Better AI Chip Stock to Buy Today: Broadcom vs. Marvell

Broadcom is presented as the more attractive AI infrastructure investment versus Marvell, supported by its estimated 60% AI-ASIC market share in 2026 and guidance for ASIC revenue to double in each of the next two fiscal years, reaching $230 billion in fiscal 2028. Marvell expects fiscal 2028 data-center revenue to rise 60% and its custom-chip business to more than double, but the stock has already gained more than 200% this year. Broadcom, up about 3%, trades at 18x fiscal-2027 and 11.5x fiscal-2028 forward earnings estimates, versus Marvell at 37.5x fiscal-2028 and 25x fiscal-2029 estimates, underpinning the article's preference for Broadcom.

Analysis

The relevant divergence is not simply ASIC exposure but earnings quality: AVGO monetizes scarce design capability, switching costs, and a broader networking attach, while MRVL remains more exposed to program timing, customer concentration, and component-level pricing pressure. A hyperscaler chip win can create a large revenue ramp but carries poor near-term visibility until tape-out, qualification, and volume deployment; therefore, MRVL's higher multiple embeds a narrower execution path. AVGO's lower valuation should also be viewed against its more diversified cash-flow base, which can absorb a delayed customer deployment without requiring a major estimate reset.

Near-term, the key read-through is hyperscaler capex guidance and evidence that custom silicon is incremental rather than displacing merchant GPU spend. Over 1-3 months, AVGO can rerate if management converts design-win rhetoric into backlog, shipment visibility, or raised AI networking guidance. MRVL needs a sharper catalyst—confirmed optical/interconnect demand, sustained gross-margin expansion, and disclosed production ramps—to justify its premium; otherwise, any delay in a single customer program could drive disproportionate multiple compression.

The consensus may be underestimating the bottleneck created by scale-up networking. As clusters become larger, Ethernet switching, optical DSPs, and interconnect content can grow independently of accelerator unit growth, favoring AVGO's switch silicon and selectively benefiting MRVL's optical franchise. Conversely, custom accelerators improve hyperscaler compute economics but can concentrate negotiating leverage in a handful of buyers; long-run ASIC revenue growth may not translate one-for-one into supplier margins once programs mature.

Key Decisions for Investors

• Initiate a 6-12 month pair: long AVGO / short MRVL in equal dollar amounts. The trade captures AVGO's superior execution buffer and valuation support against MRVL's elevated program-timing risk; target 15-20% relative outperformance, with a stop if MRVL reports two consecutive quarters of data-center growth and gross-margin expansion materially above guidance while AVGO misses AI-networking expectations.
• Add AVGO on post-earnings weakness only if AI semiconductor or networking guidance is maintained and the drawdown is attributable to broad semiconductor de-risking rather than a customer-specific order reduction. Use a 12-month horizon; thesis is falsified by a meaningful reduction in disclosed AI backlog, evidence of a major TPU/customer insourcing shift, or weakening switching revenue.
• Maintain MRVL as a watch rather than a fresh directional long until management provides independently measurable program milestones: customer volume production, backlog conversion, and gross-margin trajectory. A confirmed ramp with improving mix would invalidate the short leg of the pair and warrant rotating into MRVL or a long MRVL / short a higher-beta semiconductor basket.
• For a lower-beta expression of the networking bottleneck, overweight AVGO versus NVDA over the next two quarters only if hyperscaler capex remains intact but accelerator supply and pricing normalize. This is not a structural NVDA short; cover the relative short if GPU demand or pricing accelerates while Ethernet/optical order growth fails to inflect.