Chip Stocks Are Tumbling. Is It Time to Sell Cerebras, Broadcom, and Nvidia?

The article argues that recent sell-offs in Nvidia, Broadcom, and Cerebras are largely valuation-driven rather than a deterioration in fundamentals. Cerebras is down about 18% since its May 15 IPO despite OpenAI’s planned $20 billion spend and Amazon AWS integration, while Broadcom reported $10.8 billion in AI semiconductor revenue, up 143%, and still guides to $100 billion in AI chip sales for the year. Nvidia remains dominant with 88% data center GPU share and trades at about 30x earnings, though competition and AI-spending sustainability remain key risks.

Analysis

The market is still pricing AI as if only one chip architecture can win, but the more important takeaway is that the accelerator stack is fragmenting by workload. Training remains GPU-dominated, but inference is where hyperscalers can optimize for cost per token, latency, and power, which is exactly where custom ASICs and wafer-scale designs can steal share. That means the next leg of AI capex may not be linear for NVDA; the mix shifts toward lower-margin, customer-specific silicon as buyers try to de-risk dependence on a single vendor.

AVGO looks best positioned for the second-order benefit of this transition because it monetizes the “build-it-for-me” trend without needing every new model to run on its own platform. The recent reset matters less on the headline multiple than on positioning: if AI guidance stopped surprising upward, the stock can de-rate while earnings still compound. The real risk is not demand decay, but expectation compression—if hyperscaler capex growth normalizes over the next 2-4 quarters, AVGO likely trades more like a high-quality cash compounder than an AI scarcity asset.

Cerebras is the most speculative expression of the inference trade. The question is not whether its architecture is clever, but whether a few anchor customers can turn a technical edge into repeatable unit economics before the market loses patience; that typically takes 12-24 months, not weeks. The contrarian angle is that the recent pullback may be a better entry than the IPO pop suggested, but only as a small, venture-like position because the path to monetization is still binary and capital intensity is high.

NVDA remains the cleanest long-duration winner, but the consensus may be underestimating how much of its dominance is already reflected in the stock’s relative de-risking versus peers. If AI infrastructure spend pauses, NVDA should hold up better than the market thinks because installed ecosystem lock-in and software adjacency protect share; however, upside from here likely requires either a new demand wave outside hyperscale or a reacceleration in inference deployment. In the near term, the better trade is not a blanket long semis bet, but a barbell between quality cash flow and optionality.

Key Decisions for Investors

• Long AVGO / short a basket of higher-multiple AI-adjacent semis over 1-3 months: the setup favors a quality-vs-story rotation if hyperscaler capex guidance stops accelerating; target ~10-15% relative outperformance if AI enthusiasm cools.
• Add to NVDA on 5-10% pullbacks only, with a 6-12 month horizon: the downside is more about multiple compression than fundamental deterioration, while the upside remains tied to ecosystem lock-in and eventual non-data-center AI demand.
• Initiate a small speculative long in CBRS-sized exposure only after confirmation of additional design wins or revenue conversion evidence over the next 1-2 quarters; treat as venture-style optionality with a predefined max loss, not a core position.
• Sell AVGO upside via call spreads into any post-earnings strength over the next 30-60 days: the stock can continue compounding, but the asymmetry has shifted from upside surprise to valuation normalization after a rich run.
• If you want a cleaner pair, long AVGO / short AMD for the next 2-4 quarters: AVGO’s customer diversification and margin profile give it better downside support, while AMD remains more exposed to being the alternate choice rather than the default standard.