Forget the Capex Fears: Why Alphabet and Amazon Are Must-Buys

Amazon says server and networking investments reach breakeven in under three years, supported by AI capacity contracted for at least five years; it plans roughly $220B of capex this year and reports a $496B backlog, more than double year over year. Alphabet reports server payback in under two years—and roughly one year using TPUs—while planning up to $205B in annual capex and holding a $514B backlog including TPU system sales. Both companies argue that AI-cloud demand exceeds capacity and that custom chips improve investment returns, while Amazon's North American e-commerce operating income rose 21% to $7.5B on 16% revenue growth.

Analysis

The investable implication is not simply higher cloud revenue: disclosed infrastructure returns reduce the probability that hyperscaler AI capex becomes a prolonged FCF drag. AMZN and GOOGL can sustain elevated spend while preserving incremental cloud margins if contracted utilization remains high, supporting multiple resilience versus peers whose AI monetization is less visible. The key second-order loser is NVDA at the margin: rising internal accelerator adoption shifts a larger portion of future compute value from merchant silicon into hyperscaler-owned stacks, even if aggregate GPU demand remains robust near term.

Over the next 1-3 months, the market will focus on whether backlog converts into recognized cloud revenue quickly enough to validate current depreciation and capex trajectories. GOOGL has the cleaner upside asymmetry if TPU-led inference cost advantages translate into faster Cloud margin expansion and defend Search monetization; AMZN has greater operating leverage because retail efficiency gains can compound AWS-driven FCF recovery. Treat management payback claims cautiously: the relevant verification is reported Cloud operating margin, depreciation growth, lease-adjusted FCF, and evidence that contract terms contain meaningful minimum-commitment protections.

The contrarian risk is that capacity scarcity masks weak unit economics: customers may reserve compute ahead of actual deployment, leaving a utilization air pocket once supply catches up in 2027-28. A broader enterprise inference price war would also pass hardware savings through to customers rather than shareholders. Thesis failure would be indicated by decelerating cloud backlog/revenue conversion, Cloud margins failing to expand despite scale, or capex guidance rising while lease-adjusted FCF revisions turn negative.

Key Decisions for Investors

• Maintain/establish long GOOGL versus short NVDA over a 6-12 month horizon; the pair expresses value migration from merchant accelerators to proprietary hyperscaler silicon. Target 10-15% relative outperformance; exit if NVDA data-center growth reaccelerates while GOOGL Cloud margin fails to expand for two consecutive quarters.
• Accumulate AMZN on post-earnings or capex-driven weakness rather than chase momentum; 12-18 month thesis is AWS capacity conversion plus retail operating leverage. Risk/reward is attractive only if consensus AWS growth revisions rise without a corresponding deterioration in consolidated FCF; reduce on evidence of retail margin retrenchment or slowing AWS growth.
• Use quarterly earnings as a catalyst watch: require reported cloud growth and segment-margin improvement to validate either long. If both companies guide capex higher but fail to lift cloud revenue outlook, avoid adding—this would signal utilization or pricing leakage rather than productive scarcity.
• Avoid treating ARM as a direct beneficiary without TPU/CPU shipment and royalty disclosures. ARM gains only if custom Arm CPU deployment expands faster than the associated internalization pressure on external chip vendors; monitor as an alert, not a recommendation.