Broadcom's AI Chip Revenue Grew 221% Last Quarter. A $1,000 Investment a Year Ago Would Be Worth This Much Today.

Broadcom reported fiscal Q3 AI semiconductor revenue of $16.7 billion, up 221% year over year, and guided for fiscal Q4 AI revenue of $21.7 billion, up 236%. Management forecasts AI chip revenue reaching $58 billion in fiscal 2026, $115 billion in fiscal 2027, and $230 billion in fiscal 2028, supporting an approximately $30 EPS target and an implied $135 billion in free cash flow. Shares were muted after total Q4 revenue guidance of $34.8 billion trailed some expectations above $35 billion and memory costs pressured margins, but the stock trades near 12x fiscal 2028 earnings versus roughly 25x for leading chip peers. Key risks are customer concentration, memory supply constraints, and a potential slowdown or regulatory pushback in data-center construction.

Analysis

AVGO’s setup is less a simple AI-beta trade than a test of whether custom silicon can take inference workloads from a niche alternative to a material share-taker from NVDA’s merchant accelerator stack. If hyperscalers shift even a modest portion of mature, repeatable inference workloads internally, AVGO captures both accelerator content and the higher-value networking attach; NVDA’s unit growth can remain strong while its incremental mix and pricing power deteriorate. The relevant read-through is therefore negative at the margin for NVDA’s long-duration multiple, but potentially positive for MU and SK Hynix exposure as custom systems remain memory-intensive.

The valuation argument relies on a distant management target rather than independently underwritten earnings, so the key issue is conversion: AI revenue must translate into gross profit and FCF despite memory inflation, customer-specific engineering costs, and likely lower initial margins on large ASIC ramps. Over the next 1-3 months, shipment execution and any upward revision to total revenue—not just AI revenue—are the catalysts. Over 6-18 months, the market will require evidence that TPU/OpenAI/Meta programs diversify revenue rather than merely replace one concentrated hyperscaler dependency with another.

Consensus may be underweight the strategic value of AVGO’s customer co-design model: it creates switching costs after software, packaging, networking, and power architecture are validated. Conversely, the market may be correctly discounting the headline growth because buyers can delay custom-chip deployments if data-center power availability or AI monetization weakens; ASIC programs have lumpy acceptance cycles and less resale optionality than merchant GPUs. A sequential AI revenue miss, gross-margin decline beyond memory-cost pass-through, or reduced hyperscaler capex guidance would invalidate the near-term bullish thesis.

Key Decisions for Investors

• Initiate a 12-18 month long AVGO position on post-earnings consolidation rather than chase a headline breakout; target 20-30% upside from multiple normalization plus estimate revisions, with a thesis stop if the next two quarters show AI growth deceleration without offsetting margin expansion.
• Express custom-silicon share gain via a modest long AVGO / short NVDA pair, sized beta-neutral, over 6-12 months. The trade is designed to isolate ASIC/inference substitution; exit if NVDA sustains data-center revenue upside while AVGO’s AI backlog fails to convert into total-company guidance increases.
• Add MU as a second-order beneficiary only after confirming that higher memory pricing is being passed through by AVGO rather than compressing system economics. Watch AVGO gross margin and hyperscaler commentary; a persistent margin squeeze would favor memory suppliers near term but weaken the broader AI hardware demand signal.
• Set an event alert around GOOG, META, and OpenAI-related infrastructure disclosures over the next quarter: confirmed volume ramps or incremental custom-chip commitments justify increasing AVGO exposure, while capex restraint, power delays, or deployment deferrals argue for reducing it before fiscal 2027 expectations reset.