Amazon CEO Jassy says company could sell AI chips, raising stakes for Nvidia, AMD

AWS AI revenue has a run rate of more than $15B as of Q1 2026 and Amazon’s chip business reports a $20B annual run rate (which Jassy says would be roughly $50B if standalone); Trainium is sold out and Trainium3 is nearly fully subscribed. Amazon says it added 3.9 GW of capacity in 2025, plans to double that by 2027, and may sell racks of its custom chips to third parties—posing increased competition to Nvidia and AMD. The company plans $200B of capex in 2026, expects Trainium to save “tens of billions” in capex per year and deliver several hundred basis points of operating margin advantage, though shares are down ~7% since the announcement.

Analysis

Amazon’s move into vertically integrated AI silicon is best viewed as an inflection that accelerates a multi-year price-performance war at the inference layer rather than a one-off product launch. At rack and system level, hyperscale integration advantages (thermal, power, custom interconnects) can deliver effective TCO savings that are hard for standalone GPU vendors to match without conceding margin; expect material customer re-architecture efforts over 12–36 months as buyers re-evaluate total cost of ownership, not just chip FLOPS.

Second-order supply effects matter: foundry allocation, HBM/memory sourcing and advanced packaging capacity will be the choke points that determine who scales first. Suppliers with flexible wafer allocation and advanced packaging capabilities will capture disproportionate upside, while traditional server OEMs face disintermediation risk if hyperscalers push pre-racked, vertically optimized appliances to market partners.

Key risks and catalysts are timing and software lock-in. The dominant incumbent retains a software ecosystem and model-optimization lead that can blunt hardware-cost advantages in the short term; conversely, early third-party rack deals, factory allocations or published price-performance comparisons would be immediate catalysts. Regulatory and customer contract complications represent non-linear downside events over 6–24 months that could delay or limit commercialization upside.

Key Decisions for Investors

• Core long-AMZN: Buy AMZN equity or Jan-2028 LEAPS (buy calls) sized 3–5% portfolio as a 12–36 month asymmetric bet on margin uplift and new revenue streams; set a 20% stop-loss and take partial profits on 40–60% upside. Rationale: optionality on commercial silicon sales plus AWS margin expansion outweighs near-term headline capex volatility.
• Pair trade (market-neutral): Long AMZN / Short NVDA (equal dollar) for 6–12 months to capture a re-rating if hardware pricing compression favors vertically integrated hyperscalers. Hedge: reduce NVDA notional if NVDA announces a clear generational architecture advantage; initial position size 2–4% net exposure.
• Tactical hedge: Buy 9–12 month NVDA puts (small allocation, 1–2% portfolio) to protect against a rapid share reallocation in inference GPUs driven by hyperscaler commercial silicon. Keep premium risk small—this is insurance against a quick, sentiment-driven repricing.
• Supply-chain barbell: Add exposure to foundry/packaging leaders (e.g., TSM) via equity or call spreads for 12–24 months (2–3% allocation), and avoid or underweight traditional tower/server OEMs which face disintermediation risk from rack-level productization.