Marvell: Why Jensen Huang Wants A $1T Competitor

This is less about one company “winning” and more about the market repricing the value chain behind AI inference. If custom silicon keeps taking share, the marginal winner shifts from general-purpose accelerators to the companies that can monetize design wins, software integration, and hyperscaler-specific workloads; that favors MRVL and GOOGL while putting structural pressure on NVDA’s long-duration multiple. The second-order effect is that AI capex doesn’t disappear — it migrates toward a more fragmented hardware stack, which can expand total unit demand for networking, optics, and design services even if GPU attach rates soften.

The key market risk is that consensus may be underestimating how fast hyperscalers internalize economics once a custom ASIC clears a performance-per-watt hurdle. That typically creates a 12–24 month lag where the share shift looks incremental, then becomes nonlinear as procurement teams standardize around the lower-cost architecture. The flip side is that these programs are execution-sensitive: any yield issues, software friction, or model-shift toward more general workloads could quickly push customers back toward GPUs, making NVDA’s pullback vulnerable to a sharp squeeze if the replacement cycle stalls.

The overdone/underdone question: MRVL’s move may be justified if investors now assign it a durable role as the “picks-and-shovels” partner to hyperscaler silicon, but the market may be extrapolating too much of that into a straight-line growth story. Custom ASICs are usually high-quality wins but lumpy, and the real earnings leverage is often lower than the headline TAM suggests because pricing power accrues to the customer, not the supplier. GOOGL looks like the cleaner beneficiary because internal chip savings compound over years and can be redeployed into AI capex, search, and cloud margin expansion; NVDA’s risk is not collapse, but multiple compression as the market prices a slower share of the AI dollar.