⚠ The Rack Interior May Never OpenModerate threat

Marvell sells well into the links between racks and between buildings. The larger and faster-growing market is the one inside the rack, where accelerators talk to each other at memory-like speeds — and that market has not been a merchant market at all. It has been the property of Nvidia's proprietary interconnect.

Marvell now has a route in, but read the terms carefully. In March 2026 Nvidia invested $2 billion in Marvell and brought it into the NVLink Fusion ecosystem, with Marvell supplying custom XPUs and NVLink-compatible scale-up networking while Nvidia supplies the CPUs, NICs, DPUs, switches and the interconnect itself¹. That is genuine access to the rack interior. It is also access granted by the incumbent, on the incumbent's fabric, revocable in the way that ecosystem membership always is.

The alternative path — an open Ethernet-based scale-up standard — is real but slower, and standards bodies do not set shipping dates.

The number to watch is scale-up content inside programs Marvell already holds. If a custom accelerator win also carries Marvell's interconnect, the company is becoming a systems supplier. If the accelerator ships with someone else's fabric around it, Marvell is selling a component into a rack it does not architect.