Marvell launches connectivity chips for AI infrastructure

Marvell announced two connectivity products for AI infrastructure: a 3.2T protocol-aware retimer with 16 lanes of 224G PAM4 connectivity, and a second-generation 1.6T active-electrical-cable DSP built on 3-nanometer technology. The DSP uses 12.8 watts, supports copper reaches exceeding six meters, and is said by Marvell to save eight watts versus 5-nanometer designs; both products are sampling to customers. LightCounting’s CEO forecast that the AI scale-up networking market will exceed $50 billion by 2030.

Analysis

The read-through is positive for Marvell’s position in AI connectivity, but the investment signal is product qualification—not the cited market-size forecast. Faster, lower-power copper links could let operators keep more connections electrical at rack scale, potentially shifting some shorter-reach spend away from optics while increasing content per AI rack for signal-conditioning silicon. That is a substitution risk for optical suppliers, not evidence of broad optical displacement: reach, power, thermal limits and system architecture determine the mix.

The competitive test is whether Marvell converts samples into qualified designs against Broadcom, Astera Labs and Credo, and whether customers adopt the supported protocols at scale. Multi-protocol support may widen the addressable opportunity, but also makes differentiation and pricing—not compatibility alone—decisive. The company’s power and reach claims are vendor-reported; verify them in customer qualification and deployment data.

Near term (days), the launch may support sentiment but does not establish revenue. Over 1–3 months, customer qualification, design-win disclosures and deployment schedules matter. Over 6–18 months, rack architecture and copper-versus-optical mix will determine whether this becomes material growth or simply roadmap parity. The thesis weakens if sampling fails to yield design wins, customers favor competing solutions, or management does not translate connectivity demand into relevant guidance.

Key Decisions for Investors

• Do not chase MRVL solely on the launch. Treat it as a watch item until there is evidence of customer qualification, design wins, production timing or a measurable contribution to guidance.
• Conditional trade: consider a staged MRVL long after a meaningful pullback or on verified design-win evidence; size it as an execution-sensitive position. Reassess if qualification slips, competitors win key programs, or connectivity revenue/guidance fails to reflect the opportunity.
• Track copper-versus-optical content in new AI rack deployments. A faster shift to electrical links at short reach could pressure optical exposure, but avoid a broad optical short without deployment and mix data.
• Verify independently the claimed power and reach advantages, customer adoption by protocol, expected production volumes and revenue contribution. These are the key missing inputs before underwriting earnings impact.