Better AI Chip Stock: Broadcom vs. Nvidia

Broadcom's AI semiconductor revenue surged 221% year over year to $16.7 billion in fiscal Q3, representing 56% of total revenue, and management forecasts AI chip revenue of $115 billion in fiscal 2027 and $230 billion in fiscal 2028. Custom accelerators, including OpenAI's Broadcom co-developed Jalapeño chip, could reduce Nvidia's exposure to some inference workloads while expanding Broadcom's addressable market. Nvidia remains highly resilient: Q2 revenue rose 106% to $96.2 billion, Data Center revenue increased 117% to $89 billion, and non-hyperscaler AI cloud, industrial, and enterprise revenue grew 138%; it also trades at a lower forward P/E of 14.4x versus Broadcom's 18.8x.

Analysis

The investable issue is not whether custom silicon displaces GPUs wholesale, but where the workload mix settles. ASICs win when model architectures and utilization are stable enough to amortize design/NRE costs across massive fleets; that favors a handful of hyperscalers and inference, while leaving NVDA structurally advantaged in training, rapidly changing models, sovereign AI, enterprises, and the software/tooling stack. AVGO therefore has greater concentration risk: each incremental AI program can be very large, but revenue visibility depends on a small number of customers converting internal chip roadmaps into deployed capacity.

Near term (days to 1-3 months), AVGO can outperform if management identifies additional deployed programs or raises AI backlog visibility, because the market will capitalize a perceived multi-year ASIC TAM. The key risk is that investor expectations are already embedding an unusually steep ramp; any evidence of tape-out delays, lower unit volumes, customer insourcing, or weaker networking attach rates would drive disproportionate multiple compression. For NVDA, the relevant counter-signal is not a custom-chip announcement but deterioration in hyperscaler capex, GPU lead times, or Data Center gross margin.

Contrarian view: custom accelerators may expand total AI infrastructure spending rather than simply reallocate it from NVDA. Lower inference cost drives more token consumption and new applications, increasing demand for networking, memory, power, and eventually training capacity. The cleaner second-order beneficiaries are ANET and MRVL, provided AI cluster build-outs remain broad; HBM suppliers MU and SK Hynix exposure is also more workload-agnostic than the AVGO-versus-NVDA narrative.

Key Decisions for Investors

• Initiate a 1-3 month relative-value position: long NVDA / short AVGO in equal dollar amounts after any AVGO post-guidance strength. Thesis is that NVDA's diversified demand pool and platform economics better absorb ASIC substitution; target 10-15% relative return. Stop if AVGO discloses two or more incremental large production ASIC customers or NVDA cuts Data Center gross-margin guidance.
• Maintain AVGO as a tactical long only into a verifiable backlog/customer-deployment catalyst, not on aggregate AI revenue targets alone. Size smaller than NVDA because customer concentration and program timing create binary quarterly outcomes; take profits on a 15-20% rerating without corroborating order visibility.
• Express the infrastructure-volume upside through long ANET versus short a broad semiconductor ETF (SOXX) over 6-12 months. Ethernet scaling benefits from both GPU and custom-silicon clusters, reducing exposure to the accelerator architecture winner; invalidate if hyperscaler capex guidance turns negative or AI-networking growth materially decelerates.
• Set alerts for hyperscaler capex revisions, NVDA Data Center gross margin, AVGO AI backlog conversion, and HBM pricing. A broad capex pause is the principal thesis-breaker for all AI infrastructure longs and would favor reducing beta rather than rotating between AVGO and NVDA.