If You Only Own Nvidia for AI Exposure, You're Missing Half the Trade. Here's the Other Half.

Broadcom's fiscal 2025 AI-chip revenue rose 65% to $20 billion, representing 31% of sales, and management expects nearly sixfold growth to $115 billion by fiscal 2027. The article positions Broadcom's customized ASIC accelerators as a cost-efficient AI-inference complement to Nvidia's GPU-led AI-training franchise, with major customers including Meta, Google, OpenAI and Anthropic. Analysts project Broadcom revenue and EPS CAGRs of 62% and 77%, respectively, through fiscal 2028, while the stock trades at 22x next-year earnings versus Nvidia's 23x.

Analysis

The relevant investment question is not training versus inference in isolation, but whether hyperscaler AI budgets shift from merchant GPUs toward internally designed silicon quickly enough to alter NVDA's incremental gross-margin and pricing trajectory. AVGO is the principal monetization vehicle for that shift, but its custom-silicon economics come with materially higher customer concentration, program-timing risk, and lower portability than NVDA's CUDA-led platform. The likely second-order beneficiary is TSMC (TSM), which captures wafer content regardless of accelerator architecture; networking suppliers such as ANET may also benefit if inference clusters scale as distributed deployments rather than centralized training pods.

Near term, this is more likely a sentiment and estimate-revision debate than a clean share-gain event: custom chips require long design cycles, qualified packaging capacity, and customer-specific software optimization. Over 1-3 months, the catalyst is hyperscaler capex commentary and any disclosure of ASIC volumes, yields, or incremental customer programs; AVGO's upside requires backlog conversion rather than aspirational total-addressable-market math. Over 6-18 months, sustained inference growth could compress NVDA's premium multiple only if GPU pricing, gross margin, or attach rates show tangible deterioration.

Consensus may be over-extrapolating a binary substitution narrative. Large customers can deploy ASICs for stable, high-volume workloads while retaining NVDA for model development, rapidly changing workloads, and ecosystem compatibility; total accelerator spend can rise even as GPU share falls. The key falsifier for an AVGO-over-NVDA thesis is continued NVDA data-center revenue upside accompanied by stable gross margin and evidence that inference GPU demand remains supply-constrained, while AVGO program delays or a customer capex reset would expose the concentration discount.

Key Decisions for Investors

• Initiate a modest 3-6 month long AVGO / short NVDA dollar-neutral pair only after verifying relative estimate revisions and AI-segment backlog conversion; target 10-15% relative upside with a 7% stop on pair spread. This expresses custom-silicon mix shift while reducing broad AI-capex beta.
• Maintain core NVDA exposure rather than replacing it with AVGO; trim only if the next earnings cycle shows data-center gross margin pressure or a material reduction in forward demand visibility. A clean beat-and-raise with stable margins invalidates the near-term share-loss case.
• Add TSM as a lower-idiosyncratic AI-accelerator basket expression on weakness, subject to foundry utilization and advanced-packaging capacity data. It participates in both merchant GPU and custom-ASIC demand, though geopolitical risk warrants smaller sizing than AVGO or NVDA.
• Set alerts for META and GOOG capex guidance, accelerator deployment disclosures, and AVGO customer-program commentary over the next two reporting cycles. Do not underwrite aggressive AVGO earnings growth until disclosed program ramps, customer concentration, and margin mix support it.