Custom Silicon & the Hyperscaler ProgramsNarrow moat

The business the market cares about is custom silicon: Marvell designs an accelerator, a network controller or a storage chip to one customer's specification, using that customer's architecture and Marvell's intellectual property, and manufactures it through TSMC. It runs at roughly $1.5 billion a year across 18 cloud design wins, with management describing more than 50 active AI opportunities across over 10 major customers¹.

The moat here is the engagement itself. Winning a socket means eighteen months to three years of joint engineering before any revenue arrives, deep access to a hyperscaler's architectural roadmap, and a level of trust that is not transferable. Very few companies can do this at all: it requires advanced-node access, world-class high-speed interface IP, and the engineering scale to tape out chips that cost hundreds of millions of dollars to develop — Marvell runs roughly 50 such active AI engagements across more than 10 customers².

The weakness is equally structural. The customer owns the architecture, not Marvell, which means the next generation can be moved — and has been. What Marvell sells is capability rather than lock-in, and capability is exactly what a well-funded rival can also assemble. Watch the custom-silicon run rate against the design-win count: revenue growing more slowly than wins would mean Marvell is being awarded the smaller, less profitable pieces of each program.