almost always the one about to report earnings. Nobody wants to pay up for last week's winner.
That leaves one question: who gets the next turn?
My money's on the plumbing. The custom chips and the wiring that hold an AI data center together. Marvell (MRVL) and Broadcom (AVGO) are the two big names there, and this is the week I'd start leaning in.
Nvidia's GPUs get most of the attention, but a GPU alone in a box accomplishes nothing. Tens of thousands of chips have to communicate, and the data moving between them cannot be allowed to choke the system. Marvell and Broadcom operate in that layer. They design custom AI chips for the large cloud companies, as well as the switches and high-speed connections linking rack to rack. Credo (CRDO) and Astera Labs (ALAB) sit one level farther down the chain, supplying cables and connectors.
The thesis rests on actual contracts, not a loose story. Broadcom helps design Google's in-house AI chips. Marvell has a multiyear custom-silicon agreement with Amazon. When a cloud giant decides to develop its own chip instead of buying exclusively from Nvidia, one of those two frequently wins the work. The business is growing quickly, yet the stocks barely participated while the previous three groups had their moment.
One detail sealed it for me. When optics had its monster week in early August, Marvell quietly ran 17% and Credo 21% right alongside it. Neither got a week of its own, though, and Broadcom actually slipped last week while memory ran. This group keeps showing up to the rally without ever leading it, and I think that changes at the prints.
The calendar is what makes the call timely. Nvidia reports Aug. 26, and Marvell follows the next evening on Aug. 27. Credo is due Sept. 1, Broadcom Sept. 2 and Ciena (CIEN) Sept. 3. That is a solid wall of results from the exact companies in this part of the stack. The prior three rotations followed much the same script: buyers moved in before the numbers, the companies reported and their sector led the week. This group begins reporting Aug. 26. If the sequence repeats, the time to arrive early is now.
There is a fair case against all of this. I may be looking at an earnings calendar and calling it a strategy. Whichever sector is about to report gets bought, and I have mistaken that behavior for a larger trend. Nvidia also remains the biggest risk. If its Aug. 26 results disappoint, the whole group is likely to go red together, regardless of whose turn it was supposed to be. Valuations offer little protection: Broadcom trades around 65 times earnings, and Marvell is above 75 times. A weak outlook could punish either stock quickly. Any rotation thesis lasts only until the market stops rotating.
So that's my read. This week I'm watching Marvell and Broadcom, with Credo and Astera Labs as the higher-octane way to play the same idea. I don't own any of them yet, so I'm talking my thesis and not my book. Ask me again in two weeks and the baton may already be in some other group's hands. That's the fun of this tape right now. It will not sit still.
