Marvell — AI supply-chain exposure

Marvell makes the connective silicon that moves data around an AI cluster rather than the chips that do the math. Its SerDes — the serialiser-deserialiser circuitry that pushes bits across copper and into optical links — and the DSPs inside pluggable transceivers are what let thousands of accelerators talk to each other and to storage at speed. It also designs custom compute ASICs for hyperscalers that want their own silicon. In the model's terms it sits in the infrastructure layer, a supplier of interconnect and the producer of the SerDes at its core.

The structural read is that as clusters scale, the bottleneck migrates from raw compute to moving data between chips. Every optical module carrying traffic across a data centre needs a DSP, and Marvell is one of a short list of vendors that supply them. Layered on top is its custom-silicon business, which rides the hyperscalers' drive to design accelerators in-house — giving it exposure to both the networking build-out and the merchant-GPU alternative at once.