✦ Scale-Up: The Interconnect Inside the RackThin moat

The newest networking market is the one inside the rack, where accelerators must exchange data at memory-like speeds. It has been dominated by Nvidia's proprietary interconnect, and the rest of the industry has been organising to standardise an Ethernet-based alternative — an effort involving essentially every major networking and silicon vendor.

For Marvell this is a natural extension rather than a new business. Scale-up links are exactly the very-high-speed serial interfaces its SerDes IP was built for, and a market that standardises on Ethernet is one where merchant silicon vendors can participate at all — which is precisely why the consortium exists. The same IP already sells into custom accelerators, optical DSPs and switches.

There is now a second route in. In March 2026 Nvidia invested $2 billion in Marvell and brought it into the NVLink Fusion ecosystem, where Marvell supplies NVLink-compatible scale-up networking alongside custom XPUs² — the rack interior reached on the incumbent's fabric rather than around it, a trade the Competitors pages take up. The honest caveat is that the open alternative is largely a 2027-and-beyond opportunity, the standards are still being written, and the incumbent's advantage is soldered into shipping hardware. Watch the standards timetable and, more usefully, whether Marvell's scale-up content shows up inside the custom programs it already has. Winning the interconnect alongside the accelerator at the same customer is how a component supplier quietly becomes a systems one — Marvell already holds 18 cloud design wins to build from¹.