Marvell's custom silicon business is the primary growth engine and the central investment thesis. As the #2 ASIC player behind Broadcom, Marvell offers hyperscalers a credible alternative for custom AI accelerators. The AWS Trainium partnership anchors the business, while additional design wins with Google, Samsung, and others in the pipeline provide optionality.
Scale disadvantage vs Broadcom is the central bear case
Broadcom has deeper hyperscaler co-design relationships (Google since 2014), superior 3.5D packaging, and ~60% market share that drives better unit economics. If ASIC design becomes a winner-take-most market, Marvell could be squeezed on both volume and margins.
Can Marvell close the packaging gap with Broadcom's 3.5D XDSiP?
Marvell's custom silicon customer base spans multiple hyperscalers. AWS is the anchor customer through the Trainium/Inferentia partnership. Google and Samsung are confirmed design partners. Microsoft and Meta are in the pipeline. The customer base is more diversified than initially expected, providing some insulation against single-customer concentration risk.
AWS Trainium2 uses Marvell-designed custom interconnect and silicon IP. Multi-generational partnership confirmed with Trainium3 expected to tape out in 2026. AWS is Marvell's large...
Broadcom is Marvell's primary competitor in custom AI ASICs, holding approximately 60% market share to Marvell's 35%. Broadcom has structural advantages in packaging (3.5D XDSiP), deeper co-design relationships (Google since 2014), and higher margins. However, hyperscalers are incentivized to dual-source and avoid single-supplier dependence, which provides Marvell a natural second-source opportunity.
Broadcom holds ~60% custom ASIC market share vs Marvell's ~35%. The gap reflects Broadcom's earlier entry (Google TPU since 2014), deeper co-design relationships, and superior adva...
Marvell's custom silicon business model mirrors Broadcom's: NRE fees during design phase, production royalties on volume shipments, with manufacturing outsourced to TSMC. The key difference is scale -- Marvell's smaller customer base means higher R&D per dollar of revenue and lower NRE leverage. Gross margins of 45-50% on custom silicon are solid but trail Broadcom by 15+ points. The fabless model keeps capex low but R&D intensity is higher than AVGO.
Marvell's custom silicon follows the same NRE + royalty model as Broadcom. Customers pay non-recurring engineering fees during the 2-3 year design phase, then royalties on each chi...