Andy Jassy Said Amazon Isn't Investing $200 Billion in AI "On a Hunch." Amazon's Own Trainium Chip Business Is Already Running at a $20 Billion Annual Pace.

The market should read this less as a capex warning and more as evidence that Amazon is converting AI infrastructure into a vertically integrated margin engine. The key mechanism is not revenue growth alone; it is the combination of contracted demand plus in-house silicon lowering AWS’s unit cost curve, which can translate into higher operating leverage once depreciation rolls through and utilization rises. In that framing, the stock’s medium-term upside is tied to AWS margin resilience in 2027-2028, not the optics of a bigger 2026 investment budget.

Second-order effects favor AMZN relative to the broader AI stack. If Trainium continues to take even modest share in high-volume training/inference workloads, it pressures Nvidia’s pricing power in hyperscale accounts and forces other clouds to accelerate their own custom-chip roadmaps. That does not kill NVDA’s growth case, but it can compress the incremental gross margin on the most strategic customer cohort, especially if hyperscalers become more disciplined buyers rather than pure GPU renters.

The contrarian risk is that investors may overestimate how quickly this converts to cash. The payback window is long, so the next 1-3 quarters still carry FCF and depreciation risk; if AI utilization softens or customer ramps slip, the market could punish AMZN for front-loading spend before the economics are visibly accretive. Falsifiers: AWS growth re-accelerating without margin erosion, or Trainium backlog conversion stalling below expectations; for NVDA, any evidence that cloud mix remains fully intact despite custom silicon adoption.