Alchip, GUC and MediaTek: The Attack From BelowThin moat

The rivals that have actually cost Marvell business recently are not the ones with comparable technology. They are Taiwanese design-services firms with a different business model and a lower cost structure.

Alchip has secured Amazon's Trainium 3, a 3nm part expected to reach mass production in 2026¹ — a program Marvell had been associated with, and the single most concrete evidence available that hyperscaler sockets are winnable by firms Marvell does not consider peers². MediaTek has raised its AI data-center ASIC revenue guidance to about $2 billion for 2026, and GUC is expanding alongside it. All three are competing hard enough that the capital requirements of the business have become a talking point in their own right.

The distinction that matters is what they sell. Marvell sells silicon expertise bundled with its own IP — SerDes, memory interfaces, manufacturing relationships — at a price that reflects it. A design-services firm sells engineering hours to implement the customer's architecture, and charges accordingly. For a hyperscaler whose own silicon team has matured enough to own the architecture, the cheaper option is not obviously worse. That is the attack: not a better chip, but a customer that no longer needs the expensive version.

Watch the composition of Marvell's design wins, not the count. Wins where Marvell contributes its own IP are defensible; wins that reduce to implementation services are contests it will keep losing on price.