If I Could Only Buy 1 Artificial Intelligence (AI) Semiconductor Stock to Buy Before 2027, It Would Be This One

Last week was the worst period for chip stocks in more than a year. The Philadelphia Semiconductor index fell more than 11%. A report showing high inflation made it more expensive to finance every data center being planned. Worries about CoreWeave (CRWV +1.65%) and its large debt also spread to companies that sell equipment and services for AI systems.

Marvell Technology (MRVL +1.41%) fell nearly 17% in five days. But if you want to own just one AI chip stock going into 2027, that drop is a reason to consider buying Marvell now rather than waiting. While investors were selling shares, the company spent the week sharing details about its products, offering some of its clearest updates of the year. Almost nobody seemed to notice because they were focused on stock prices.

What Marvell announced while its stock was falling

Marvell does not make graphics processing units (GPUs). Instead, it designs chips that work alongside them. These include special-purpose processors that large cloud companies and AI labs build for their own needs, optical and copper connections that link thousands of processors into groups, and memory controllers that help those chips get the data they need. Marvell showed these products at the Open Compute Project Global Summit in San Jose, held from Oct. 12 to Oct. 15.

Two announcements at the event stood out to me. On Oct. 12, Marvell introduced the second generation of its Structera X memory expanders. These chips are designed to tackle what the company calls the "memory wall" during AI inference, when an AI model uses what it has learned to answer questions or perform tasks. The problem is simple: As AI models grow and can handle longer amounts of text at once, GPUs can run out of memory before they run out of computing power. Structera X lets a server add up to 6 terabytes of standard DDR5 memory, share that memory among four hosts, and compress data as it works. Marvell says this can cut the time it takes to produce the first token, or piece of an answer, by 40% to 50%. The first-generation Structera X is already shipping to large cloud companies, and a hardware engineering director at Meta Platforms (META -1.59%) was quoted in Marvell's announcement. Test versions of the second generation will be available in Q1 2027.

On Oct. 13, Marvell announced two more chips made using 3-nanometer technology. One is a 3.2-terabit retimer, which helps keep data signals strong as they travel between devices. The other is a new chip for active electrical cables that doubles the distance copper cables can cover between AI servers and switches, to more than six meters. Both products support UALink, an open standard that lets AI accelerators made by companies other than Nvidia communicate with each other. Early samples of both chips are already available to potential customers.

That may sound complicated, so here is the simple version. An AI processor can only do so much if it cannot reach enough memory or connect quickly to nearby chips. Marvell's new products address both problems. One allows a server to add much more memory than its GPU has on its own, helping it run larger AI models without running out of memory. The other helps groups of AI processors communicate over longer, cheaper copper cables without losing the signal. These products are important parts of AI systems, and Marvell sells them.

Some numbers to consider

On Oct. 6, one week before AI stocks fell sharply, Marvell held its first investor day in years. Company leaders shared a goal worth looking at closely: They want annual revenue to reach between $70 billion and $90 billion by fiscal 2031, compared with $8.2 billion in fiscal 2026. They also expect adjusted earnings, measured using non-GAAP accounting rules, to exceed $30 per share. The main driver of this plan is what management calls "XPU attach." This means the networking, memory, and security chips that cloud companies need alongside each custom accelerator they design.

Investors should be careful about very ambitious targets. Reaching this goal would require annualized growth of 55% to 60% over five years. But even if Marvell grows at half that rate, its stock price of around $228 looks low compared with what the company expects to earn. Nvidia (NVDA -0.29%) saw enough potential to invest $2 billion in Marvell preferred stock in March. Nvidia does not make investments like that without careful consideration.

There are other risks to consider. Marvell's price is high compared with its past earnings. Revenue from custom chips can rise and fall sharply and depends on a small number of customers. If large cloud companies cut spending because borrowing becomes more expensive, Marvell could feel the impact before Nvidia does. However, last week's stock drop was driven by fears about financing, not by customers abandoning Marvell. The company continued shipping products during the sell-off. For investors who can buy the stock and hold it through 2027 or longer, the 17% price drop may offer an attractive opportunity.