Amazon: Anthropic's $110 Billion Commitment Changes The AI Capex Debate (Rating Reiterated)

Amazon's AI infrastructure capital expenditure is projected to reach $220B by 2026, supported by more than $225B in multiyear AWS and Trainium capacity commitments from Anthropic and OpenAI. The contracted demand reduces utilization risk while AWS posted 37% year-over-year Q2 revenue growth and a 39.4% operating margin. The Buy thesis is reinforced by AWS's ability to maintain high profitability while accelerating AI-related reinvestment.

Analysis

The key equity debate is shifting from whether hyperscalers will spend to whether incremental compute converts into durable, high-margin revenue before depreciation ramps. AMZN is better positioned than levered AI-infrastructure peers because retail and advertising cash flow can absorb a longer utilization ramp, limiting financing-driven multiple compression. The near-term upside is therefore less about revenue beats and more about evidence that incremental AWS capacity sustains segment-margin expansion despite depreciation and power costs.

The non-obvious read-through is mixed for semiconductor vendors. Greater adoption of Trainium improves AMZN's bargaining power against NVDA and can lower the cost of serving price-sensitive inference workloads; that is modestly negative for NVDA's long-duration pricing narrative, but supportive for AVGO, ANET and power/cooling suppliers if cluster build-outs remain broad. The critical diligence item is whether contracted capacity represents enforceable take-or-pay commitments, versus flexible consumption reservations or commitments economically supported by strategic investments; reported backlog, cancellation terms and customer concentration matter more than headline contract values.

Over the next 1-3 months, AWS backlog disclosure, management commentary on useful lives/depreciation, and any evidence of Trainium workloads displacing GPU instances are the catalysts. Over 6-18 months, the risk is that a small number of frontier-model customers optimize utilization or shift workloads across clouds, turning apparent demand visibility into a pricing concession cycle. The consensus likely underweights this concentration risk, but also overweights capex as a standalone negative: utilization and unit economics, not absolute spend, determine whether the multiple rerates.

Key Decisions for Investors

• Accumulate AMZN on a 5-8% market-driven pullback with a 6-12 month horizon; target a relative rerating versus mega-cap cloud peers if AWS margin holds while depreciation rises. Falsify if AWS growth decelerates for two consecutive quarters while segment margin falls more than 200 bps, indicating capacity is being monetized through price rather than mix.
• Express a relative-quality AI infrastructure view: long AMZN / short ORCL in equal dollar amounts over 3-6 months. AMZN has more internally funded capex capacity, while ORCL is more exposed to financing, customer-concentration and residual-value concerns; exit if ORCL demonstrates sustained cloud-margin expansion and materially reduces financing dependence.
• Do not add a directional NVDA short solely on custom-silicon headlines. Instead, monitor disclosed Trainium utilization and GPU-versus-custom-silicon pricing; consider a 6-month AMZN long / NVDA short pair only if AWS identifies material inference migration and NVDA's data-center gross-margin outlook weakens.
• Ahead of the next earnings release, use defined-risk AMZN call spreads rather than outright calls only if implied volatility is below its one-year earnings-event range. The trade requires confirmation that management provides backlog, utilization or depreciation guidance; absent that disclosure, treat the event as informational rather than a catalyst.