Nvidia vs. Broadcom: Whose AI Bull Case is Better?

Broadcom projects AI semiconductor revenue of $115 billion in FY2027 and $230 billion in FY2028, reflecting a further two years of rapid growth in its custom AI-chip business. Nvidia remains substantially larger, reporting $96.2 billion of Q2 revenue and guiding for $108 billion in Q3, with more than 90% tied to data-center operations. The article views both as attractive AI investments but favors Nvidia because of the broad applicability and continued dominance of its GPU platform, while acknowledging Broadcom's faster growth and ASIC market-share gains.

Analysis

The relevant investable issue is not GPU-versus-ASIC substitution in aggregate, but the split between training, inference, and captive hyperscaler demand. Custom silicon can take meaningful share in stable, high-volume inference workloads, yet it is unlikely to displace NVIDIA in frontier-model training or rapidly changing architectures, where software portability and time-to-deployment matter more than chip-level unit cost. This favors AVGO revenue growth but does not mechanically imply NVDA revenue loss; hyperscaler capex can support both while compute demand expands.

AVGO's risk is materially more concentrated than its headline AI growth rate suggests. A small number of large custom-chip programs likely drive most incremental AI semiconductor value, making design delays, customer capex pauses, or an internal ASIC roadmap change capable of producing abrupt estimate resets over the next 1-3 quarters. NVDA has broader customer/workload diversification, but faces a different risk: a sustained mix shift toward inference and internally designed accelerators could compress its long-term platform multiple even if absolute revenue continues rising.

The market may be underpricing the networking and memory bottleneck created by heterogeneous accelerator fleets. More ASIC deployment increases the need for high-bandwidth interconnect, optical components, and memory, supporting selective exposure to AVGO's networking franchise, ANET, MRVL and MU. Conversely, the simplistic view that each ASIC dollar is a dollar removed from NVIDIA misses that lower-cost inference can stimulate model deployment and total token consumption, potentially enlarging the accelerator market over 6-18 months.

Near term, this is low-novelty commentary rather than a standalone catalyst. The next decisive data points are hyperscaler capex guidance, disclosed custom-silicon production ramps, NVDA data-center gross margin, and evidence that ASICs are moving beyond captive workloads. A broad reduction in cloud capex or NVDA margin/guidance resilience despite ASIC ramp would respectively validate or falsify the substitution thesis.

Key Decisions for Investors

• Maintain NVDA as the core AI compute exposure over a 6-18 month horizon; do not reduce solely on ASIC-share headlines. Reassess if data-center gross margin declines by more than 300 bps sequentially alongside weaker forward revenue guidance, which would indicate pricing/mix pressure rather than normal product-transition noise.
• Add AVGO only on post-earnings volatility or verified customer-ramp disclosures, not on long-dated company targets alone. Size below NVDA because customer concentration creates asymmetric downside; require evidence of multiple production programs and AI semiconductor backlog conversion over the next 2-3 quarters.
• Express the heterogeneous-compute second-order theme via a basket long ANET/MRVL/MU rather than a binary NVDA short. Hold 3-9 months, with the thesis supported by improving Ethernet/optics orders and HBM pricing; exit if hyperscaler capex guides materially lower.
• For relative-value portfolios, consider long NVDA / short an equal-dollar broad semiconductor ETF (SOXX) rather than short AVGO. NVIDIA's software ecosystem and training exposure should outperform commodity analog and memory-heavy constituents if AI capex remains concentrated; stop out if NVDA underperforms SOXX by 10% after its next earnings report despite raised guidance.