⚠ The Contribution ShrinksHigh threat

Marvell's contribution to a custom accelerator is real and unevenly distributed: the high-speed interfaces, the memory subsystem, the physical implementation, the manufacturing relationship. The customer typically brings the compute architecture — the part it considers proprietary.

The trouble is the direction of travel. Every one of those hyperscaler silicon teams is better staffed and more experienced than it was two years ago, and the pieces Marvell supplies are exactly the pieces a maturing team learns to bring in-house or buy as licensed IP. What was a partnership in the first generation can become a supply agreement in the third and a licence in the fifth, with the revenue falling at each step even as the relationship technically continues.

That progression would not show up as a lost customer. It shows up as content per chip — the dollar value Marvell captures on each accelerator shipped — quietly declining while unit volumes rise, which reads as growth right up until it doesn't.

Custom silicon runs at roughly $1.5 billion a year across 18 wins¹. Watch that figure against the number of accelerators the industry is shipping. If Marvell's custom revenue grows more slowly than hyperscaler accelerator volumes, its content per chip is falling, and the contribution is being eroded from the inside.