Semiconductor Sell-Off: 5 Stocks to Buy on the Dip

The article recommends Nvidia, Broadcom, AMD, Cerebras Systems and TSMC as potential buys during a semiconductor-sector pullback, arguing that strong cloud demand supports continued AI infrastructure investment despite concerns about its long-term economics. It cites Broadcom’s projection that custom-chip revenue will quadruple to $230 billion over the next two years, and notes Cerebras shares have fallen below their recent IPO price. The outlook is positive but depends on sustained AI demand and, for Cerebras, wider adoption of its systems.

Analysis

The key underwriting question is not whether AI compute demand is growing, but whether paid workload growth can keep pace with committed capacity. Hyperscaler spending is a stronger demand signal than neocloud financing, but neither proves attractive returns: utilization, customer monetization, and depreciation-adjusted returns matter. If financing tightens first at neoclouds, orders and sentiment could weaken before hyperscaler capex does.

The architecture shift is not purely zero-sum. Custom ASIC adoption can pressure merchant accelerator share or pricing, yet expand total silicon demand and benefit TSM across both designs. NVDA retains software and ecosystem advantages; AMD’s inference opportunity depends on software maturity and sustained deployments, not commitments alone. TSM is the broadest volume beneficiary, but also carries concentrated Taiwan-related tail risk. ASML can benefit from capacity expansion with a longer, less direct earnings transmission; cancellations or delayed fab plans would expose that lag.

Over days, sentiment and positioning can dominate fundamentals. Over 1–3 months, watch hyperscaler capex guidance, accelerator utilization, order conversion, and evidence that AI services generate revenue. Over 6–18 months, capacity additions may normalize lead times and weaken supplier pricing power even if unit demand rises. The article’s large company revenue and performance claims require verification against reported results and deployed systems; treat them as promotional until then. Consensus may underweight the distinction between cash-funded hyperscaler investment and debt-dependent neocloud expansion, while overestimating how quickly either becomes durable end-customer revenue.

Key Decisions for Investors

• Prefer a staged long in TSM over a broad basket of high-beta AI chip names: it has exposure across accelerator architectures. Add on sector weakness only while foundry utilization, advanced-node pricing, and customer capex plans remain intact. Falsify on sustained order deferrals or deterioration in those metrics; Taiwan risk remains a material gap-down exposure.
• For relative exposure, favor NVDA over CBRS as the core inference/accelerator position; keep CBRS on a small, event-driven watchlist rather than treating claimed speed advantages as proof of scalable economics. Require disclosed deployments, repeat orders, and evidence on system cost, power, and utilization before upgrading. Failure to convert its OpenAI relationship into deployments would weaken the upside case.
• Treat AVGO and AMD as conditional diversifiers, not automatic beneficiaries: monitor reported custom-chip revenue and customer concentration for AVGO, and software adoption plus repeat GPU shipments for AMD. If custom silicon gains share while total AI spending holds, AVGO may outperform NVDA; if spending pauses or commitments do not convert, both are exposed.
• Avoid buying the entire dip solely on valuation multiples cited against distant estimates. Reassess after the next hyperscaler capex and supplier earnings updates; a downward revision to spending, utilization, or order conversion is a thesis break, while stable guidance and improving workload monetization would support adding risk.