Nvidia May Have Won the AI Training Race, but the Bigger Opportunity Is Likely Still Ahead

The article argues Nvidia's AI growth opportunity could extend beyond model training into inference, with AI agents potentially generating billions of more complex compute-intensive workloads. Nvidia is positioning its Vera Rubin GPU-and-CPU architecture for long-running inference tasks, but its ability to capture the opportunity is uncertain as Amazon and Alphabet develop lower-cost custom AI chips. The outlook is constructive for Nvidia's long-term AI infrastructure demand, while emphasizing market-share and pricing risks in inference.

Analysis

Inference growth is directionally bullish for AI compute, but it is not automatically incremental to NVDA earnings at training-era gross margins. Inference buyers optimize for cost per token, latency, power and utilization; this shifts bargaining power toward hyperscalers with proprietary silicon and large captive workloads. The key valuation variable over the next 1-3 quarters is therefore not aggregate inference demand, but whether NVDA can preserve accelerator ASPs and networking attach rates as inference mix rises.

AMZN and GOOG have asymmetric strategic upside even if their custom chips do not become merchant products: each unit of workload migrated from NVDA-based instances to Trainium/Inferentia or TPU improves cloud gross margin and reduces capex concentration. A broader inference boom also supports AMZN and GOOG cloud revenue, but creates a near-term capex/FCF tension that markets may punish if utilization and AI monetization lag infrastructure commitments. NFLX is a second-order beneficiary through content localization, recommendation and advertising productivity, though its compute spend is too small to be a primary AI-infrastructure expression.

Consensus likely overstates agentic inference as a linear demand multiplier. Agent workflows can increase token consumption, but model distillation, caching, routing to smaller models and software optimization can reduce compute per task faster than task volume expands. The clean falsifier for the bearish margin view is evidence that NVDA inference revenue and networking attach accelerate while gross margin remains stable despite a larger inference mix; conversely, hyperscaler disclosures of rising internally sourced AI capacity would signal share pressure before it is visible in NVDA reported revenue.

Immediate price impact should be limited because the article supplies no incremental order, utilization or pricing data. Over 6-18 months, inference is more likely to broaden the value pool across cloud platforms, networking and power infrastructure than to sustain a single-vendor hardware rent; monitor AWS/Google Cloud growth, capex-to-revenue ratios, NVDA data-center gross margin, and disclosed custom-silicon adoption at each earnings cycle.

Key Decisions for Investors

• Maintain NVDA as a core long only if sizing reflects margin-risk rather than pure demand exposure; add on post-earnings evidence of stable data-center gross margin and accelerating networking revenue. Reduce if gross margin guidance falls materially while management attributes it to product or customer mix, as that would validate inference commoditization risk.
• Initiate a 6-12 month relative-value basket: long AMZN and GOOG versus short a proportionate NVDA hedge. The thesis is that internal silicon captures cloud-margin upside while limiting dependence on merchant accelerators; exit if AWS and Google Cloud growth decelerate while capex remains elevated, or if NVDA reports continued share gains in hyperscaler inference deployments.
• Use a catalyst watch rather than an outright NFLX trade: add only if management quantifies AI-driven advertising yield, content-production savings, or margin uplift. Without disclosed economics, AI enthusiasm is unlikely to move the earnings power enough to justify a dedicated position.
• At each AMZN and GOOG report, track AI capex growth against cloud revenue acceleration and any disclosed custom-chip utilization. Rising capex without monetization is a 1-3 quarter risk to both longs; accelerating cloud growth plus improving operating margin would strengthen the pair trade.