Measurement Challenges: Data Availability and Estimation Methods

Date: November 2025

Executive Summary

Measuring AI capex ROI faces fundamental data limitations: AI-attributed revenue and NOPAT are not separately disclosed in financial statements. This document catalogs what’s available, what requires estimation, and quantifies uncertainty.

Key Finding: 97% of enterprises struggle to demonstrate GenAI business value despite 65% adoption (McKinsey, 2024).

1. Data Availability Matrix

1.1 HIGH QUALITY (Disclosed in 10-K/10-Q)

Example (Microsoft FY2024): - Total Capex: $48.4B (disclosed, precise) - Intelligent Cloud Revenue: $109.6B (disclosed, precise) - Intelligent Cloud Operating Income: $52.1B (disclosed, precise)

1.2 MODERATE QUALITY (Voluntary Disclosures)

Example (Microsoft Q2 FY2025 Earnings): - AI business: “$13B annual run rate” (unaudited, not in 10-Q) - Azure AI growth: “175% YoY” (qualitative, no segment breakdown)

Issue: Definitions vary across companies; no GAAP standardization.

1.3 REQUIRES ESTIMATION (Not Disclosed)

2. Estimation Methods and Uncertainty

2.1 AI-Attributed NOPAT (Numerator of ROIC^AI)

Method 1: Top-Down Trend Break

[ NOPAT^{AI}_t = NOPAT_t - _t^{} ]

Uncertainty Sources: - Macro conditions (interest rates, demand shocks) confound trend - Other strategic initiatives (non-AI) also drive NOPAT - Pre-AI trend window choice (2020-2022 includes COVID distortions)

Typical Error Bounds: ±30% (comparing to disclosed AI metrics when available)

Method 2: Bottom-Up Unit Economics

[ NOPAT^{AI}_t = _t ]

Uncertainty Sources: - AI revenue is voluntary disclosure, inconsistently defined - Incremental margin requires assumptions (20-40% range typical for cloud/SaaS) - Depreciation allocation to AI unclear

Typical Error Bounds: ±40% (wide margin assumption range)

Method 3: Cost Savings (Accelerated Computing)

[ NOPAT^{AI}_t = _t - ^{GPU}_t ]

Uncertainty Sources: - Baseline CPU cost is counterfactual (what would’ve been spent) - Nvidia’s “90% savings” claims are workload-specific (ML training), not universal - Allocation of shared infrastructure costs

Typical Error Bounds: ±50% (highly workload-dependent)

Consensus Approach: Report bounds across all three methods:

[ ROIC^{AI}_t ]

2.2 AI-Specific CapEx (Denominator of ROIC^AI)

Method 1: Footnote Extraction

If PP&E schedule breaks out “AI/ML servers”: - Uncertainty: ★★☆☆☆ LOW (when disclosed) - Availability: ~30% of hyperscalers provide this level of detail

Method 2: Qualitative Disclosure Percentage

“Approximately 70% of capex for AI infrastructure” (MD&A): - Uncertainty: ★★★☆☆ MODERATE (management estimate, unaudited) - Availability: ~60% provide qualitative percentage

Method 3: Proxy Ratio

[ _t = _t ]

• Uncertainty: ★★★★☆ HIGH (assumes capital intensity scales linearly with revenue)
• Availability: Always possible (last resort)

Method 4: Industry Benchmark

Median AI capex as % of total (peers): 66.5% (2024) - Uncertainty: ★★★☆☆ MODERATE TO HIGH (company-specific strategies vary) - Use Case: Private companies or non-disclosers

2.3 GPU Utilization Rates

Industry Evidence: - 85%+ idle time (multiple surveys, academic studies) - 10-15% typical utilization in hybrid systems - 40% considered good

Measurement Challenge: Companies don’t disclose; requires inference from: - Academic research (NERSC Perlmutter study, METR analysis) - Third-party monitoring tools (when available) - Industry practitioner surveys

Financial Impact: [ IC^{AI, } = ]

At 15% utilization: Effective capex is 6.7× nominal, crushing ROIC^AI.

Uncertainty: ★★★★☆ VERY HIGH (ranges 10-50% depending on workload mix)

Conservative Assumption: Use 20-30% utilization for production inference workloads; 10-15% for research/training.

3. Quality Score Incorporation

3.1 Accounting Quality Metric (Q_t)

From base framework: [ Q_t = ]

Interpretation: Higher Q_t → more “fixes” needed → lower reporting quality.

3.2 Haircut Application

For companies with noisy accounting (high Q_t):

[ NOPAT^{AI, }_t = NOPAT^{AI}_t (1 - Q_t) ]

where () is severity parameter (default: 0.5).

Rationale: If a company can’t maintain clean roll-forwards for basic assets, their aggressive AI ROI claims deserve skepticism.

Example: - Company A: Q_t = 0.02 (clean) → Haircut = 1% → Accept AI claims at 99% face value - Company B: Q_t = 0.15 (sloppy) → Haircut = 7.5% → Discount AI claims by 8%

4. Distributed Lag Uncertainty

4.1 Parameter Selection

Academic Literature (Bom & Ligthart, 2014; Kumar, 2024): - Infrastructure shows 3-5 year lags to full productivity impact - Peak coefficient at K=4 quarters (1 year lag)

AI-Specific Evidence (McElheran et al., 2025): - J-curve: Initial productivity decline 1.33-60% - Recovery timeline: 3-5 years

ARDL Model: [ NOPAT_t = + _{k=0}^{K} k {t-k} + _t ]

Uncertainty: Choice of (K) (4 vs. 8 quarters) changes NPV by 10-15%.

Sensitivity Analysis: Report NPV for K=4, K=6, K=8 to bound outcomes.

5. Measurement Controversies

5.1 The “$600B Question” (Sequoia Capital)

Argument: Infrastructure capex requires commensurate revenue to justify ROI. - Current GenAI revenue: ~$100B (estimated) - Infrastructure build-out implies: ~$600B revenue needed - Gap: $500B

Rebuttal: Natural 5-10 year time lag between infrastructure and application revenue (distributed lag defense).

Resolution: Our framework models this explicitly via ARDL → NPV over 5-year horizon.

5.2 Acemoglu vs. Goldman Sachs (10× Disagreement)

Acemoglu (MIT): 0.5% productivity increase over 10 years (only 4.6% of tasks exposed) Goldman Sachs / McKinsey: 15% productivity increase (30% of work hours automatable)

Source of Disagreement: 1. Task exposure rate: 4.6% vs. 30% 2. Productivity gain per task: Modest vs. transformational 3. General equilibrium effects: Deflationary pressures vs. margin expansion

Framework Approach: We measure firm-by-firm, quarter-by-quarter from filings, not macro forecasts. Aggregate ex-post.

5.3 Utilization Rate Controversy

Industry Claims: “Capacity constrained”, “sold out of GPUs” (implies near-100% utilization) Empirical Evidence: 85%+ idle time measured

Possible Explanations: 1. Peak demand != average utilization (provisioned for spikes) 2. Development/testing workloads highly bursty 3. Inefficient orchestration / scheduling

Financial Implication: If true utilization 10-15%, effective ROIC^AI is 6.7× lower than nominal.

6. Best Practices for Practitioners

6.1 Transparency Checklist

When reporting AI ROI analysis, disclose:

1. Data Sources: Which metrics from 10-K vs. estimated?
2. Estimation Methods: Which of the 3 NOPAT methods used?
3. Assumptions: Utilization rate, incremental margin, distributed lag K
4. Sensitivity: Tornado chart showing impact of key assumptions
5. Bounds: Report ranges, not point estimates
6. Quality Adjustment: Apply Q_t haircut for noisy reporters

6.2 Red Flags

Avoid these pitfalls: 1. Precision Illusion: Reporting ROIC^AI to 0.1% when underlying estimates have ±30% error 2. Cherry-Picking: Using Method 2 (bottom-up) because it gives highest number 3. Ignoring Utilization: Assuming 100% when industry average is 10-15% 4. Zero Lag: Expecting immediate payback when infrastructure takes 3-5 years 5. Quality Blindness: Trusting aggressive AI claims from companies with high Q_t

6.3 Example Disclosure (Good Practice)

MSFT AI ROIC Estimate (FY2024)
───────────────────────────────────────────────────────────
AI Invested Capital:        $111.0B ending ($98.0B average)
Δ NOPAT (AI, 3 methods):    $5.3B – $8.9B  (±40% uncertainty)
ROIC^AI (nominal):          5.4% – 9.1%
ROIC^AI (20% utilization):  1.1% – 1.8%  ← Effective, after util. adjustment
WACC:                       8.5%

Status:  ⚠ AMBIGUOUS (lower 5.4% < WACC 8.5% < upper 9.1%)
         🔴 BELOW HURDLE at industry utilization (10-15%)

Sensitivity:
  +10pp utilization →  +2.7pp ROIC^AI
  +5pp margin        →  +1.9pp ROIC^AI
  +2 qtrs lag (K)    →  -1.3pp NPV

Data Quality (Q_t):         0.03  (low, clean accounting)
Quality Haircut:            Minimal (1.5%)

Conclusion: ROI highly dependent on achieving >50% utilization;
            current industry benchmarks (10-15%) imply value destruction.

7. Future Improvements

7.1 Standardized Disclosure (Advocacy)

Proposal: SEC require separate AI capex and AI revenue segment disclosure for companies with >10% capex in AI.

Benefits: - Reduces estimation uncertainty from ±40% to ±10% - Enables better capital allocation across economy - Prevents AI hype from obscuring weak fundamentals

7.2 Third-Party Verification

Opportunity: Independent auditors verify: - GPU utilization rates (via DCGM telemetry) - Model training costs (FLOPs × time × power) - Incremental revenue attribution (A/B test results)

Precedent: Energy industry audits reserves; financial services audits risk models.

8. Citations

Brynjolfsson, E., Li, D., & Raymond, L. R. (2023). Generative AI at Work (NBER Working Paper No. 31161).

Cahn, D. (2024). AI’s $600B Question. Sequoia Capital Blog. https://www.sequoiacap.com/article/ais-600b-question/

McElheran, K., Yang, M., Brynjolfsson, E., & Kroff, Z. (2025). The Rise of Industrial AI in America. Census Working Paper CES-WP-25-27.

McKinsey & Company (2024). The State of AI in Early 2024. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-2024

Module: src/ai_roi/quality_adjust.py Recommended Practice: Always report bounds, never point estimates.