Rambus and Fabrinet: Small-Cap AI Picks-and-Shovels and the Asymmetric Basket

In June, the AI Trade Looked Like It Was Breaking

In June, in a matter of days, the chip index fell about 10%. The nuclear stocks that had tripled gave back 30%, and the quantum names got cut nearly in half. It looked like the AI trade was finally breaking.

Then Micron reported. Revenue was up 346% from a year ago, with the best gross margins ever, near 85%. Its most advanced AI memory was already sold out through 2026, with the shortage running all the way into 2028. Those are not signs of a bubble beginning to cool — they are clear signs the AI build-out is actually speeding up.

If that is true, the pullback is not a problem, it is an opportunity. And down at the small-cap end, where both the risk and the reward step up, I see two picks-and-shovels names worth a closer look.

Rambus: The Shovel Seller of the Memory World — With a Twist

Every high-speed memory module going into an AI server needs a small set of traffic-control chips, and Rambus makes the whole set. Its flagship is the conductor chip, the RCD, in what is basically a three-company club. So it gets paid on the memory in the AI build-out no matter which manufacturer wins.

The business underneath is completely transformed. A decade ago Rambus was a money-losing patent-licensing shell. Today it is a debt-free, profitable chip company whose operating margin swung from negative 38% to positive 37%.

But there is a twist that runs opposite to what you would expect. Rambus gets paid on the number of modules that ship, not on the price of the memory. So when memory goes into shortage and prices spike — fantastic news for Micron — fewer modules actually ship, and Rambus can feel that as a headwind.

That is part of why this is the highest-risk name on the list. It trades near 60 times earnings with a PEG well above two, and it is working through a few overhangs at once. It is responding to a document request in a Justice Department investigation, where it is not charged and not a defendant; a securities inquiry that so far is just a law firm fishing for plaintiffs rather than a filed lawsuit; and a chief financial officer who left earlier this year, though the company says there was no disagreement. So this is a small, high-risk position, sized down on purpose, and only worth it on a real pullback like the ones we have been getting.

Fabrinet: The Purest Shovel of All

The purest picks-and-shovels play in the whole basket is Fabrinet. It is the factory that physically builds the optical parts that move data through AI networks — the transceivers that connect the GPUs and the switches inside a data center. It builds them for Cisco, Nvidia, Coherent, and Lumentum, so it wins no matter whose brand is on the box.

Customers cannot easily leave, either. Aligning lasers and fibers to within a fraction of a human hair takes years to qualify, and once a program is certified in Fabrinet's clean rooms, moving it is slow and risky.

To me the demand is the real story. Its data-center interconnect business grew 90% in the past year, and what is holding it back is not orders but capacity — demand is running far ahead of what it can physically build, which is exactly why it is racing to build a giant new plant.

This is a thin-margin business by nature, though. Because the customer owns the designs, do not look at the 12% gross margin and expect much more. The quality shows up elsewhere: revenue more than doubled over six years while operating margin climbed from under 8% to nearly 10%, and it earns about 20% on its capital with no debt. The catch is that it is small and concentrated, with Nvidia and Cisco together about half of sales, and even after a 30% pullback it is not exactly cheap at a PEG around 1.5. So this is the small, higher-risk end of the optical trade — sized accordingly, with a pullback like this one making the entry a little easier.

So How Do You Own Them — the Asymmetric Basket

If I were putting money into these today, here is how I would balance the basket: heavy on the megacaps, scaling down through to the small caps. That is exactly how I run my own portfolio, and it is built to be asymmetrical.

The small bets, if they pay off, great — and if they do not, they do not drag the whole thing down. Names like Rambus and Fabrinet have real upside but real risk, so the point is to keep them from dominating the portfolio. A pullback just opens the window to own the best companies at a better price.

FAQ

Q: If small caps have the most upside, why not own more of them? A: Because the same thing that gives them upside gives them downside and uncertainty. Rambus is carrying a DOJ document request, a securities inquiry, and a CFO departure all at once, and Fabrinet has about half its sales concentrated in just two customers, Nvidia and Cisco. Names like these can reward you handsomely when right, but the core of an asymmetric design is sizing them small enough that being wrong does not shake the whole portfolio.

Q: Why is Rambus the riskiest name here? A: Because its business model runs counter to intuition. Rambus earns on the number of modules shipped, not the price of memory, so when memory goes into shortage and prices spike, fewer modules shipping can actually become a headwind. Layer on a rich valuation near 60 times earnings and several legal and personnel overhangs, and the risk is clear.

Q: Why do Micron's earnings matter so much? A: Because they were the signal that decided whether June was an AI bubble bursting or a healthy pullback. Revenue up 346%, best-ever gross margins near 85%, advanced AI memory sold out through 2026, and a shortage running into 2028 all show demand accelerating rather than cooling. Those numbers are the basis for treating the pullback as a buying opportunity.