Move Over Nvidia, This Custom Chip Stock Could Be the Next Big Winner

Marvell (MRVL) is highlighted as a potential AI-chip winner as the market shifts from AI training to lower-power inference, where custom AI processors can outperform GPUs on efficiency. The article claims Marvell controls ~20% to 25% of the custom AI chip market and cites a potential $75B lifetime revenue from 50+ opportunities across 10+ customers, versus Nvidia’s ~80% AI chip share that is “gradually slipping.” Marvell shares are up 144% in 2026, and the piece argues Marvell could multibag relative to Nvidia as earnings growth accelerates over the next three years.

Analysis

The important shift is not that AI spend is slowing; it’s that economics are moving from general-purpose acceleration to workload-specific silicon. That transfers bargaining power away from the incumbent GPU supplier toward hyperscalers that can amortize design costs across enormous fleets, and toward chip architects with strong networking/optics attach. Near term, that is mostly a valuation and mix story: the market will pay less for a company whose growth is increasingly tied to a mature, highly visible platform, while rewarding names that can show incremental revenue from custom deployments and adjacent interconnect content.

The first-order winners are AMZN and MSFT if custom silicon lowers inference cost per query and improves gross margin on AI services; the second-order winners are the plumbing layers around the racks, especially network silicon and optical components. The biggest loser is not necessarily NVDA earnings today, but its multiple if investors conclude the long-duration growth slope is flattening. However, that thesis breaks if training demand re-accelerates, if new models remain GPU-intensive for longer than expected, or if custom programs stay concentrated and lumpy enough that only a few hyperscalers capture the savings.

Contrarian view: the crowd may be underestimating how much of the “custom AI” opportunity is just internal capex optimization, not incremental market expansion. That limits upside for pure-play beneficiaries if total AI budgets plateau. Over 1-3 months, the key catalyst is guidance from hyperscalers and semis on production ramps and 2027 capex plans; over 6-18 months, the question is whether custom silicon meaningfully compresses NVDA’s share of wallet or merely adds another layer of spending. Falsifier: if NVDA data-center growth re-accelerates and gross margins hold while MRVL commentary shows long lead times or customer concentration risk, the relative-value bearish NVDA trade should be reduced.

Key Decisions for Investors

• Initiate a modest long MRVL / short NVDA pair trade into the next earnings cycle; use it as a relative-value expression of custom-silicon share shift, not a blanket call that NVDA’s franchise is impaired. Falsify if NVDA data-center growth and margins re-accelerate faster than expected.
• Overweight AMZN and MSFT on a 6-18 month horizon as the main economic beneficiaries of lower inference cost; the trade is about capex efficiency and margin expansion, not headline AI enthusiasm.
• Avoid chasing MRVL after a large run-up; prefer buying only on post-earnings volatility or a 10-15% pullback, since the market may already be capitalizing much of the custom-silicon optimism.
• Watch SOXX and optical/networking proxies for follow-through; if hyperscaler capex commentary supports wider AI infrastructure spend, a basket long in semis/interconnects is cleaner than a single-name bet.