Last week was the worst stretch for chip stocks in over a year. The Philadelphia Semiconductor index dropped more than 11%, a hot inflation print raised the cost of funding every data center on the drawing board, and the impact of worries about CoreWeave's (CRWV -10.74%) debt load spread to every company selling into AI infrastructure.
Marvell Technology (MRVL -9.84%) sank by nearly 17% in five days. And yet, if your goal is to own one AI chip stock heading into 2027, that decline is the reason why you should look into picking up shares of Marvell now rather than later. Even as investors were selling, the company spent the week revealing its most specific, product-level work of the year, and almost nobody noticed because they were watching the tape.
CRWV CoreWeave
NASDAQ · As of 3:50 PM ET 10/22/26
What Marvell announced while its stock was falling
Marvell is not a GPU company. It designs chips that sit around GPUs: the custom accelerators that hyperscalers and AI labs are building for their specific workloads, the optical and copper links that connect thousands of processors into clusters, and the memory controllers that keep those chips fed with data. That is the business the company put on display at the Open Compute Project Global Summit in San Jose from Oct. 12 to Oct. 15.
MRVL Marvell Technology
NASDAQ · As of 3:48 PM ET 10/22/26
Two announcements from the event stood out to me. On Oct. 12, Marvell unveiled the second generation of its Structera X memory expanders -- chips built to attack what the company calls the "memory wall" in AI inference. The problem is simple: As models get bigger and context windows get longer, GPUs run out of memory before they run out of math. Structera X lets a server bolt on up to 6 terabytes of standard DDR5 memory, share it across four hosts, and compress it on the fly, which the company says cuts time to first token by 40% to 50%. The first generation Structera X is already shipping to hyperscalers, and a Meta Platforms (META -7.66%) hardware engineering director was quoted in the company's release. Test models of the second generation will be available in Q1 2027.
Image source: Getty Images.
Then, on Oct. 13, Marvell followed by unveiling two 3-nanometer connectivity chips, a 3.2-terabit retimer and a new active electrical cable chip that doubles copper reach to more than six meters between AI servers and switches. Both support UALink, the open standard that non-Nvidia accelerators use to talk to each other. Early samples of both are already available for potential customers.
That sounds like a lot of jargon, so here is the plain version. An AI processor's usefulness will always be limited by the memory it can reach and the capacity of the connections that link it to neighboring chips. The new products Marvell announced last week address both of those bottlenecks: one lets a server strap on far more memory than the GPU came with, so it can handle bigger models without choking, and the other helps racks of accelerators talk to each other over longer, cheaper copper cables without the signal falling apart. These are tools crucial for AI infrastructure, and Marvell is selling them.
META Meta Platforms
NASDAQ · As of 3:48 PM ET 10/22/26
Some numbers to consider
On Oct. 6 -- a week before the AI sector rout -- Marvell held its first investor day in years, and management used the event to publicize a target worth reading twice: It's aiming for revenue in the range of $70 billion to $90 billion by its fiscal 2031, up from $8.2 billion in fiscal 2026, with implied non-GAAP earnings above $30 per share. The engine driving it toward those targets is what management calls "XPU attach," the networking, memory, and security chips that get pulled along with every custom accelerator a cloud provider designs.
Big targets deserve some skepticism, and this one assumes 55% to 60% annualized growth for five years. But even if it achieves growth at half that rate, the stock near $228 is cheap against what the company expects to earn. Nvidia (NVDA -4.55%) saw enough to buy $2 billion of Marvell preferred stock in March, and Nvidia does not make such investments casually.
NVDA Nvidia
NASDAQ · As of 3:49 PM ET 10/22/26
Other things to consider: Marvell trades at a high trailing earnings ratio; custom chip revenue is lumpy and tied to a few customers; and if hyperscalers pull back due to higher borrowing costs, Marvell will feel the pain before Nvidia does. But last week's sell-off was driven by financing fears, not by any customers walking away, and Marvell kept shipping through it. For anyone able to buy this stock and hold onto it through 2027 or longer, the 17% discount the market just gave it makes it look like a steal.
