The Co-Design RelationshipNarrow moat

A custom silicon engagement does not resemble selling a chip. Marvell's engineers work inside a hyperscaler's architecture team for eighteen months to three years, translating a customer's design intent into something manufacturable at an advanced node — choosing memory interfaces, laying out high-speed serial links, closing timing, and managing a tape-out whose mask costs alone run to tens of millions of dollars.

That long engagement is the closest thing this business has to a switching cost. A customer three years into a joint design does not casually restart with someone else, and the institutional knowledge Marvell accumulates about a particular architecture makes it the path of least resistance for derivative parts and mid-life revisions.

But the lock is on the program, not the customer. When the architecture changes generation — new process node, new memory standard, new interconnect — the advantage resets substantially, and the customer is free to run a fresh competition using everything it learned from the last one. Watch whether Marvell retains programs across generational boundaries. Holding a socket through a node transition is the only real evidence that the relationship, rather than the price, is what won it — and Marvell had 18 such cloud design wins to defend¹.