Nvidia Started the AI Boom. These 2 Stocks Could Power the Next Phase.

Broadcom and Marvell are presented as controlling roughly 95% of the custom AI ASIC market, with custom ASICs projected to reach 27.8% of AI server compute in 2026 and grow 44.6% year over year. Broadcom highlighted long-term TPU and networking deals with Alphabet through 2031 plus expansions with Anthropic and Meta, while Marvell guided to up to $11 billion of AI ASIC revenue in 2026 and a pipeline of 50+ opportunities worth an estimated $75 billion in lifetime revenue. The piece argues hyperscalers are increasingly shifting to custom silicon, which should support continued upside for both names despite Broadcom's recent 15% post-earnings pullback and Marvell's sharply higher valuation after Jensen Huang's endorsement.

Analysis

The market is underestimating how quickly hyperscaler bargaining power turns into a supplier oligopoly. Once a custom chip design is embedded in a fleet, the switching costs are not just engineering-related; they become software, networking, validation, and supply-chain integration costs, which tends to extend revenue visibility far beyond a single tape-out cycle. That makes the more important margin story not gross margin expansion, but duration of demand and the right to sell adjacent sockets in optics, switches, and rack-level interconnect.

The second-order winner is not only the ASIC architect, but also the ecosystem that standardizes around the architect's networking stack. As AI clusters scale, the bottleneck shifts from raw compute to data movement, so suppliers with credible attach in switch silicon and optical interconnect should see mix improve faster than consensus models assume. That matters because it creates a levered earnings stream even if unit chip growth decelerates later in the cycle.

The main risk is that this narrative becomes self-defeating if hyperscalers successfully diversify supplier exposure or internalize more design work than expected. A near-term pullback in shares after strong positioning is usually the market's way of saying the order book is known, but the sustainment rate is not; the next catalyst is not another partnership headline, but evidence that 2027-2028 revenue converts from pipeline to shipped volume without margin dilution. Over the next 3-6 months, any sign of capex discipline at the big cloud names, or a pause in AI rack buildouts, would hit these names harder than model revisions imply.

Consensus is still treating this as a linear share-gain story, but the better framing is a platform-tax replacement cycle: customers are not avoiding spending, they are reallocating spend from merchant GPU rents into owned infrastructure economics. That is bullish for the custom-silicon winners, but it also means the market may be overpaying for the easiest part of the transition and underpricing the operational execution risk. The cleanest setup is to own the names with the longest contracted visibility and the broadest network attach, while being selective on valuation where upside already reflects the “next trillion-dollar” narrative.

Key Decisions for Investors

• Long MRVL on a 3-6 month horizon on any post-rally digestion; target a re-rate if pipeline converts into confirmed design wins, but trim if the stock starts discounting flawless 2027 execution.
• Maintain a core long AVGO / long MRVL basket, with AVGO as the lower-volatility compounder and MRVL as the higher-beta catch-up trade; use MRVL for upside capture and AVGO as downside ballast.
• Pair trade: long MRVL vs short a basket of merchant GPU exposure on valuation-sensitive strength, sized for 6-9 months; thesis is that custom silicon share gains can outgrow the market's current pricing more efficiently than broad AI beta.
• Buy AVGO calls into any weakness tied to quarterly noise, using 6-12 month maturities; the risk/reward improves when the market focuses on near-term guide quality rather than multi-year contract duration.
• Set a tactical hedge on GOOGL if hyperscaler capex commentary turns cautious; custom silicon is a long-duration winner, but the first-order revenue path still depends on cloud spend remaining elevated over the next 2-4 quarters.