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

Broadcom is presented as a potential AI winner as inference demand shifts from Nvidia’s GPUs toward custom ASICs, with Broadcom said to control about 70% of the ASIC market. The article highlights Broadcom’s AI chip sales rising 65% to $20 billion in fiscal 2025, representing 31% of revenue, and management/analysts forecasting $100 billion in AI chip sales by fiscal 2027. The piece is largely opinion-driven, but it frames Broadcom as attractively valued at 16x next year’s adjusted EBITDA and potentially better positioned for the next phase of AI spending.

Analysis

The market is likely underpricing the mix shift from training capex to inference opex. That matters because inference is not just a bigger TAM; it is a more repeatable, utilization-driven demand stream that should compress decision cycles and favor vendors with the lowest total cost per query rather than the best peak FLOPS. In that regime, the moat migrates from raw compute to system-level optimization, which is structurally better for custom silicon providers and the software/ecosystem vendors that can bundle chip, networking, and control plane into one procurement decision.

AVGO’s real edge is not simply share in ASICs, but its ability to capture budget that would otherwise fragment across GPU, networking, and integration layers. If hyperscalers keep pushing inference efficiency, Broadcom can take content even when unit volumes flatten because the chip becomes the center of a broader rack-level redesign. That creates a second-order winner set: optical interconnect, high-speed networking, and memory-adjacent suppliers should benefit as inference clusters are deployed closer to end users and scaled in smaller, more distributed footprints.

The main risk is that the inference thesis is consensual enough to be crowded but still early enough that revenue visibility can lag sentiment by several quarters. If Nvidia responds with more competitive inference economics, the “ASIC beats GPU” narrative could be delayed rather than disproven, especially for customers that value software compatibility and fast deployment over marginal cost savings. Another reversal vector is customer concentration: if a few hyperscalers slow custom silicon rollout after a budgeting reset, AVGO’s growth multiple could compress quickly despite long-dated secular demand.

The contrarian miss is that this may be less a zero-sum replacement of NVDA and more a portfolio optimization across workloads. Nvidia likely remains the standard for frontier training, while AVGO monetizes mature, high-volume inference; investors may be too eager to frame it as one winner. That argues for relative-value positioning rather than an outright anti-Nvidia bet.

Key Decisions for Investors

• Go long AVGO vs. short NVDA over the next 3-6 months: use a pair trade to isolate the inference re-rating; target 15-20% relative outperformance if hyperscaler custom-silicon commentary stays constructive.
• Buy AVGO on pullbacks below the post-earnings trendline with a 6-12 month horizon: asymmetric setup is favorable because valuation still leaves room for multiple expansion if AI mix keeps rising, but size modestly due to customer concentration risk.
• Own GOOGL and META as indirect inference beneficiaries over 6-12 months: they are best positioned to internalize custom silicon savings, so their ad/AI economics can expand even if headline chip vendors underwhelm.
• Avoid an outright short in NVDA; instead use covered calls or a call spread if you want to express skepticism over the next 1-2 quarters, since training demand and software lock-in can keep the stock resilient.
• Watch for confirmation in networking and interconnect suppliers as a follow-on trade; if inference ramps faster than expected, these names should inflect before the chip vendors do.