2026 Agg. Capex
$725B
77% YoY Increase
Capex / OCF
1.21x
Surged from 0.80x in 2025
Cum. Capex (2031)
$7.6T
Modeled projection
Required 2036 FCF
$4.36T
Implied by current multiples
Executive Summary
- ↓ROIC is projected to compress sharply across all hyperscalers as the massive expansion of the invested capital denominator outpaces incremental after-tax profits.
- ↓Accounting adjustments extending server useful lives to 5.5–6 years are temporarily masking economic decay, deferring an estimated $176B earnings contraction.
- !Current market capitalizations (totaling $30.19T) demand a 27.4% annual compounding of free cash flow for nine consecutive years to avoid multiple compression.
- ↓Grid interconnection queues (5–8 years) and thermodynamics present physical limits, rendering significant portions of near-term capex stranded.
The Accounting Mirage: Depreciation vs. Economic Reality
- Between 2020 and 2024, a coordinated industry shift extended the estimated useful lives of cloud servers and network equipment from 3–4 years out to 5.5–6 years.
- This extension artificially inflates operating income (EBIT) and NOPAT without altering underlying cash flows.
- In 2023–2025, this maneuver added $3.7B to Microsoft's operating income, $3.2B to Amazon's, and $2.9B to Meta's.
- Hardware obsolescence cycles directly contradict these schedules: GPUs (H100, B200, Rubin) operate on accelerated 12-to-18-month cycles.
Management View: GPU Value Cascade
- •Years 1-2: Foundational model training requiring peak bandwidth.
- •Years 3-4: Repurposed for premium real-time inference (amortizing sunk capital).
- •Years 5-6: Relegated to batch inference, RAG, and utility workloads.
Quantitative Risk: Rapid Obsolescence
- •New architectures (B200, Rubin) offer 40x performance leaps over previous generations.
- •Data center space and power limits render older GPUs economically unviable to operate.
- •Premature hardware retirement triggers severe impairment charges.
| Hyperscaler | Historical Life | Claimed Life (2026) | Estimated Overstatement Risk (2028) |
|---|---|---|---|
| Amazon (AWS) | 3–4 Years | 5 Years (Rev. from 6) | Contributor to $176B sector-wide understatement |
| Microsoft (Azure) | 4 Years | 6 Years | Contributor to $176B sector-wide understatement |
| Alphabet (Google) | 3–4 Years | 6 Years | Contributor to $176B sector-wide understatement |
| Meta Platforms | 3 Years | 5.5 Years | 20.8% Earnings Overstatement |
| Oracle | 4–5 Years | 6 Years | 26.9% Earnings Overstatement |
Quantitative ROIC Decay and EVA Compression
- As cash is aggressively depleted to fund GPU procurement, the invested capital denominator is expanding rapidly.
- Derived WACC sits at 9.87% for MSFT, 10.17% for GOOGL, and 10.17% for AMZN.
- While AWS, Azure, and Google Cloud clear their cost of capital (2025–2030), the EVA spread is narrowing dramatically.
- By 2030, ROIC is modeled to fall roughly 19 percentage points at Microsoft and 21 points at Alphabet.
Marginal Capital Productivity = Δ NOPAT / Cumulative Capex
Capital Efficiency Collapse
Microsoft (2022–2025): Every $1.00 of capex generated $0.283 in new recurring after-tax profit.
Microsoft (2026–2030): Every $1.00 of capex generates $0.130 in new recurring after-tax profit.
| Hyperscaler | Historical ΔNOPAT/Capex | Projected (2026–2030) | Relative Degradation |
|---|---|---|---|
| Microsoft | 28.3% | 13.0% | -54.0% |
| Alphabet | 26.4% | 11.8% | -55.0% |
| Amazon | 13.2% | 11.6% | -12.0% |
Featured Infographic
Physical Bottlenecks: Power Economics & Thermal Constraints
- AI data centers require between 20 MW and 1 GW of electricity, acting as heavy industrial manufacturing plants.
- Power requirements shatter historical envelopes: traditional server racks consume 15–25 kW; B200 architectures approach 250 kW per rack.
- Grid interconnection queues in major corridors (e.g., PJM) stretch from 5 to 8 years.
- A three-year energization delay on a 100 MW campus incurs a $70M/year penalty in lost power allocation and capital carrying costs.
Legacy Air-Cooled Facilities
- •Cooling Efficiency (PUE): 1.35 to 1.50.
- •Wastes up to 33% of incoming grid power on fans and ambient chillers.
- •Incapable of sustaining silicon without catastrophic thermal failure.
Liquid-Cooled AI Campuses
- •Cooling Efficiency (PUE): 1.08 to 1.15.
- •Direct-to-chip liquid cooling required to prevent silicon throttling.
- •Reduces grid draw by ~14 MW on a 100 MW site (saving $10M/year overhead).
AI Workloads and Unit Economics
- Large language model inference is memory-bound (constrained by High Bandwidth Memory capacity), not purely compute-bound.
- The H200 features a 76% VRAM upgrade (141 GB HBM3e) over the H100, resulting in 25% to 60% higher token throughput.
- Despite a higher hourly rental rate ($28.72/hr vs $21.52/hr), the H200 costs ~7% less per million tokens due to throughput advantages.
- Utilization Risk: Cost per million tokens varies by up to 36.3x purely based on request rate. If utilization drops below 40%, fixed capex becomes a permanent balance sheet drag.
Reverse DCF Analysis: The Monetization Hurdle
- At Q3 2026, the core AI infrastructure complex market capitalization exceeded $30.19 Trillion (~40% of the S&P 500).
- A Reverse DCF indicates the market prices in a 27.4% annual compounding of free cash flow for nine consecutive years.
- This requires these companies to generate $4.36 Trillion in pure free cash flow by 2036 (representing ~94.5% of all current US corporate profits).
- The industry currently generates a mere $50B to $150B annually in AI-attributable revenue; the 2026 capex wave roughly doubles the depreciable asset base.
| Metric | MSFT | GOOGL | AMZN | META |
|---|---|---|---|---|
| Market Capitalization (Q3 '26) | $3.71T | $4.14T | $2.79T | ~$1.65T |
| Implied FCF CAGR | 22.1% | 24.3% | 21.8% | 19.5% |
| Base AI Revenue Gap (6-Yr D&A) | $145B/Yr | $162B/Yr | $120B/Yr | $75B/Yr |
| Burry AI Revenue Gap (3.5-Yr D&A) | $192B/Yr | $215B/Yr | $158B/Yr | $105B/Yr |
Institutional Credit Risk & The Macroeconomic Mandate
- Tech sector bond issuance reached 16.7% of global nonfinancial bond supply in 2025 to fund infrastructure with severe obsolescence risks.
- NBER research (w35290) models that assuming zero-NPV marginal investments, AI-sector productivity must multiply by exactly 2.7x just to justify current capex commitments.
- If the expected software licensing cash flows fail to materialize against fixed debt/lease obligations, the capital deepening becomes highly destructive.
| Macro Scenario (NBER) | Implied TFP Gain | Hyperscaler Consequence |
|---|---|---|
| Baseline (Zero-NPV) | 2.7x Multiplier | Avoids insolvency; WACC elevation & minor multiple contraction. |
| Transformative | 5.0x+ Multiplier | Generates $4.36T FCF required by Reverse DCF; justifies valuation. |
| Failure to Launch | < 1.0x (Standard) | Catastrophic capital misallocation; deep market crash risk. |