Marvell Technology does not make the AI chip you have read about in the headlines. It makes the chips that let those chips talk to each other - the optics, the switches, the interconnect and, increasingly, the custom accelerators that hyperscalers design for their own data centers.
Founded in 1995 in Sunnyvale, California, Marvell began life as a storage-chip company. Its first product was a CMOS read-channel chip for hard disk drives, and its first customer was Seagate, which took volume shipments beginning in 1996. For years the company was a broad merchant-silicon supplier, its parts tucked inside consumer devices, storage gear and networking equipment that carried its name on almost nothing.
That quiet breadth has narrowed into something sharper. Today Marvell is a fabless semiconductor company - it designs chips and hands manufacturing to foundries such as TSMC - focused on data infrastructure. Its portfolio spans custom XPUs and ASICs, PAM4 electro-optics, high-speed Ethernet switching, data processing units, storage controllers and interconnect silicon such as PCIe retimers and CXL memory controllers.
The through-line is movement. Wherever data has to travel - inside a server, across a rack, between buildings in a hyperscale campus - Marvell sells the silicon that moves it faster and with less power. As AI models have grown past the limits of a single chip, that plumbing has become the bottleneck, and the bottleneck has become the business.
The company is headquartered in Santa Clara, California, employs roughly 7,480 people, and holds more than 10,000 patents. It trades on Nasdaq under the ticker MRVL.
Marvell describes its strategy plainly: invest only in markets where it can be number one or number two. That is a harder rule than it sounds, because it means walking away from revenue it could have kept.
On Marvell's stated capital-allocation discipline
