Qualcomm vs. Marvell: Which Underdog AI Chip Stock Has More Upside From Here?

The article argues Qualcomm offers greater long-term AI upside than Marvell, citing multi-generation data-center partnerships with Meta and Amazon. Qualcomm's Amazon arrangement includes warrants for up to 25 million shares tied to as much as $60 billion in potential chip purchases, while Meta server-CPU production is slated to begin in 2H 2028. Marvell's AI data-center revenue is projected to rise from roughly $2.2 billion in 2023 to $15 billion-$16 billion next year, but its nearly 250% stock gain over the past calendar year is presented as limiting incremental upside.

Analysis

The investable distinction is timing, not simply AI exposure. QCOM's prospective hyperscaler programs could justify a higher terminal multiple, but meaningful revenue is unlikely to affect consensus earnings for several years; the nearer-term valuation driver remains handset/Android premium-tier demand and licensing. Any market re-rating before server-volume evidence will therefore be narrative-led and vulnerable to a weak smartphone cycle or delayed silicon qualification.

MRVL's exposure is more immediate but carries a more acute design-win and customer-concentration risk: custom silicon and optical/connectivity revenue can fall sharply if a single hyperscaler changes architecture, defers capex, or vertically integrates. The non-obvious read-through from QCOM entering server CPUs and interconnect is incremental competition for ARM-based compute and AI fabric budgets, challenging not only MRVL but also AVGO's custom-silicon franchise; it does not directly displace NVDA's full-stack CUDA moat absent credible software and rack-scale performance benchmarks.

The Amazon warrant structure should not be treated as booked backlog. Its economics depend on exercise price, vesting milestones, purchase mix, and whether the stated purchasing ceiling represents incremental spend versus reallocated supplier share; until disclosed, it is both a demand signal and a potential dilution/price-concession mechanism. Contrarian view: MRVL's established deployment position may outperform QCOM over the next 12 months if AI capex remains strong, because QCOM's upside case has a long duration and substantial execution risk.

Catalysts over 1-3 months are detailed design-win disclosures, server benchmark/power data, and FY guidance changes; over 6-18 months, watch hyperscaler qualification milestones and initial revenue contribution. Falsification for a QCOM re-rating is no quantified data-center revenue target or customer deployment update by the next two earnings cycles, coupled with renewed handset margin pressure.

Key Decisions for Investors

• Do not chase QCOM solely on the partnership narrative; build a 6-18 month starter long only on pullbacks, with add criteria tied to disclosed server revenue/volume commitments or independently validated performance-per-watt. Risk: long-dated revenue never reaches earnings estimates while handset weakness dominates.
• Express relative valuation dispersion via long QCOM / short MRVL in equal-dollar size only after MRVL outperforms QCOM by another 10-15% without upward revisions; target 15-20% spread normalization over 6-12 months. Stop if MRVL raises AI revenue guidance materially or QCOM delays qualification milestones.
• Maintain a watch item on AMZN warrant filings and QCOM disclosures: initiate no dilution-adjusted revenue trade until exercise price, vesting triggers, and expected annual purchase cadence are known.
• For a more defensive AI-infrastructure posture, favor NVDA over both names on 3-6 month horizons if hyperscaler capex guidance accelerates; QCOM and MRVL are more exposed to component substitution and individual design-win timing.