Nvidia's Not Going to Be King for Life, So I Buy This Now

The article makes a bullish long-term case for Broadcom as a custom-AI-silicon beneficiary, citing fiscal Q3 2026 revenue of $29.59B (+85.5% YoY), AI semiconductor revenue of $16.70B (+221%), and $13.67B of free cash flow (46% of revenue). Management guided Q4 AI semiconductor revenue to $21.7B and projects roughly $230B in fiscal 2028 AI semiconductor revenue, supported by custom accelerators for customers including Google, Meta and OpenAI. Broadcom also reported $8.8B in Q3 software revenue (+29% YoY), targets more than $30 in fiscal 2028 EPS, and pays a $0.65 quarterly dividend after a 10% increase; key risks are concentrated hyperscaler demand and the potential for order cuts.

Analysis

The investable implication is not a zero-sum ASIC-versus-GPU shift; it is a change in who captures the system-level profit pool. AVGO can monetize both custom accelerators and the high-speed networking/content required to connect increasingly heterogeneous AI clusters, while NVDA retains the premium segment where software portability, rapid model iteration, and time-to-deployment matter more than unit cost. The more direct competitive read-through is negative for MRVL, whose custom-silicon narrative competes for the same hyperscaler design wins, while TSMC and HBM suppliers remain volume beneficiaries regardless of accelerator architecture.

The key near-term risk is that investors capitalize management's long-dated AI targets before independently verifiable customer commitments emerge. Custom chips have long design cycles and lumpy recognition; a single program slip, lower utilization at a major cloud customer, or an inability to translate chip shipments into profitable networking/software attach would compress AVGO's AI premium quickly. Over the next 1-3 months, track hyperscaler capex guidance, disclosed accelerator deployment volumes, and AVGO AI backlog/conversion commentary; over 6-18 months, the decisive metric is whether ASICs expand total AI infrastructure spend rather than merely displace GPU purchases.

Consensus likely overstates the benefit from NVDA supply constraints: constrained GPU supply can accelerate in-house silicon qualification, but it also gives NVDA pricing power and preserves its software moat for training. The contrarian risk to the bullish AVGO thesis is that frontier-model architectures change faster than ASIC design cycles, favoring programmable GPUs; this would show up in delayed customer ramps, weaker semiconductor gross margin, or customers emphasizing flexibility over inference cost. Treat the article's customer, revenue, and EPS assertions as management-linked claims until confirmed in filings and customer disclosures.

Key Decisions for Investors

• Initiate a 6-12 month AVGO/NVDA relative-value long/short only after validating ASIC backlog and customer concentration in reported filings; size modestly because both remain positively exposed to AI capex. Thesis target: AVGO outperforms if custom-silicon ramps convert into revenue and networking attach; exit if AVGO reports a material AI-ramp delay or NVDA materially raises supply availability.
• Maintain NVDA as the core AI compute exposure rather than replacing it outright with AVGO. Add selectively on evidence that hyperscaler training demand remains GPU-led; the falsifier is consecutive quarters of declining data-center gross margin or customer commentary showing meaningful workload migration away from CUDA-based training.
• Watch MRVL as the higher-beta negative read-through from confirmed AVGO hyperscaler ASIC wins. A short is not yet recommended without design-win and valuation data, but initiate an alert if AVGO identifies additional customers or raises custom-accelerator outlook while MRVL fails to raise its own data-center revenue trajectory.
• Use TSM and HBM supply-chain exposure as a lower-idiosyncratic way to express continued accelerator proliferation over 6-18 months. Reduce if leading-edge wafer utilization or HBM contract pricing weakens, as those signals would indicate AI capex digestion rather than a vendor-share rotation.