How does Marvell make money from custom AI chips if it doesn't design them?

Marvell supplies the surrounding infrastructure semiconductors: high-speed SerDes interfaces, optical DSPs, and interconnect PHYs that connect custom AI chips to each other and to the network. Every AWS Trainium or Google Axion deployment needs Marvell components to function at scale.

Has the Inphi acquisition delivered on its original thesis?

Yes. Inphi's optical DSP portfolio was acquired for approximately $10B in 2021 with the thesis that data center interconnect bandwidth would explode. The AI training cluster buildout—requiring ultra-high-speed connectivity between thousands of accelerators—has validated that thesis through FY2024-FY2026 revenue contributions.

What is Brightlane and how significant is the automotive opportunity?

Brightlane is Marvell's automotive Ethernet semiconductor brand. As EVs and ADAS systems demand far more in-vehicle bandwidth than legacy CAN bus can provide, Marvell's 100BASE-T1 and 1000BASE-T1 PHYs are designed into new vehicle platforms. Once designed in, volumes are stable for several years.

MRVL Stock Outlook 2026: The Custom ASIC Infrastructure Bet

Nvidia dominates the AI accelerator conversation, but underneath the GPU layer sits a dense ecosystem of semiconductors that nobody talks about—and Marvell Technology has been quietly building dominant positions in several of them. Custom ASIC connectivity, optical DSP for hyperscale interconnects, PAM4 Ethernet switching silicon, and automotive Ethernet all represent segments where Marvell competes with real technical depth rather than as a commodity supplier. The 2026 investment case is whether these segments compound together into a durable above-market growth profile.

Custom ASIC Wins: The Infrastructure Behind AWS Trainium and Google Axion

Hyperscalers are designing their own AI chips for a straightforward reason: at their scale, custom silicon beats off-the-shelf alternatives on performance-per-dollar for specific workloads. AWS Trainium, Google’s TPU family and Axion Arm-based CPU, and Meta’s MTIA represent billions of dollars of custom semiconductor investment.

What most coverage misses: each of these custom chips requires high-speed SerDes (serializer/deserializer) interfaces, optical DSP components, and interconnect fabric to move data between chips and to the network. Marvell designs and manufactures these enabling components. The more hyperscalers invest in custom AI silicon, the larger Marvell’s addressable revenue from surrounding connectivity semiconductor becomes.

Marvell’s earnings calls reference custom ASIC-related revenue as among the fastest-growing segments in the data center business, with named hyperscaler design wins contributing materially to the revenue trajectory.

Related: AI Stocks Investment Guide 2026 →

Inphi: The $10 Billion Bet That Paid Off

Marvell acquired Inphi in 2021 for approximately $10 billion in a deal that looked expensive to some analysts at the time. Inphi was the leading supplier of DSPs (digital signal processors) for high-speed optical transceivers—the components that convert electrical signals to optical signals for long-distance data center interconnects.

The investment thesis was that data center interconnect bandwidth requirements would grow exponentially. The AI training cluster buildout validated that thesis faster than anticipated. Training a modern large language model requires thousands of accelerators to exchange gradients at ultra-high speed, making interconnect bandwidth one of the primary performance constraints. Demand for 800G and emerging 1.6T optical modules has pulled Marvell’s optical DSP revenue to material scale within the data center segment.

PAM4 Switching Silicon: Competing in Broadcom’s Core Market

PAM4 (four-level pulse amplitude modulation) is the signaling technology that enables 400G and 800G Ethernet at data center scale. As hyperscalers and enterprise data centers upgrade their ToR (top-of-rack) and spine switches to support AI cluster connectivity, PAM4 PHY semiconductors are a required upgrade.

Broadcom holds a dominant position in this market with its Tomahawk and Trident merchant silicon lines. But Marvell competes in PHY-layer Ethernet components with genuine technical credibility, holding share in segments that Broadcom’s vertically integrated switch ASICs do not fully cover. The AI infrastructure upgrade cycle running through 2025-2027 is a sustained demand driver for Marvell’s Ethernet portfolio.

Brightlane: The Long-Duration Automotive Option

Brightlane is Marvell’s automotive Ethernet brand, covering 100BASE-T1, 1000BASE-T1, and multi-gigabit PHYs for in-vehicle networks. The problem Brightlane solves: modern electric vehicles have over a hundred electronic control units, advanced driver assistance cameras, LiDAR sensors, and high-resolution displays that require far more bandwidth than the CAN bus architecture of the 2010s can provide. Automotive Ethernet is the upgrade path.

Automotive semiconductor design-ins operate on a different timeline than data center silicon. The design cycle from specification to volume production spans three to five years, and once designed in, a supplier typically supplies that platform for the vehicle’s production lifetime—often seven to ten years. Brightlane wins confirmed in 2022-2024 are now entering volume production, providing a multi-year revenue floor for the automotive segment.

Bull Case vs Bear Case

Bull case

AWS Trainium3 and Google’s next-generation custom silicon projects confirm expanded Marvell optical DSP and SerDes design wins, lifting data center revenue materially
PAM4 switching upgrade cycle peaks in FY2026-FY2027, accelerating Ethernet PHY revenue
Brightlane achieves its first $1B+ quarterly automotive revenue run rate as new EV platform volumes ramp

Bear case

Hyperscalers internalize optical interconnect design or diversify to Coherent (formerly II-VI), reducing Marvell’s share in next-generation modules
An AI capex spending pause by hyperscalers delays data center semiconductor demand
Automotive market slowdown compresses EV production volumes, pushing Brightlane ramp timing to the right

Related: AVGO Broadcom Stock Outlook 2026 →

Tax Considerations for US Investors

MRVL pays a small quarterly dividend but at a minimal yield—treat it as a growth stock for tax planning purposes. Qualified dividends are taxed at 0-20% depending on income bracket. Capital gains on shares held over 12 months qualify for 15-20% LTCG rates. Given MRVL’s history of significant earnings-day moves, Roth IRA placement is attractive for long-term holders who believe in the AI infrastructure buildout thesis.

Position sizing relative to other semiconductor holdings matters: MRVL has meaningful correlation with NVDA, AVGO, and ASML during semiconductor sector risk-off events. Investors who want semiconductor exposure without concentration might consider an ETF sleeve alongside single-stock positions.

Bottom Line

Marvell enters 2026 as one of the cleaner stories in AI infrastructure semiconductors. The Inphi thesis has been validated by the optical interconnect demand wave. Custom ASIC design wins with top hyperscalers provide multi-year revenue visibility. PAM4 switching and automotive Ethernet add diversification that pure-play AI chip exposure lacks. The key risk is hyperscaler supplier diversification or internalization reducing Marvell’s unit economics per custom chip generation.

This article is for informational purposes only and is not investment advice.