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

Qualcomm is positioning for a multi-year AI data-center expansion through its Dragonfly C1000 server CPU and AI300 accelerator, supported by a Meta server-CPU supply agreement beginning in 2H 2028 and an AWS custom-chip collaboration. Qualcomm issued Amazon-linked warrants for up to 25 million shares, contingent on as much as $60B in potential chip purchases, creating a significant demand-alignment mechanism. The article argues Qualcomm offers greater upside than Marvell because its AI pivot may be underappreciated, whereas Marvell shares have risen nearly 250% over the past calendar year despite projected data-center revenue growth from $2.2B in 2023 to $15B-$16B next year.

Analysis

QCOM's prospective cloud exposure should not be capitalized like near-term accelerator revenue: the meaningful production window is roughly two years out, and customer-linked warrants create an important offset to headline purchase commitments through dilution and potentially favorable pricing. The nearer valuation catalyst is evidence that Qualcomm can translate co-development into funded design wins, disclosed backlog, or non-handset revenue guidance—not partnership language. Until then, handset royalty/chip cyclicality and Android premium-unit trends remain the earnings driver.

MRVL is more exposed to the current AI networking/optics capex cycle, but that also makes it vulnerable to hyperscaler digestion, custom-silicon program timing, and customer concentration. A shift from scale-up GPU clusters toward more distributed inference would sustain optical-content demand, while vertically integrated alternatives from NVDA and internally designed hyperscaler networking silicon could compress its addressable margin pool. The relevant comparison is therefore QCOM as a low-expectations, long-dated option on cloud compute versus MRVL as a higher-beta play on present deployment intensity.

Consensus may underappreciate that hyperscalers use supplier warrants to secure capacity and lower effective procurement cost; this is not automatically proof of superior economics for QCOM. Conversely, QCOM's broad edge/client footprint offers a potentially less appreciated inference distribution channel: if on-device AI materially reduces cloud inference load, QCOM could gain in endpoints even as centralized accelerator demand normalizes. That outcome favors QCOM's diversification but weakens the simple "next data-center winner" framing.

Over the next 1-3 months, monitor QCOM's disclosed data-center R&D intensity, gross-margin trajectory, warrant accounting, and any quantified customer commitments. Over 6-18 months, the thesis is falsified if cloud-related design wins fail to expand beyond anchor customers, or if data-center investment requires material margin dilution without a visible revenue ramp; it is validated by backlog disclosure and an upward revision to non-handset revenue or long-term operating-margin targets.

Key Decisions for Investors

• Initiate a modest long QCOM / short MRVL pair over a 6-12 month horizon, sized beta-neutral. The trade monetizes lower embedded cloud expectations at QCOM versus MRVL's sensitivity to AI-capex deceleration; reassess if MRVL delivers sustained order growth above expectations or QCOM cannot quantify incremental cloud backlog by its next two reporting cycles.
• For directional QCOM exposure, stage entries around earnings rather than chase partnership headlines. Add only if handset guidance remains stable and management identifies funded data-center milestones; target a 15-20% upside on multiple re-rating, with a 8-10% stop if non-handset margin guidance deteriorates or warrant dilution/economic concessions exceed expectations.
• Maintain MRVL as a tactical, not structural, AI holding into the next earnings cycle. Reduce on upside moves unsupported by backlog or gross-margin expansion; a 10-15% downside is plausible on a single hyperscaler program delay given elevated expectations and concentrated customer exposure.
• Set an alert for disclosures from AMZN or META on internal CPU/accelerator deployment, optical interconnect spending, and supplier commitments. These are the missing independent data points needed to convert QCOM's long-dated collaborations into a higher-conviction revenue estimate.