The AI Boom Is Entering A New Phase -- And Nvidia May Not Be The Biggest Winner

Broadcom’s AI chip sales rose 65% to $20 billion in fiscal 2025, and management expects that figure to reach $100 billion in fiscal 2027. The article argues the AI market is shifting from training to inference, where Broadcom’s custom ASICs and 70% market share could challenge Nvidia’s GPU dominance. Broadcom trades at 16x next year's adjusted EBITDA, framing it as a relatively attractive AI growth play.

Analysis

The market is likely underestimating how quickly inference spending can migrate from a capex story to an operating-efficiency story. That matters because the first wave of AI infrastructure spend was tolerance-driven, but the next wave is ROI-driven: hyperscalers will optimize for cost per token, latency, and utilization, which structurally favors custom silicon once workloads stabilize. If that transition holds, AVGO becomes the toll collector on a broader monetization cycle, while NVDA’s mix may shift more toward premium training plus software lock-in rather than pure chip volume.

The second-order winner may be the ecosystem around inference scale-out: advanced packaging, HBM, network interconnect, and power-management suppliers should see incremental demand even if GPU share peaks. However, the trade is not linear, because ASIC adoption can compress total silicon BOM cost and reduce the dollar intensity of compute buildouts over time. That means the bullish case for AVGO can coexist with a more muted medium-term revenue growth path for some adjacent names that depend on maximal-capex architectures.

The main risk is timing. Inference adoption can take 12-24 months to translate into durable revenue reallocation, and hyperscalers often dual-source to preserve negotiating leverage, which limits near-term share gains. A meaningful reversal would come if model architectures remain frontier-limited and training demand re-accelerates, or if ASIC designs underperform on flexibility and software integration, forcing customers back toward GPUs for evolving workloads.

Consensus may be too focused on who wins the chip socket and not enough on margin shape. AVGO’s valuation looks optically cheap because the market is still pricing it as a diversified semis/software compounder, but if AI-driven mix continues improving, the multiple can rerate before the full revenue target is achieved. The contrarian nuance is that NVDA is not necessarily the loser; its moat may simply shift from hardware scarcity to platform control, making the relative trade more about AVGO multiple expansion than NVDA collapse.

Key Decisions for Investors

• Long AVGO vs short NVDA as a 6-12 month relative-value trade; thesis is not outright GPU decay, but faster multiple expansion at AVGO as inference monetization becomes more visible. Risk: if training spend re-accelerates or ASIC ramps slip, the pair compresses.
• Buy AVGO on a 5-8% pullback or on confirmation of next hyperscaler capex guide; target 12-18 months. Use a trailing stop if the market starts discounting the $100B AI revenue path as aspirational rather than executable.
• Add a basket long in AI infrastructure enablers with exposure to inference scaling, funded by trimming high-beta semiconductor names most dependent on generalized GPU intensity. Best expression is 3-6 months out as the market reprices workload mix rather than raw AI enthusiasm.
• Use NVDA calls, not stock, to keep upside exposure while limiting downside if inference substitution pressures multiple expansion. Favor 6-9 month tenors; risk/reward improves if the market overreacts to ASIC headlines.
• Watch for any sign of hyperscaler dual-sourcing discipline breaking; if custom ASIC wins broaden beyond the current few customers, increase AVGO weight. If adoption remains concentrated, treat AVGO as a strong business with a capped near-term rerating until the customer base broadens.