⚠ Broadcom's Three-to-One Design ShareHigh threat

Marvell's problem in custom silicon is not that the market is small — it is that someone else has most of it. Broadcom holds an estimated 70% of custom-accelerator design share against Marvell's 20-25%¹, and it holds the programs that matter most: Google's TPU across seven generations, Meta's accelerator work, and OpenAI's ten-gigawatt program.

Scale compounds in this business in ways that are hard to reverse. More programs mean more IP amortised across more tape-outs, more advanced-node experience, and more credibility in the room when the next hyperscaler decides who to trust with a billion-dollar silicon bet. Broadcom can also bundle its networking silicon with the accelerator, which Marvell can partly but not fully match.

What keeps this a moderate rather than severe threat is that hyperscalers actively want a second source and will pay to sustain one — nobody building a strategic silicon program wants a single supplier with pricing power. That structural preference is most of Marvell's opportunity. Watch the design share figures over several years: a Marvell that holds a quarter of the market is a viable second source, and one that drifts toward a tenth is a specialist waiting to be bypassed.