Nvidia Is Up 7% in 2026 While Marvell Technology Has Nearly Doubled. Here's Why.

Marvell Technology is benefiting from hyperscalers' shift toward custom AI accelerators, with its data center revenue up 46% to $6.1 billion in fiscal 2026. The article says custom ASIC shipments could triple by 2027 versus 2024, supporting Marvell and Broadcom as demand grows for semi-custom AI infrastructure. Nvidia remains strong, but the piece argues Marvell may outperform over the next five years.

Analysis

The market is re-rating AI capex from a single-vendor GPU cycle to a barbell of general-purpose compute plus bespoke silicon. That is structurally bullish for the semiconductor “picks-and-shovels” layer, but the second-order winner is whoever controls the design-in workflow and interconnect standards; once a hyperscaler commits to a custom roadmap, switching costs rise sharply and revenue visibility extends well beyond the initial chip shipment. The implication is that the value pool shifts away from pure performance leadership toward integration, packaging, and semi-custom system content.

This is not a clean displacement of NVDA. Custom accelerators typically expand total AI spending by lowering inference cost per token, which can increase deployment intensity and keep overall compute demand growing faster than unit share loss. The risk for NVDA is mix, not collapse: if hyperscalers use custom silicon for steady-state inference while reserving GPUs for frontier training, Nvidia’s pricing power remains intact but its attach rate in large accounts can compress over 12–24 months.

MRVL’s upside is real, but consensus may be underestimating how cyclically levered its rally has become to a handful of design wins and customer concentration. A delay, scope reduction, or in-house replacement at one large cloud customer would hit near-term sentiment harder than revenue because the stock is now trading as a secular AI compounder, not a cyclical network semiconductor name. The contrarian risk is that the custom-chip narrative becomes crowded precisely as hyperscalers optimize capex, causing estimates to outrun actual shipment ramps over the next 2–3 quarters.

Broadcom remains the cleaner way to express the custom-ASIC trend because it has the broadest embedded base and more diversified software/cash flow support, while AMZN and GOOGL are the best hidden beneficiaries if lower-cost inference improves cloud economics and ads/search throughput. INTC is the weakest read-through here: if the market validates design-heavy custom silicon, it further highlights Intel’s lack of differentiated AI pull-through. On balance, the trade is not “short Nvidia,” but “own the custom-silicon enablers and fade overextended valuation in the pure beneficiaries if execution slips.”

Key Decisions for Investors

• Long AVGO vs short NVDA for a 3–6 month relative-value trade: AVGO has more durable custom-ASIC exposure and better downside protection if AI spend normalizes; use a 1:1 dollar-neutral structure and take profit if NVDA re-accelerates on Blackwell guidance.
• Add to MRVL only on a pullback of 8–12% or after the next hyperscaler commentary cycle; upside remains meaningful over 12 months, but the position should be sized as a momentum trade with a hard stop if any major design win is delayed.
• Long GOOGL and AMZN into 6–12 month horizon as indirect beneficiaries of cheaper inference economics; the best risk/reward is that custom silicon improves unit economics faster than consensus models, supporting margin expansion.
• Avoid or underweight INTC in this theme basket; if you need semiconductor exposure tied to AI infrastructure, prefer names with embedded design-win visibility and interconnect leverage rather than turnaround optionality.
• For options, consider a call spread on AVGO or MRVL over 9–12 months rather than outright equity to cap valuation risk; the thesis is durable, but the stocks are now vulnerable to any capex digestion quarter.