The Depreciation Disconnect: Hyperscaler CapEx

01. Executive Summary: The "Phantom Earnings"

The current valuation premiums for Hyperscalers (Microsoft, Meta, Alphabet) implicitly assume that the massive surge in AI CapEx—projected to exceed $600B+ aggregate by 2026—will yield long-term ROIC comparable to traditional Cloud infrastructure.

This assumption is flawed. Our analysis identifies a structural mismatch between Accounting Useful Life (6 Years) and Economic Useful Life (~3 Years). By failing to recognize the rapid obsolescence of GPU clusters, Hyperscalers are creating a "Depreciation Shield" that we estimate inflates current EPS by 10-15%.

02. The Core Disconnect: Accounting vs. Physics

To support short-term earnings amidst rising CapEx, major players have extended their depreciation schedules. This accounting maneuver boosts reported Net Income but diverges from engineering reality.

03. Technical Obsolescence: The "3-Year Wall"

While a server may physically function for 6 years, an AI Training Cluster hits economic obsolescence much faster due to the "Performance-per-Watt" curve. Nvidia's accelerated 1-year roadmap renders older chips cost-prohibitive to power.

The Verdict: By 2027, training on 2023-vintage H100s will be effectively "scrap value" for low-end inference. The "Prime Economic Life" is 3 years, not 6.

04. Quantitative Analysis: The $15B Earnings Gap

We modeled the P&L impact of "Marking-to-Market" the depreciation schedule for a hypothetical Hyperscaler with $100B in AI Assets.

*Implication: For every $100B in AI assets, the company potentially overstates Pre-Tax Income by $15B annually.

05. Investment Verdict

We maintain a Short Bias on hardware-heavy Hyperscalers. The market is capitalizing these firms on P/E ratios that assume current margins are sustainable. As the 2023-2024 vintage of GPUs hits the "3-Year Wall" in 2026-2027, we expect a massive write-down cycle and margin compression.