Broadcom Owns the AI Accelerator Scale, While Marvell Builds the Links
Broadcom’s custom AI accelerators are already a $16.7 billion quarterly business, while Marvell is growing around optics and networking. The comparison turns on scale, supply commitments and customer concentration.
Broadcom and Marvell both benefit from hyperscalers building custom artificial-intelligence infrastructure, but they occupy different positions in that spending chain. Broadcom has become a large-scale supplier of custom accelerators and networking. Marvell is smaller and is expanding through optical links, switches and chips that connect accelerators to memory, storage and networks.
Broadcom reported fiscal third-quarter revenue of $29.6 billion, up 86% from a year earlier. AI semiconductor revenue reached $16.7 billion, up 221%, and management expects it to rise to about $21.7 billion in the fourth quarter. The company generated $13.7 billion of free cash flow in the quarter. Custom accelerators supplied most of the AI revenue, while networking products such as Tomahawk switches support the surrounding data-center fabric.
That scale is Broadcom’s advantage. It can co-design chips with a small group of large customers and fund multi-generation road maps. It also creates concentration risk. A delayed deployment by one hyperscaler can affect a large revenue pool, and the growing mix of memory-heavy accelerators has pressured gross margin even while operating margin remained strong.
Marvell reported $2.74 billion of revenue in the period described by the source, up 36.5% year over year, with data centers accounting for 79% of sales. Its opportunity is less about displacing Broadcom’s leading custom accelerators and more about supplying the links around them. Management expects 1.6-terabit optical products and 51.2-terabit switches to drive growth as data centers move more information among large clusters.
Marvell’s relationship with Google illustrates both the opportunity and the tradeoff. It supplies networking and controller products around Google’s custom systems, while a warrant could allow Google to acquire up to 7% of Marvell if revenue milestones are met. That aligns a major customer with growth but could dilute existing shareholders.
Supply is a constraint for both companies. Broadcom has identified power, high-bandwidth memory and data-center shells as possible bottlenecks. Marvell plans roughly $1 billion of supplier prepayments to secure capacity. Those commitments make sense if demand materializes, but they increase the cost of a slowdown or project delay.
The investment distinction is therefore straightforward. Broadcom offers proven scale, high cash generation and direct exposure to custom accelerators. Marvell offers faster percentage growth from a smaller base in the connectivity layer, with greater execution and dilution risk. The next reports should show whether Broadcom can convert its $21.7 billion AI outlook into shipments and whether Marvell’s optical ramp appears in reported revenue rather than only design wins.



