The Glaring Reason Microsoft Is Falling Behind Alphabet and Amazon

Amazon and Alphabet are pulling ahead in AI data-center efficiency through custom chips, with Alphabet saying TPU 8t delivers 3x the training power of the prior generation and TPU 8i offers 80% better inferencing performance per dollar. Microsoft is still behind in custom silicon, but its Maia 200 is live in two data centers and improving tokens per dollar by 30%, while the stock trades at 24.4x forward earnings versus 34.2x for Amazon and 34.9x for Alphabet. The article argues Microsoft’s valuation discount makes it a compelling AI buy despite near-term capex pressure and weaker positioning in custom chips.

Analysis

The real earnings power shift is not “more AI spend” but vertical integration of the data-center stack. The companies with in-house silicon are converting capex from a pure cost center into a controllable margin lever: better tokens-per-dollar, lower vendor lock-in, and the ability to monetize surplus compute through third parties. That creates a flywheel for AMZN and GOOGL, while MSFT is still paying an external-tax premium during the transition period.

The second-order winner is Broadcom, not Nvidia. Every hyperscaler that adopts custom accelerators still needs interconnect, packaging, networking, and design support, which should keep AVGO’s content per dollar of AI capex high even as GPU share mix moderates. By contrast, NVDA’s near-term risk is not a demand collapse but a slower marginal share-of-wallet gain as hyperscalers increasingly split workloads between proprietary silicon and general-purpose GPUs.

The market is likely underestimating timing asymmetry. AMZN and GOOGL should see operating leverage improve over the next 2-4 quarters as custom silicon scales across more workloads, while MSFT’s payback is delayed by still-ramping internal chips and heavier exposure to OpenAI platform uncertainty. The discount on MSFT may be justified for now, but if Maia/Cobalt adoption accelerates, the re-rating could be sharp because the stock already embeds a lot of execution skepticism.

Contrarian view: the consensus may be too complacent on capex intensity. If demand growth decelerates even modestly, the companies with the heaviest 2026 build plans will face the most FCF compression, and the market will punish any hint that AI spend is front-loading future returns rather than accelerating them. That means the next catalyst is not another chip announcement; it’s utilization and monetization metrics over the next two reporting cycles.

Key Decisions for Investors

• Long AMZN vs MSFT for 3-6 months: buy the cloud operator with the clearest path to AI unit-cost deflation and pair against the one still in transition; target a 10-15% relative move if custom silicon adoption continues to show up in margins.
• Long GOOGL, preferably on pullbacks after earnings-driven volatility: custom TPU economics plus external customer deployment is a cleaner operating leverage story than MSFT’s current AI stack; risk/reward improves if the stock remains discounted to secular growth quality.
• Maintain/overweight AVGO on weakness: it is the toll collector on the custom-silicon buildout even as hyperscalers internalize more compute; use 1-2 quarter horizon with downside limited by diversified networking and design exposure.
• Reduce or hedge NVDA into strength with a partial covered-call structure over the next 1-3 months: not a structural short, but a relative-share-risk hedge against hyperscaler silicon substitution pressure.
• For MSFT, prefer a staged entry only after evidence of Maia/Cobalt utilization broadening: buy on 10%+ drawdowns, but keep position size smaller until capex intensity starts translating into visible opex leverage.