Andy Jassy is so bullish on Amazon's chips that he took a rare shot at Nvidia

Amazon's chip business is running at an annual revenue run rate of over $20 billion and CEO Andy Jassy says Trainium3 is 30-40% more price-performant than the prior generation. Jassy claims Trainium will save Amazon 'tens of billions' in capex annually at scale and deliver several hundred basis points of operating-margin advantage versus third-party chips; a significant chunk of Trainium4 capacity is already reserved ahead of general availability in ~18 months. Amazon continues to sell access to its chips to third parties (e.g., Anthropic, Uber) and signals potential future rack sales, while positioning AWS as both a partner to and competitor with NVIDIA. Stock performance: -1.2% YTD, +25% over the past year.

Analysis

Amazon’s push to internalize inference/training hardware is less a one-off product win and more a structural margin arbitrage that reorders vendor bargaining power. By controlling silicon, software and datacenter deployment, Amazon can convert variable third-party chip spend into internal capital investment and operating leverage — the real optionality is that a modest percentage shift in cloud compute mix can deliver several hundred basis points to AWS margins over a multiyear horizon.

Second-order winners and losers extend beyond obvious chip suppliers. ODMs and chipset ecosystems (foundry capacity, advanced packaging and memory suppliers) face demand reallocation; conversely, customers who run large-scale ML workloads gain negotiating leverage and optionality to migrate between stacks, accelerating commoditization of inference. The software / stack advantage remains a key moat for incumbent accelerators: absent near-term parity on frameworks, performance per watt and end-to-end developer productivity, displacement is multi-quarter to multi-year and dependent on defensible benchmarks and customer migrations.

Key catalysts to watch are public, reproducible benchmarks and major customer migration announcements, any aggressive pricing moves from entrenched suppliers, and disclosures around rack-level sales or OEM deals which change the TAM calculus. Tail risks include a renewed performance lead from incumbents, foundry constraints that retard Amazon’s rollout, or regulatory pushback if selling racks is construed as anti-competitive — each could materially reverse momentum within quarters rather than years.

Key Decisions for Investors

• Long AMZN (size: core overweight, horizon 9–12 months): buy a 9–12 month call spread (long delta ~0.35 / short higher strike to fund) to capture margin expansion optionality while capping cost. Rationale: asymmetric upside from operating-leverage improvement; downside limited to premium paid if macro multiple contracts.
• Pair trade — Long AMZN / Short NVDA (size ratio ~1:0.35, horizon 6–12 months): use a relative-weighted equity pair to express share-shift risk without gross market directional exposure. Risk management: tighten if NVDA posts >15% outperformance vs AMZN in 4 weeks or if AMZN reports guidance shortfall.
• Income trade on NVDA (horizon 4–8 weeks): sell a small OTM call spread to harvest premium anticipating transient re-pricing risk following competitive noise. Keep position size limited and overlay a protective long-dated put if conviction on downside increases — worst-case is assignment exposure into a continued rally.
• Tactical long UBER (size: satellite, horizon 3–6 months): accumulate calls or small equity position to play marginal cost relief on inference-intensive products. Upside is operational margin relief; downside tied to demand/ride volatility so keep as a low-conviction add-on position.