NVIDIA vs. Marvell: Which AI Infrastructure Stock Is the Better Buy?

NVIDIA is presented as the preferred AI-infrastructure investment after fiscal Q2 2027 revenue rose 106% year over year to $96.22 billion and data-center revenue increased 117% to $89.02 billion. The article cites NVIDIA's 16.78x forward P/E versus Marvell's 41.74x, despite NVIDIA's stronger margins (75% versus 58.9%) and consensus growth outlook of 87.4% revenue growth in fiscal 2027. Marvell remains a high-growth AI networking and custom-silicon beneficiary, projecting roughly $18 billion of fiscal 2028 revenue, but its premium valuation, customer concentration and execution risks make NVDA the stronger risk-reward choice.

Analysis

The relevant market debate is not GPU versus custom silicon, but whether hyperscalers can convert AI capex into enough inference revenue to sustain current order intensity. NVDA's integrated compute-networking-software stack makes it the cleaner beneficiary if spending remains broad-based; MRVL has materially higher operating leverage to a smaller number of design wins, where a single program delay can impair both revenue and the multiple. The apparent valuation discount in NVDA is also partly a durability signal: recurring software and platform attachment support earnings visibility that ASIC vendors lack.

Over the next 1-3 months, the key catalyst is hyperscaler capex commentary and evidence that memory availability/costs are constraining system shipments rather than merely reducing gross margin. Higher HBM pricing is a second-order positive for SK Hynix and Micron (MU), but it can transfer economics away from system vendors and cap the upside from NVDA's revenue mix. For MRVL, optics and scale-up connectivity can outperform accelerators as cluster architectures become more bandwidth-intensive, but that upside requires verified production ramps rather than management pipeline language.

Contrarian view: a long NVDA/short MRVL expression is increasingly consensus and could reverse sharply if a major hyperscaler discloses a successful internal ASIC ramp. The more asymmetric risk is MRVL's premium multiple meeting a normalization in custom-silicon expectations; ASIC programs have long qualification cycles and lower gross-margin ceilings, so revenue growth alone may not produce proportional EPS upside. Falsify the relative-value thesis if MRVL reports sustained custom-chip gross-margin expansion and backlog conversion while NVDA's networking attach rate or gross-margin recovery misses guidance.

Key Decisions for Investors

• Initiate a 6-12 month market-neutral pair: long NVDA / short MRVL, sized dollar-neutral. Target 15-25% relative return from MRVL multiple compression and NVDA estimate resilience; exit if MRVL delivers two consecutive quarters of custom-silicon upside with expanding gross margin or if NVDA guides data-center growth materially below hyperscaler capex growth.
• Add NVDA on post-earnings volatility rather than chase pre-results strength; use a 6-month call spread financed by selling an upside call. The thesis requires confirmation that supply constraints defer demand rather than destroy it; a sustained gross-margin trajectory below management's recovery range is the core fundamental stop.
• Maintain a watch, not a directional position, in MRVL into hyperscaler earnings. Upgrade only after customer-specific evidence of production ASIC volume, backlog duration, and optical-content attach; absent that disclosure, the risk/reward remains unfavorable at a premium earnings multiple.
• Monitor AMZN, GOOG, META and MSFT capex guidance as the leading indicator for both names over the next two reporting cycles. A synchronized capex deceleration would pressure MRVL disproportionately and likely shift AI spending toward utilization and inference efficiency, favoring NVDA's installed-base software economics.