Not Broadcom. Not AMD. Nvidia's Biggest Threat Continues to Be Something Near and Dear to Its Heart.

Nvidia reported fiscal Q2 data-center revenue of $89.0B, up 117% year over year and above the $85.8B estimate, while total revenue of $96.2B beat consensus and Q3 revenue guidance was $108.0B plus or minus 2%. The article argues that AMD and Broadcom are not material threats to Nvidia's general-purpose AI compute leadership, but identifies Nvidia's largest customers—Amazon, Microsoft, and Meta—as its primary longer-term risk. Their expanding in-house AI-chip programs could reduce Nvidia's data-center footprint, ease hardware shortages, and pressure the company's pricing power and gross margins.

Analysis

The relevant risk is not that hyperscaler silicon displaces Nvidia at the frontier, but that it captures the lower-return inference and internally standardized workloads that determine fleet utilization. Even modest substitution matters because Nvidia's valuation embeds sustained mix and gross-margin durability; a 300-500bp structural margin reset from lower scarcity rents would have a larger EPS/multiple effect than a few points of GPU share loss. This is principally a 6-18 month risk, since internal ASIC ramps require software maturity, packaging capacity, and proven performance-per-watt at scale.

AMZN, MSFT, and META gain bargaining leverage before their chips become material volume substitutes. The near-term economic benefit is likely lower cost-per-token and reduced supply-constrained capex, but this could be reinvested into more compute rather than released as free cash flow; do not assume custom silicon automatically lifts hyperscaler margins. AVGO is the cleaner second-order beneficiary if custom programs scale, through ASIC design/IP and networking content, while TSM and advanced packaging capacity remain the practical bottlenecks.

Consensus likely overstates the binary "customer becomes competitor" framing. Proprietary accelerators are optimized for narrow workloads, whereas Nvidia's CUDA ecosystem, networking stack, and deployment velocity preserve a premium in training and heterogeneous enterprise demand. The falsifier for a bearish Nvidia margin thesis is continued data-center gross-margin resilience alongside accelerating Blackwell/Rubin system sell-through; conversely, watch hyperscaler disclosures for declining Nvidia capex mix, falling accelerator lead times, or explicit cost-per-inference improvements from internal chips over the next two earnings cycles.

Key Decisions for Investors

• Maintain a 6-12 month relative-value position: long AVGO / short NVDA in equal dollar beta-adjusted exposure. The trade monetizes custom-silicon adoption without requiring an AI-capex slowdown; reassess if NVDA sustains data-center gross margin above 75% and guides supply materially ahead of demand for two consecutive quarters.
• Do not short NVDA outright into the next earnings print: near-term shipment visibility and platform lock-in can overwhelm margin concerns. Use a 3-6 month put spread only after a post-results strength move if implied volatility is favorable; target downside is a valuation de-rating on gross-margin guidance, with risk capped to premium.
• Add AMZN and META only selectively on evidence that internal inference silicon reduces unit costs without accelerating total capex. Monitor quarterly capex guidance and useful-life assumptions; rising capex despite internal-chip deployment means savings are being competed away into incremental compute.
• Set an alert around AVGO custom-ASIC revenue guidance and TSM advanced-packaging commentary. A confirmed acceleration in both is the investable validation of the substitution thesis; absent that evidence, treat customer-silicon narratives as a medium-term watch item rather than a catalyst.