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

Amazon expects to spend about $220 billion in 2026 capex and Alphabet up to $205 billion, with both companies citing AI-cloud demand that exceeds available capacity. Amazon says server and networking investments reach breakeven in under three years, backed by a $496 billion backlog and AI capacity contracted for at least five years; Alphabet reports server payback in under two years and roughly one year using TPUs, with a $514 billion backlog. Amazon's North American e-commerce revenue grew 16% last quarter while segment operating income rose 21% to $7.5 billion, while Alphabet cites TPU-driven AI cost advantages and Gemini-supported search-query growth.

Analysis

The investable implication is not headline capex volume but whether incremental capacity converts into cloud-margin durability before depreciation catches up. Alphabet appears better positioned to defend AI pricing while preserving gross margin because vertical integration lowers its cost floor; this creates a potential share-gain loop against MSFT Azure and AWS in price-sensitive inference workloads. The offset is that lower customer pricing can expand demand without expanding absolute dollars per workload, so revenue mix—not claimed performance benchmarks—will determine equity upside.

For AMZN, the more material earnings sensitivity may be the combination of AWS utilization and retail fixed-cost absorption: sustained AI demand supports cloud returns, while automation can lift North America margins independently of cloud. However, the reported payback periods are management metrics rather than disclosed project-level cash returns; they are highly sensitive to accelerator utilization, power availability, depreciation-life assumptions, customer contract cancellation rights, and the eventual mix of lower-margin inference capacity.

Custom accelerators are a structural competitive issue for NVDA, but not necessarily a near-term short catalyst. Hyperscaler silicon chiefly displaces the most standardized internal workloads; frontier training, ecosystem lock-in, and external enterprise demand can still support GPU spending. The cleaner second-order beneficiary is ARM: broader adoption of internally designed CPUs expands Arm architecture volumes even if individual hyperscalers capture most of the economic surplus.

Consensus may be too focused on a binary "AI capex bubble" framing. The relevant 6-18 month risk is a return-on-capital dispersion: operators with proprietary silicon, power access, and monetizable distribution can sustain investment, while less-integrated AI infrastructure buyers face pricing pressure and multiple compression. Near term, this article alone is not a catalyst; the next earnings disclosures on cloud revenue growth, segment margin, depreciation, and committed-capacity conversion matter more.

Key Decisions for Investors

• Accumulate GOOG over a 3-6 month horizon, preferably on broad AI-capex drawdowns, contingent on Cloud revenue growth remaining above company-wide growth and Cloud operating margin holding or expanding. Thesis is custom-silicon-enabled margin defense plus AI-supported Search monetization; reassess if Cloud margin declines for two consecutive quarters despite revenue growth.
• Maintain AMZN as a core long but do not add solely on reported infrastructure economics. Add after evidence that AWS growth re-accelerates while consolidated capex-to-revenue stabilizes; upside comes from simultaneous AWS utilization and retail-margin leverage, while falsification is AWS deceleration with rising depreciation and weaker North America operating margin.
• Establish a modest 6-12 month long ARM / short NVDA relative-value position only if hyperscaler disclosures show rising internal CPU/accelerator mix and NVDA data-center revenue growth decelerates materially. This expresses architecture proliferation versus merchant-accelerator concentration; stop out if NVDA maintains growth and gross margin while ARM royalty growth fails to accelerate.
• Monitor MSFT, ORCL and CoreWeave-related credit/equity proxies for the negative read-through: if AI capacity pricing falls while their capex commitments rise, they lack the same degree of silicon and consumer-ad monetization offsets. Treat this as a watchlist rather than a short until contract duration, utilization, and funding data confirm a widening return-on-capital gap.