The Two Chip Stocks That Kept Rising as the Sector Fell: Broadcom and Marvell

On the days the semiconductor trade cracked this year, nearly every chip stock fell at once. July was the sector's worst month in roughly four years, down about 13%, and sessions like this week reopened the slide with the Nasdaq futures knocked down as chip names led the way down. Yet in the middle of a tape where the whole group moved like a single bet on AI spending, the same two designers kept landing on the wrong side of the red. Broadcom and Marvell. The question is not whether they were lucky. It is what they hold that the rest of the pack does not.

Broadcom and Marvell do not win the same dollar that Nvidia does. Nvidia sells the merchant GPU that dominates the training stage of AI — the part of the cycle where its CUDA software moat is strongest. Broadcom and Marvell sit on the other side of that cycle: they design the custom chips, called ASICs, that hyperscalers commission for the inference stage, the part of running a model that happens over and over and covers the ongoing cost. Broadcom designs the Tensor Processing Unit chips at the heart of Google's AI infrastructure. Marvell has built the same relationship with Google into a much larger one.

That difference showed up in operating results, not just in pricing. In the quarter Broadcom reported at the start of September, its AI accelerator revenue hit $16.7 billion, up 221% from a year earlier and up 54% sequentially — while its chief executive guided to roughly doubling that stream again next fiscal year. Marvell closed the same stretch with record quarterly revenue, data-center revenue up about 46% year over year, and guidance that came in ahead of expectations. And in August, Google expanded its custom-silicon relationship with Marvell through an agreement that carried a warrant tied to as much as $120 billion of future purchases; Marvell's stock jumped roughly 10% on the news while several rivals slipped.

Watch the two companies' revenue and margins, and these are not flattering-spec stories. Broadcom converts an unusually large share of sales into cash — an operating margin near 43% and a similar free-cash-flow margin — which is why a company carrying a roughly $1.7 trillion market value can still be adding value. Marvell is smaller and thinner by comparison, around a $200 billion market value with a 16% operating margin and roughly 31% revenue growth, but it is being valued on acceleration, not on the margin it has reached yet.

Why the exceptions point somewhere real

The reason the same two names keep breaking from the falling pack is that they are being priced for a different stage of the cycle. The contested share of value in AI is shifting from training to inference. Training runs on Nvidia's CUDA moat; inference is decided on cost per output, latency, and efficiency. That is precisely the dispute where custom silicon wins: a hyperscaler running inference at scale can design a chip that does the job a merchant GPU does for a fraction of the ongoing cost. Building your own chip is the buyer's answer to the chip seller's pricing power, and it hands the design work — and the economics — to Broadcom and Marvell. When the crowded AI trade unwinds, the selling concentrates in the liquid, dominant names, while the custom pair keeps being valued on revenue that is still compounding.

None of this makes them immune. Broadcom and Marvell were both down again this week alongside the group, and the exceptions can become the next crowd. Marvell already trades at roughly 21 times sales, a multiple that demands the acceleration keep arriving. Broadcom, despite the cash flow, sits near 45 times trailing earnings. A rotation out of the merchant-GPU trade can easily become a crowded rotation into custom silicon, complete with the same habit of pricing in the next quarter before it is reported.

What separates a migration from a crowded trade

Here is the discipline that matters. The reason to respect the two names is not that they look cheap — they do not. It is that they sit on the stage of the cycle the market is trying to price next. The custom-chip story only stays real for as long as the guidance reaches operating results: utilization, delivery, and margin on the chips hyperscalers actually deploy, not the warrants and projections on paper. Google's commitment to Marvell, and Broadcom's guidance of roughly doubling its AI revenue, are conditional on the same thing every supply commitment is conditional on — whether the customer deploys, runs them, and keeps buying at the promised rate.

The pair soared on down days because they represent where the value in this compute cycle is migrating, and the market had not crowded into them as fully as it had into the leaders. That is exactly the moment I have spent this cycle looking for — a dark horse on the next generation, priced on the economics rather than the headline. But the same judgment that put capital there applies on the way out: the exceptions are where the value is going, and also where the next overcrowding will form. What changes the conclusion is not another price target. It is whether the revenue that made Broadcom and Marvell the two names that wouldn't fall keeps showing up in the numbers.