Amazon CEO takes aim at Nvidia, Intel, Starlink, more in annual shareholder letter

Key numbers: Amazon says its Trainium family has reached a ~$20B annual revenue run rate with Trainium3 capacity nearly sold out and Trainium4 (available ~18 months) also nearly sold out; Jassy estimates a hypothetical $50B ARR if Amazon sold chips externally. Amazon committed ~$200B in capex for 2026 (primarily AWS data centers) and cites an OpenAI pledge to spend $100B on AWS as an example of contracted demand, though Jassy acknowledges execution/verification risk. Operational highlights: Graviton CPUs are used by 98% of the top 1,000 EC2 customers and Amazon Leo (satellite service) has won multiple carrier and enterprise contracts, supporting upside to AWS but against the backdrop of Amazon’s share price weakness (<$200).

Analysis

Amazon’s move to internalize critical infrastructure (silicon, connectivity, robotics) is less about immediate revenue capture and more about changing margin and capacity dynamics across the cloud stack. If proprietary accelerators and CPUs take even a low-double-digit share of large-model inference and general-purpose server workloads over 12–24 months, the implied reallocation of fab capacity will create a choke point at foundries that benefits suppliers (capacity-constrained equipment/TSMC-like vendors) while mechanically capping the growth and pricing power of external GPU/CPU vendors.

A second-order effect is counterparty concentration: a handful of hyperscalers or model providers pre-committing to capacity means AWS’s capital intensity is de-risked on paper but concentrated in execution — missed spend or slower model training cadence from those customers could flip capacity from scarcity to stranded within 6–18 months, producing volatile utilization and the risk of asset write-downs. Regulatory risk is non-trivial; aggressive vertical integration invites antitrust scrutiny that can force non-discriminatory access or behavioral remedies on a 12–36 month timeline, which would change upside timing and upside magnitude.

Net-net, the fastest-moving opportunities are relative-positioning trades rather than binary calls on winners: favor exposure to the owner/operator that controls demand and distribution economics, pair that with protective hedges against an execution or regulatory reversal, and size for asymmetric outcomes — a rerating if integration succeeds vs a sharp re-pricing if capacity ends up underutilized or regulated away.

Key Decisions for Investors

• Go modestly long AMZN via a 12‑month call spread (allocate 3–5% of book): buy a near‑ATM 12m call and sell a +25–35% OTM 12m call to fund cost. Rationale: captures re‑rating if integration-driven margin accrual shows up in two sequential quarters; target 40–60% upside, stop and unwind on 20% drawdown in the spread or if regulatory action is announced.
• Hedge AWS exposure with a small NVDA bearish options position (buy a 9–12m put spread, size ~1% of book): designed to pay off if external GPU TAM growth decelerates. Reward limited but high-impact if GPU demand derails; cap loss to premium paid and reassess at 6‑month mark.
• Express conviction on x86 erosion with a protective INTC put spread (6–12m, size ~1%): buy puts financed by selling nearer OTM puts to limit cost. Expect material share shifts to show in server revenue mix within 12–18 months; keep position small given turnaround and IDM execution risk.
• Buy portfolio insurance: purchase a 3–6 month NASDAQ downside protection (ATM-to-slight-OTM puts sized to ~0.5–1% of book cost) ahead of expected AWS execution/earnings cadence. This caps tail risk from a sudden deleveraging of cloud capex expectations or regulatory shock that would re-rate tech multiples quickly.