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

Broadcom reported fiscal Q2 2026 revenue of $22.2 billion, up 48%, with AI chip and networking revenue jumping 143% to $10.8 billion and more than $30 billion in AI orders booked. Marvell posted fiscal Q1 2027 revenue of $2.4 billion, up 28%, and expects its custom silicon business to exceed $10 billion in annual revenue by fiscal 2029. The article is constructive on AI-custom chip demand but notes both stocks already trade at elevated valuations, limiting near-term upside.

Analysis

The core second-order read is that AI capex is shifting from a single-vendor GPU stack toward a bifurcated architecture: frontier training remains Nvidia-led, but hyperscalers are increasingly monetizing inference economics through custom silicon. That matters because the next leg of spend is less about raw FLOPs and more about cost-per-token, power efficiency, and network topology—areas where Broadcom and Marvell get pulled into the value chain even when they are not the headline AI brand.

Broadcom looks like the cleaner exposure because its AI mix is already large enough to benefit from operating leverage while the rest of the franchise cushions any digestion phase in cloud capex. The key nuance is that its AI revenue is not just chips; it is also attached networking, which should prove stickier if customers continue multi-vendor hedging and cluster expansion. Marvell is more leveraged to the same theme but with a narrower customer base and a higher multiple, making it more vulnerable to any delay in design wins converting into sustained production ramps.

The market may be underpricing the margin geometry. Custom silicon can grow fast while still diluting consolidated margins versus software or mature networking, so the fundamental upside can coexist with multiple compression if investors extrapolate revenue rather than earnings power. Conversely, if hyperscalers prove that in-house accelerators lower inference costs enough to expand model usage, the supply chain beneficiaries are not just AVGO/MRVL—optical, packaging, and power-management names should see a second wave of demand.

The contrarian risk is that the current enthusiasm assumes uninterrupted customer concentration and spending cadence. If even one large cloud customer stretches qualification cycles or dual-sources more aggressively, both names could see a sharp sentiment reset within 1-2 quarters despite healthy headline bookings. Nvidia is the relative loser on mix, but not necessarily on absolute demand; the bigger risk to NVDA is valuation and margin expectations, not unit volume.

Key Decisions for Investors

• Long AVGO on pullbacks over the next 2-6 weeks; use the lower multiple versus MRVL to own the same custom-silicon trend with better downside protection. Target 15-20% upside if AI bookings convert into another quarter of upward guidance revisions; trim on any evidence of margin compression accelerating.
• Pair trade: long AVGO / short MRVL for the next 3-6 months. Rationale: AVGO has broader earnings support and lower valuation, while MRVL’s near-100x multiple leaves less room for any design-win delays. Risk is a sharp rerating if MRVL secures a major new hyperscaler ramp.
• Maintain a tactical underweight or hedged position in NVDA versus the hyperscaler ecosystem for 1-2 quarters. This is a relative-mix call, not a demand-collapse call: custom silicon adoption is a slow bleed on share, and the risk/reward favors owning the enablers of diversification rather than the incumbent whose premium already embeds dominance.
• Use downside calls or collars on MRVL into post-inclusion strength over the next 1-2 weeks. The stock has momentum and index-flow support, but the valuation leaves it exposed to any guidance hiccup; defined-risk structures make sense until the market digests the S&P event.
• Look for a secondary long basket in networking/power infrastructure over 3-12 months rather than chasing the headline chip names alone. The next-order winners should be the picks-and-shovels around custom ASIC clusters, where demand is less customer-concentrated and the valuation reset has been smaller.