Marvell's Moat Is Connectivity, Not Custom Silicon
The case for owning Marvell almost always opens with custom AI chips, and that is the part of the story where Marvell is structurally second. The company and Broadcom are the two names that turn hyperscaler in-house silicon designs into manufacturable chips, the application-specific processors that Amazon, Microsoft, Google and Meta increasingly favor over off-the-shelf GPUs for inference. But Marvell holds roughly 15 percent of that market against Broadcom’s 55 to 60 percent, and Counterpoint expects Broadcom to command about 60 percent of custom AI silicon by 2027. Broadcom also owns the Google TPU relationship for years to come and sells an end-to-end networking-and-ASIC bundle Marvell cannot match. Anyone buying the stock as the other custom-chip company is buying the smaller horse in a two-horse race. That is not where the edge is.
The edge is in connectivity, and it is the part of the franchise that gets the least attention. Marvell is the leader in the optical interconnect that moves data between chips, between racks and between data centers — 800-gigabit and 1.6-terabit scale-out optics, 51.2-terabit Ethernet switches, co-packaged optics, and datacenter interconnect modules. The custom-silicon platform sits on top of intellectual property that is genuinely hard to replicate: ultra-high-speed SerDes, custom high-bandwidth memory, advanced packaging and chiplets, with 5-nanometer designs shipped, 3-nanometer in progress and 2-nanometer in development. The Celestial AI acquisition added a photonic fabric for optical interconnect; the XConn acquisition added PCIe and CXL switching. Taken together, these are not a side business. They are the reason to own the stock.
The argument is physics, not sentiment. As AI clusters scale up inside a rack and out across a building, the binding constraint stops being raw compute and becomes the bandwidth between compute. Every additional accelerator multiplies the interconnect a system needs, and that demand accrues regardless of whose GPU or XPU ends up inside the box. Marvell sells the connective tissue of the AI factory, which makes it largely agnostic to the compute war raging one layer up. It is the picks and shovels of the picks and shovels.
The most important recent development reframes the entire position. In March, Nvidia invested $2 billion in Marvell and entered a partnership spanning NVLink Fusion, custom XPUs and silicon photonics. That converts Marvell from a presumed anti-Nvidia bet — the ASIC alternative to Nvidia’s merchant GPUs — into a component supplier sitting inside Nvidia’s own rack-scale ecosystem, with its silicon and optics made NVLink-compatible. It is simultaneously a validation of the technology and a measure of ecosystem lock-in that did not exist a year ago.
The growth supports the multiple, if anything does. Fiscal first-quarter revenue set a record at $2.418 billion, up 28 percent year over year, and management raised its fiscal 2028 revenue target to $16.5 billion, ten percent above the guide it had offered only three months earlier. Data center is now better than 70 percent of the business. The stock joined the S&P 500 in April, and the most bullish sell-side targets reach into the low-$300s. This is a company growing faster than Broadcom off a smaller base, which is the entire appeal.
It is also the entire risk. The valuation is rich even by AI-infrastructure standards, which makes the stock a high-beta instrument on the way down as much as up — it fell more than 6 percent in a single June session as the AI trade wobbled, and it will not be the name that holds a fear tape. The structural risk is customer concentration: revenue leans on a handful of hyperscalers, and a single design win lost to in-sourcing or a competitor is material to the model. And scale itself is a question, because Marvell is far smaller than Broadcom and Nvidia and must keep pace on research and capital against companies many times its size.
The bull case for Marvell is not that it is a cheaper Broadcom. It is that Marvell owns the connectivity layer that scales with cluster size, now carries Nvidia’s endorsement, and grows faster off a smaller base — and that you accept a premium multiple, hyperscaler concentration and real volatility to own it. Strip the story down and the least replaceable asset is the optics, not the ASICs. The market is paying for the chips. The moat is in the wires.