Broadcom Builds Custom Chips for Google, Meta, Anthropic, and OpenAI. At 25 Times Forward Earnings, It's the Cheapest Mega-Cap AI Stock Nobody Talks About.

Broadcom’s AI chip sales surged 65% to $20 billion in fiscal 2025, and management expects at least $100 billion in AI chip revenue by fiscal 2027, implying AI will exceed 58% of projected revenue. Analysts expect revenue and EPS to grow at 53% and 66% CAGRs from fiscal 2025 to fiscal 2028, while the stock still trades at 25x next year’s earnings and 16x next year’s adjusted EBITDA. The article is broadly bullish on Broadcom’s long-term AI growth and valuation, though it is opinion-driven rather than a new company announcement.

Analysis

The market is still treating AI compute as a one-line Nvidia story, but the more durable trade may be the migration from general-purpose training silicon to customized inference stacks. That shift matters because once a hyperscaler standardizes on an ASIC design, switching costs move from hardware pricing to software, validation, and supply-chain co-design, which can lock in share for years and compress competitors’ pricing power. The second-order implication is that AI capex is becoming more segmented: training spend may remain cyclical, while inference capex becomes a higher-visibility, lower-volatility annuity.

The bigger tell is not the growth rate itself, but the implied mix shift in the ecosystem. If large model operators are optimizing for total cost per token, Broadcom’s architecture should increasingly win inside production deployments where latency and unit economics matter more than flexibility; that can pull demand away from GPU clusters at the margin without requiring a headline collapse in Nvidia demand. The risk for Nvidia is not near-term share loss across all workloads, but slower incremental wallet share in mature AI deployments, which can cap multiple expansion even if absolute revenue remains strong.

The valuation argument is more interesting than the headline multiple suggests because the market is still underpricing durability versus growth. If the AI revenue base becomes more recurring and bundled with networking/software relationships, the earnings stream should warrant a premium to cyclical semiconductor peers, not a discount to mega-cap software; that leaves room for upside if management guides to sustained customer concentration without order volatility. The main tail risk is customer insourcing: if the biggest hyperscalers overbuild their own designs and delay ramps after an initial rollout, the growth curve can cliff 12-18 months later, especially if AI utilization fails to keep pace with capex.

Key Decisions for Investors

• Go long AVGO on 3-6 month horizon on pullbacks, targeting a re-rating if management confirms inference-led backlog visibility; use a 12-15% downside stop because the main risk is hyperscaler timing slippage, not demand destruction.
• Pair trade: long AVGO / short NVDA for 6-12 months to express the view that production AI spend migrates toward custom inference silicon faster than the market expects; keep sizing moderate because NVDA remains the higher-quality growth franchise.
• Sell NVDA out-of-the-money calls or run a call spread against existing longs into strength over the next 1-2 quarters to hedge the risk of relative multiple compression as ASIC adoption rises.
• Add GOOGL and META as indirect beneficiaries of lower AI unit costs on a 12-month view, since improved inference economics should expand model deployment economics and support higher monetization without proportional capex growth.
• Avoid chasing NFLX here; it is not a direct beneficiary unless AI infrastructure cost savings materially lift content recommendation efficiency, so the risk/reward is weak relative to semiconductor exposure.