Three must-own stocks to play the second wave of AI

The article says AI capital is shifting away from Nvidia’s dominant 90% share of model-training GPUs toward inference and agentic AI deployment. That implies a broader second phase of AI investment, with opportunities potentially expanding beyond a single hardware winner. The piece is largely thematic commentary rather than a catalyst, so near-term price impact looks limited.

Analysis

The key shift is not a change in AI demand, but a change in who captures the economics: training was a one-time capex burst, while inference creates a recurring utilization stream that is harder to monopolize and more sensitive to deployment efficiency. That usually broadens the winner set from the dominant silicon supplier into networking, memory, power, cooling, and cloud orchestration layers, which is where the next incremental margins should migrate over the next 6-18 months. For NVDA, the risk is less outright demand loss than multiple compression as the market prices in slower share gains and more value capture moving downstream.

Second-order effects favor firms that can sell picks-and-shovels into every inference node rather than only top-end accelerators. As agentic workloads proliferate, the constraint shifts toward distributed compute density, interconnect latency, and energy per inference, which can support attach rates for networking and adjacent infrastructure even if GPU unit growth normalizes. The supply chain implication is that any bottleneck in HBM, advanced packaging, or power delivery becomes a near-term catalyst for earnings revisions across the AI stack within the next 1-2 quarters.

The contrarian read is that the market may be overconfident in a clean handoff from training to inference leadership: inference economics are real, but enterprise deployment is typically slower and more fragmented than model enthusiasm implies. If monetization lags, the current re-rating in AI infrastructure could overshoot before usage data catches up, especially if cloud customers optimize for lower-cost models and squeeze hardware intensity. That creates a window where NVDA can still grow, but not justify premium expansion unless agentic workloads visibly increase token throughput and capex intensity by mid-year.

Tail risk for the bull case is that inference shifts from proprietary high-margin stacks to cost-optimized, multi-vendor deployment faster than expected, compressing GPU pricing power. Tail risk for the bear case is that demand remains so strong that supply constraints re-accelerate pricing and extend NVDA's moat, making any short too early. The highest-probability reversal catalyst is evidence of slowing enterprise adoption or a meaningful reduction in hyperscaler AI capex growth, which would hit sentiment within days but fundamentals over several quarters.

Key Decisions for Investors

• Trim NVDA into strength over the next 2-4 weeks; keep a core long only if hyperscaler capex commentary remains accelerating, since the main risk is multiple compression rather than near-term revenue miss.
• Express the next leg of AI spend via a pair: long AVGO / short NVDA for a 3-6 month horizon, betting that inference monetization shifts value toward networking, custom silicon, and system integration with less valuation risk.
• Buy an AI infrastructure basket on pullbacks — e.g., long AVGO and SMCI with a tighter stop than NVDA — to capture broader deployment spend if inference ramps faster than expected over the next 1-2 quarters.
• For convexity, consider NVDA put spreads 2-4 months out funded by out-of-the-money calls only if you expect a capex slowdown headline; payoff is best if sentiment de-rates before earnings revisions do.
• Watch hyperscaler capex and inference usage metrics closely; if those flatten for one quarter, reduce NVDA exposure aggressively because the market can reprice the story in days even though the fundamental impact unfolds over months.