The 3-Stock Custom Silicon Basket That Could Outperform Nvidia by 2030

The article argues that AI custom silicon is becoming a major growth driver, with Broadcom’s ASIC sales doubling to $8.4 billion in Q1 and management forecasting AI revenue of $100 billion by next year. Marvell reported total sales up 42% to $8.2 billion, while TSMC posted first-quarter sales up 41% to $35 billion and net income up 58% to $3.49 per ADR. The overall message is constructive for Broadcom, Marvell, and TSMC as AI hardware demand broadens beyond Nvidia GPUs.

Analysis

The market is starting to separate AI compute into two layers: model training still needs brute-force GPUs, but inference and hyperscaler-specific workloads are migrating toward custom silicon. That is a structural margin transfer away from the most crowded part of the AI trade and toward the picks-and-shovels layer that captures design wins, wafer starts, and packaging demand regardless of who wins at the architecture level.

Broadcom and Marvell benefit because custom-chip economics are sticky once embedded in a hyperscaler roadmap; the switching costs are not just technical, but organizational and supply-chain wide. The second-order effect is that their growth may prove less cyclical than headline AI capex suggests, because these projects are typically multi-year and tied to cost-per-token reduction targets rather than experimental budgets.

TSMC is the cleanest expression of the trend because it monetizes both the incumbent GPU stack and the custom ASIC wave, while also gaining leverage from advanced-node scarcity and advanced packaging bottlenecks. The key risk is not demand collapse, but execution slippage: if custom chips underdeliver on power/performance economics, hyperscalers could temporarily reaccelerate GPU orders, which would favor Nvidia near term and compress relative upside for the custom silicon names.

The consensus may be underestimating how quickly hyperscalers will dual-source architecture to preserve bargaining power versus Nvidia. That argues for a relative-value trade, not an outright anti-Nvidia bet: the downside to NVDA is valuation and position crowding, while the upside to AVGO/MRVL/TSM is a long-duration re-rating if custom silicon captures even a modest share of AI inference growth over the next 12-24 months.

Key Decisions for Investors

• Long AVGO / short NVDA as a 6-12 month relative-value pair; thesis is multiple compression on the crowded incumbent versus continued estimate revisions in custom silicon. Stop if GPU supply tightness or stronger-than-expected NVDA guidance re-accelerates near-term momentum.
• Add to TSM on weakness for a 12-24 month horizon; it is the lowest-regret way to own the AI hardware cycle because it monetizes both GPUs and ASICs, with asymmetry from advanced-node and packaging scarcity.
• Initiate MRVL as a tactical long into any post-earnings consolidation; this is a higher-beta beneficiary if hyperscaler custom-chip deployments broaden, but keep sizing smaller given customer concentration risk.
• Use call spreads in AVGO rather than stock if chasing the theme; the probability-weighted upside is attractive over 6-9 months, but the stock can digest good news quickly after a strong run.
• Avoid outright shorting NVDA here; better expression is underweight or pair it versus the custom-silicon basket, because a temporary reacceleration in training demand could squeeze a naked short.