⚠️ AI ROI Extension

Will the $1-3T AI Capex Wave Earn Its Cost of Capital?

Phase 8 Extension | November 2025

Wall Street's Burning Question

Hyperscalers are investing $2.8 trillion in AI infrastructure through 2029. Goldman Sachs warns of "too much spend, too little benefit." The IMF and Bank of England raise bubble concerns comparable to the 2000 dot-com crash.

This framework provides the first LP-based system to measure AI capex ROI from public filings, turning trillion-dollar narratives into testable, auditable economics.

Key Metrics

$2.8T projected AI infrastructure investment through 2029 (Citigroup)
55 GW of new power capacity required by 2030
10-15% typical GPU utilization (85%+ sits idle)
97% of enterprises struggle to demonstrate GenAI business value
3-5 years typical infrastructure payback (J-curve effect)
$600B revenue gap between current AI revenue and infrastructure requirements (Sequoia Capital)

What This Framework Measures

📊

Incremental ROIC on AI Assets

Calculate return on AI-specific capex $ROIC^{AI}$ using three attribution methods. Compare to WACC to determine value creation vs. destruction.

📝

AI Asset Ledger

Track $IC^{AI}$ (GPU servers, data centers, capitalized software) from 10-K footnotes. Roll-forwards respect conservation constraints.

⚡

Power Constraints

Validate if disclosed PPA capacity supports implied AI power demand. 55 GW gap by 2030 binds growth forecasts.

⏱️

Distributed Lag Model

Infrastructure takes 3-5 years to generate returns (J-curve). ARDL models capture time lag between capex and NOPAT.

🎯

Terminal Value Consistency

Physics-consistent EV/EBITDA multiples enforce $WACC > g$ and $g \leq ROIC$. Prevents internally inconsistent valuations.

💰

Cost Parity Analysis

Test Nvidia's "90% cost savings" claims. Shadow P&L validates accelerated computing ROI with utilization adjustments.

⚠️

Bubble Risk Scoring

Assess systemic risk via concentration, valuation stretch, and circular financing. Compare to 2000 dot-com bubble.

🔍

Quality-Adjusted Metrics

Apply $Q_t$ haircuts to AI cash flows. Penalize aggressive ROI claims from reporters with loose accounting discipline.

Theory Documentation

📘

AI ROI Framework

Complete mathematical framework: $ROIC^{AI}$ formulas, distributed lags, terminal value consistency, quality adjustments.

📒

AI Ledger Methodology

Building $IC^{AI}$ from 10-K footnotes. Roll-forward constraints, XBRL extraction, PPA tracking, estimation methods.

⚡

Power Constraints

The critical bottleneck: 55 GW gap, PPA extraction, utilization mysteries, regional constraints, onsite generation trends.

📐

Measurement Challenges

Data availability matrix, estimation uncertainty, the $600B question, utilization controversy, best practices.

🔬

Case Studies

Meta ad ROI (2.6-10.1%, sub-hurdle at lower bound), Azure AI revenue (15.2%, above-hurdle), AWS growth, Nvidia claims, bubble warnings.

📚

Bibliography

60+ sources: Academic papers (Acemoglu, Brynjolfsson), investment bank reports (GS, MS, Citi), central bank warnings, case studies.

Example: Microsoft FY2024 Analysis

Metric	Value	Status
AI Invested Capital $(IC^{AI})$	$111.0B (ending), $98.0B (average)	DISCLOSED
$\Delta$ NOPAT (AI-attributed)	$5.3B – $8.9B	ESTIMATED
$ROIC^{AI}$	5.4% – 9.1%	AMBIGUOUS
WACC	8.5%	CALCULATED
Implied EV/EBITDA $(g=6.5\%)$	11.7×	CONSISTENT
Analyst Multiple	15.0×	TOO HIGH (+3.3×)
PPA Capacity	0.8 GW	CONSTRAINED (104× gap)
Implied Power Demand	83.3 GW	INFEASIBLE

Key Finding: $ROIC^{AI}$ is ambiguous (5.4-9.1% straddles 8.5% WACC) in FY2024. Lower bound below hurdle but upper bound slightly above; J-curve effect suggests full returns by 2027-2028 if distributed lags play out. Critical constraint: power capacity 104× short of implied demand (0.8 GW disclosed vs. 83.3 GW required)—must expand PPAs or cap growth forecasts.

Key Innovations

vs. Traditional DCF

Conservation-Consistent Terminal Multiples: EV/EBITDA must respect $WACC > g$ and $g \leq ROIC$ (physics constraints)
Distributed Lag Model: Explicit 3-5 year payback timeline vs. assuming immediate returns
Power Feasibility Check: Growth forecasts capped by disclosed PPA capacity (55 GW gap by 2030)
Utilization Adjustment: Effective ROIC at 15% utilization is 6.7× lower than nominal
Quality Haircut: $Q_t$ penalizes aggressive AI claims from companies with sloppy accounting
Multi-Method NOPAT Attribution: Report bounds across trend break, bottom-up, and cost savings approaches

Wall Street Debate

Bears (IMF, BoE, Goldman Sachs)

Valuations "comparable to 2000 dot-com peak" (BoE Financial Stability Report, October 2024)
"$1T spend with little to show for it" (Goldman Sachs, June 2024)
Daron Acemoglu (MIT): Only 0.5% productivity gain over 10 years; 4.6% of tasks exposed to AI
Sequoia Capital: $600B revenue gap—current AI revenue ~$100B vs. $600B needed to justify infrastructure
Circular financing concerns (Nvidia invests in OpenAI → OpenAI buys Nvidia GPUs)

Bulls (Morgan Stanley, Tech Companies)

Profitability starts 2025 (34% margin projected, vs. dot-com unprofitability)
Azure AI $13B run rate (+175% YoY), AWS fastest growth since 2022, Meta 22% ad efficiency gains
Natural 5-10 year time lag between infrastructure and application revenue (distributed lag defense)
Unlike dot-com: Companies are profitable and cash-generative ($100B+ free cash flow annually)
Joseph Briggs (Goldman): 15% labor productivity increase possible if adoption scales

Framework Assessment

Early-stage $ROIC^{AI}$ (7-10%) is sub-hurdle for many but J-curve effect justifies patience
Power constraints are binding (3-30× gaps between disclosed PPAs and implied demand)
Utilization mystery is critical (industry 10-15% vs. hyperscaler "capacity constrained" claims)
Valuations moderately stretched (28× P/E vs. 15-17× historical) but not extreme (dot-com was 200×)
Systemic risk material due to concentration (Mag-7 is 30% of S&P 500) but fundamentals stronger than 2000

Python Implementation

Module Overview

ai_ledger.py: Build $IC^{AI}$ from PP&E roll-forwards and footnote disclosures
incremental_roic.py: Calculate $ROIC^{AI}$ with three attribution methods, report bounds
lag_model.py: ARDL distributed lag model (3-5 year infrastructure payback)
terminal_multiple.py: Validate conservation-consistent EV/EBITDA multiples
power_analyzer.py: Check power constraints, estimate GPU utilization, extract PPAs
quality_adjust.py: Apply $Q_t$ haircuts to AI cash flows for noisy reporters
cost_parity.py: Test Nvidia's "90% cost savings" claims with shadow P&L
bubble_risk.py: Systemic risk scoring (concentration, valuation, circular financing)

Status: Python implementation delegated to Codex via CODEX_TASKS/TASK_12_AI_ROI_IMPLEMENTATION.md. Target: 8 modules (~1,200 LOC), 85%+ test coverage, oracle fixtures for MSFT/META/GOOGL/AMZN.

Data Sources

Data Type	Source	Quality
Total CapEx, Segment Revenue, Operating Income	10-K/10-Q (audited)	HIGH
PP&E Roll-Forwards, PPA Commitments	Footnotes (audited)	HIGH
AI Business Metrics ($13B Azure AI run rate)	Earnings Calls, Press Releases	MODERATE
AI-Attributed NOPAT	Estimated (3 methods)	LOW (±30-40%)
GPU Utilization Rates	Industry Benchmarks	LOW (±50%)

Measurement Controversies

📉

Macro Impact (10× Disagreement)

Acemoglu: 0.5% productivity gain. Goldman Sachs: 15% gain. Source: task exposure rate (4.6% vs. 30%), complementary investments, GE effects.

💻

Utilization Mystery

Industry claims "capacity constrained" (implies 100%). Measured: 85%+ idle time. At 15% utilization, effective capex is 6.7× nominal.

💰

Revenue Attribution

Sequoia's "$600B Question": Current AI revenue ~$100B vs. $600B needed. Rebuttal: Natural 5-10 year lag between infrastructure and applications.

🎈

Bubble vs. Boom

Bears: Valuations at dot-com levels, circular financing. Bulls: Unlike 2000, companies profitable with $100B+ FCF. Framework: Moderately stretched, not extreme.

Metric	Value	Status
AI Invested Capital \((IC^{AI})\)	$111.0B (ending), $98.0B (average)	DISCLOSED
\(\Delta\) NOPAT (AI-attributed)	$5.3B – $8.9B	ESTIMATED
\(ROIC^{AI}\)	5.4% – 9.1%	AMBIGUOUS
WACC	8.5%	CALCULATED
Implied EV/EBITDA \((g=6.5\%)\)	11.7×	CONSISTENT
Analyst Multiple	15.0×	TOO HIGH (+3.3×)
PPA Capacity	0.8 GW	CONSTRAINED (104× gap)
Implied Power Demand	83.3 GW	INFEASIBLE