Custom AI Chips Are Coming for Nvidia's Crown. Here Are 2 Companies Quietly Cashing In.

Broadcom's AI chip and networking revenue jumped 143% to $10.8 billion in its fiscal Q2 2026, with the company booking more than $30 billion in AI orders and guiding AI revenue to more than triple year over year next quarter. Marvell also reported strong AI-driven demand, with revenue up 28% to $2.4 billion and management targeting over $10 billion in custom silicon revenue by fiscal 2029. The article is constructive on both names but notes valuation risk, with Broadcom trading in the low 60s P/E range and Marvell near 100.

Analysis

The incremental winner is not “AI semis” broadly but the suppliers that sit one layer above the accelerator market: the co-design houses and the networking stack that becomes more valuable as clouds diversify away from Nvidia. That creates a subtle second-order effect: every hyperscaler that insists on a bespoke XPU also expands demand for adjacent silicon, packaging, validation, and high-speed interconnect, so the pie grows even if Nvidia loses share at the margin. In practice, this favors AVGO and MRVL over generic GPU exposure, while also pulling through more custom demand for foundry capacity and advanced substrate suppliers over the next 12-24 months.

The market is correctly rewarding growth, but it is likely underpricing customer concentration as a sequence risk rather than a static risk. When a small set of cloud buyers controls the order book, the revenue profile can look annuity-like right up until one design cycle slips, a hyperscaler standardizes internally, or capex budgets pause for a quarter; then the multiple compresses faster than fundamentals deteriorate. That matters more for MRVL, where expectations embed a cleaner pathway to scale, and less for AVGO, whose diversified software/semis mix gives it more downside absorption if AI spend normalizes.

The contrarian read is that the real competitive threat to Nvidia is not a direct replacement but the commoditization of “good enough” inference silicon at the margin. If custom chips continue to proliferate, the industry could shift from a winner-take-most model to a portfolio model where hyperscalers optimize per workload, which caps Nvidia’s pricing power even if unit demand remains strong. The market may be overestimating how fast that transition converts into margin-rich profits for the designers themselves, since the economic surplus can be competed away by customers who have more leverage than the headline growth suggests.

Key Decisions for Investors

• Long AVGO / short NVDA as a 6-12 month relative-value trade: AVGO has the cleaner diversified earnings base and better downside buffer if AI capex pauses, while NVDA faces increasing pricing/mix pressure from custom silicon adoption.
• Buy MRVL only on weakness or after a post-inclusion pullback; treat it as a 3-6 month momentum name with high execution risk. Upside remains meaningful if the custom-silicon cycle extends, but the ~100x multiple leaves little room for a design win miss.
• Use call spreads instead of outright longs for AVGO into the next earnings cycle: 3-6 month AVGO bull call spreads can capture continued AI order strength while limiting valuation compression risk.
• Pair long AVGO against short a high-beta AI hardware basket on any further 5-10% rally: the trade is a hedge against multiple de-rating if the market starts distinguishing between revenue growth and durable margin capture.
• Add AMZN as a secondary beneficiary via cloud-specific custom silicon leverage, but size modestly; this is a 12-18 month thesis and the upside is indirect, so it should be funded as a lower-volatility expression of the same trend.