Meet the Company That's Quintupled Its Share of AI Chip Shipments in 2 Years

Amazon’s custom AI chips business is now running at an annualized revenue rate above $20 billion and, according to CEO Andy Jassy, could generate $50 billion if sold as a standalone operation. Demand for Trainium3 is strong, with nearly all 2026 capacity sold, and Amazon says Trainium adoption should save tens of billions of capex dollars per year while improving AWS margins. The article highlights rising competition for Nvidia from hyperscaler-designed accelerators, which is positive for Amazon’s cloud economics and strategically important across the AI chip market.

Analysis

The key implication is not that custom silicon is ‘good for Amazon’ in a generic sense; it is that AWS is turning capex intensity into a moat. If Amazon can keep shifting internal workloads from merchant-priced GPUs to near-cost accelerators, the operating leverage on every incremental AI dollar is structurally better than peers who remain dependent on NVDA pricing power. That should widen the dispersion between cloud leaders with captive silicon roadmaps and everyone else, especially as inference becomes the larger share of AI workloads.

The second-order winner is not just AMZN equity holders but the entire AWS ecosystem: lower compute costs should preserve customer unit economics, encourage more model deployment, and improve retention of large enterprise workloads. The loser is NVDA at the margin, but more specifically the market’s assumption that every AI dollar has to transit through Nvidia at full gross margin. The risk is that the competition is self-defeating only if demand growth slows; if AI usage keeps compounding, custom chips can expand the pie while compressing the supplier’s take rate.

The timing matters: this is a 6-18 month earnings/cash flow story, not a next-week catalyst. Near term, headline capex can still pressure sentiment because investors see the spend before they see the margin benefit, so AMZN can remain misunderstood even as the economics improve. The contrarian miss is that custom chips are not just a cost-cutting tool; they are a strategic pricing weapon that could eventually force Nvidia to defend share with more flexible pricing or faster product cadence, which would compress industry margins more broadly.

The cleanest expression is long AMZN vs short NVDA on a 6-12 month horizon, because the market is likely still underestimating how much of AWS margin expansion is coming from chip substitution rather than simple scale. A secondary trade is long AMZN / short a basket of GPU-exposed hardware beneficiaries that rely on merchant AI capex, since the value capture is migrating inward to hyperscalers. For more convexity, buy AMZN call spreads into any post-capex selloff; the setup is asymmetric if the market is overreacting to near-term spend while underpricing 2026 margin expansion.

Key Decisions for Investors

• Go long AMZN vs short NVDA on a 6-12 month horizon; target an AWS margin re-rating as custom silicon adoption shifts AI economics away from merchant GPU spend, with risk centered on NVDA pricing cuts and AMZN capex disappointment.
• Buy AMZN 9-12 month call spreads on any 5-8% pullback tied to capex headlines; use this to express a delayed-margin-inflection view with defined downside and upside tied to 2026 earnings revisions.
• Underweight NVDA on a relative basis versus hyperscalers for the next 2-4 quarters; the thesis is not absolute demand weakness but take-rate compression as customers internalize more of the silicon stack.
• Long META and MSFT only as a second-order beneficiary basket versus NVDA, but avoid making them the primary expression; their custom silicon narratives are earlier and less monetizable than AMZN’s AWS-driven economics.