AI ROI Framework: Conservation-Consistent Measurement
Date: November 2025 Status: Phase 8 Extension
Executive Summary
This framework extends the discrete accounting conservation system to answer the central question facing investors and analysts:
Will the $1–3 trillion wave of AI capital expenditures earn their cost of capital?
By embedding AI-specific metrics within the 15-constraint accounting core, we transform qualitative claims about AI productivity into testable, filing-grounded economics that survive diligence.
Key Innovation: Traditional DCF models treat AI capex as generic capital. This framework enforces incremental ROIC on AI assets (ROIC(^{AI})) must respect the growth-reinvestment identity and conservation constraints, preventing internally inconsistent valuations.
1. The Wall Street Question
1.1 Investment Scale
From 2024–2029, hyperscalers (Microsoft, Amazon, Google, Meta, Oracle) are projected to invest:
- $2.8 trillion in AI infrastructure (Citigroup, 2025)
- $55 GW of new power capacity required (Bloom Energy, 2025)
- $1.4 trillion in U.S. power infrastructure alone
Q3 2024 Snapshot: Top 4 hyperscalers spent $58.9B in capex (63% YoY growth), with capex-to-revenue ratios reaching 22% vs. historical 11–16%.
1.2 Return Uncertainty
Bearish View (Goldman Sachs, June 2024): - ~$1T spend with “little to show for it” beyond developer efficiency - Daron Acemoglu (MIT): 0.5% productivity increase over next decade - Only 6.1% of U.S. businesses currently using AI for production
Bullish View (Morgan Stanley, 2024): - GenAI expected to generate profits starting 2025 (34% margin) - $153B revenue (2025) → $1.1T (2028), a 20× increase - Joseph Briggs (GS): 15% labor productivity increase possible
Bubble Warnings (IMF, Bank of England, October 2024): - Stock valuations “comparable to the peak” of 2000 dot-com bubble - “Growing risk that AI bubble could burst” with systemic implications - High concentration in small cluster of AI-heavy companies
1.3 Measurement Gap
Despite massive investment: - 97% of enterprises struggle to demonstrate GenAI business value (2024) - 65% adoption but only 11% use at scale - AI-attributed revenue and NOPAT not separately disclosed in 10-K/10-Q filings
This framework provides the first LP-based system to reconcile AI assumptions with accounting conservation, enabling firm-by-firm, quarter-by-quarter ROI measurement from public filings.
2. Conservation Framework Extension
2.1 Accounting Core (Review)
The base system uses a 15-constraint integer matrix (A ^{15 m}) encoding double-entry conservation:
[ A = ]
where () contains line items from the balance sheet, income statement, and cash flow statement. Any reported filing (_t) with non-zero residual (_t = A _t) is projected to the feasible set:
[ t = { } |( - _t)|_1 A = ]
Quality Metric: [ Q_t = ]
Higher (Q_t) indicates more “fixes” needed to satisfy conservation, signaling lower reporting quality.
2.2 AI-Specific Augmentation
We extend the constraint matrix to track AI assets as a sub-ledger within PP&E and intangibles:
[ IC^{AI}_t = NWC^{AI}_t + NFA^{AI}_t ]
where: - **(NFA^{AI}_t): GPU servers, data center infrastructure, capitalized AI software (from footnote roll-forwards) - (NWC^{AI}_t)**: Working capital directly attributable to AI operations (typically small)
Roll-Forward Constraint (extends existing PP&E conservation): [ IC^{AI}t = IC^{AI}{t-1} + ^{AI}_t - ^{AI}_t - ^{AI}_t ]
This embeds AI capex within the existing conservation matrix, ensuring AI asset tracking respects double-entry physics.
3. Incremental ROIC on AI Assets
3.1 Definition
Return on Invested Capital for AI infrastructure:
[ ]
where: - (NOPAT_t): Incremental Net Operating Profit After Tax attributed to AI (see §3.2) - **(_{t-1:t})**: Average AI invested capital over the period
Threshold Test: If (ROIC^{AI} < WACC), the AI capex is value-destructive.
3.2 NOPAT Attribution Methodology
AI-attributed NOPAT is not directly disclosed. We estimate via three approaches:
Method 1: Top-Down Trend Break
Compare actual NOPAT growth to pre-AI trend:
[ NOPAT^{AI}_t = NOPAT_t - _t^{} ]
where (_t^{}) is extrapolated from 2020–2022 trend (before major AI capex).
Advantages: Simple, uses only financial statements Disadvantages: Attributes all trend deviation to AI (may miss other factors)
Method 2: Bottom-Up Unit Economics
For disclosed AI revenue segments (e.g., Microsoft’s “$13B AI business run rate”):
[ NOPAT^{AI}_t = _t ]
Incremental Margin Estimation: - Use management-disclosed margins (e.g., Meta: “strong ROI from core AI”) - Or apply segment-level operating margin adjusted for infrastructure depreciation
Advantages: Ties to disclosed revenue Disadvantages: Requires margin assumptions; AI revenue not always broken out
Method 3: Cost Savings (Accelerated Computing)
For GPU-based compute cost reductions:
[ NOPAT^{AI}_t = _t - _t - ^{GPU}_t ]
where: - Baseline CPU Cost: Estimated from prior workload requirements - Accelerated Cost: GPU opex (power, maintenance) - D&A(^{GPU}): Depreciation on GPU assets
Nvidia’s Claim: “90% cost savings” from accelerated computing. We test this with disclosed unit economics (see §6).
Advantages: Validates vendor claims with filings Disadvantages: Requires workload-specific modeling
3.3 Multi-Method Consensus
In practice, we compute all three estimates and report bounds:
[ ROIC^{AI}_t ]
If the upper bound < WACC, AI capex is definitively value-destructive. If the lower bound > WACC, AI capex is definitively value-creative. Otherwise, ROI is ambiguous given data limitations.
4. Distributed Lag Model
4.1 Infrastructure Payback Timeline
AI infrastructure (data centers, GPU farms) requires 3–5 years to generate full returns, evidenced by:
- Academic Literature (Bom & Ligthart, 2014): Public infrastructure shows short-run elasticity 0.083, long-run 0.122 (5–10 year lag)
- AI-Specific Evidence (McElheran et al., 2025): Initial productivity decline of 1.33% to 60% (J-curve), recovering over 3–5 years
- Hyperscaler Timeline (Morgan Stanley, 2024): Profitability expected 2025 for early movers
We model AI capex impact with Autoregressive Distributed Lag (ARDL) structure:
[ NOPAT_t = + _{k=0}^{K} k {t-k} + _t + _t ]
Typical Parameters (from infrastructure literature): - (K = 4) to (8) quarters (1–2 years lag) - Peak coefficient at (k = 4) (1-year lag) - 50% of total effect within 6 months, 90% within 3 years
4.2 Net Present Value Calculation
Over a 5-year horizon:
[ NPV^{AI} = {t=0}^{20} - {t=0}^{20} ]
where (t) indexes quarters.
Decision Rule: Invest if (NPV^{AI} > 0), accounting for distributed lags.
Implication: A project showing negative annual ROIC in years 1–2 may still be NPV-positive if the J-curve effect is temporary.
5. Terminal Value Consistency
5.1 Conservation-Consistent Multiple
The implied enterprise value-to-EBITDA multiple must satisfy:
[ ]
where: - (): Effective tax rate - (g): Terminal growth rate - (ROIC): Return on invested capital (including AI assets) - (WACC): Weighted average cost of capital
Feasibility Bounds: 1. (WACC > g) (else perpetuity diverges) 2. (g ROIC) (implied by growth-reinvestment identity (g = s ROIC) with (s )) 3. (g ) for mature firms (negative terminal growth requires explicit justification) 4. () (empirical tax rate bounds for developed markets)
5.2 AI Growth Constraints
If AI capex drives terminal growth assumptions:
[ g = + ]
Power Constraint Check: - If (g) implies (X) GW of compute capacity - And only (Y < X) GW is feasible given power constraints - Then (g) must be reduced to (g’ = g (Y/X))
This prevents analysts from forecasting growth rates that physically cannot be supplied with electricity.
5.3 Solver Output
For a given set of inputs ((g, ROIC, WACC, )), the validator:
- Checks feasibility bounds (reports violations as
terminal_value_infeasible) - Computes implied EV/EBITDA
- Compares to analyst’s chosen multiple
- Reports inconsistency gap (^*_{} = | - |)
If (^* > 0.5) (tolerance: 0.5 turns of EBITDA), the model is internally inconsistent.
6. Cost-of-Compute Parity Test
6.1 Accelerated Computing Claims
Nvidia and partners claim: - “90%+ cost/energy savings” from GPU vs. CPU for certain workloads - “Best ROI computing infrastructure investment” (Jensen Huang, CEO)
We formalize as a shadow P&L:
[ _t = _t - _t ]
where: - Baseline CPU Cost: (_t ) - Accel Cost: (_t + _t + _t)
Data Sources: - FLOPs per workload: Published in academic papers (e.g., GPT-4 training: (2 ^{25}) FLOPs) - CPU $/FLOP: AWS EC2 pricing ($0.096/hr for c6i.xlarge, 19.2 GFLOPS ≈ $5/GFLOP-hr) - GPU $/FLOP: AWS p4d.24xlarge pricing ($32.77/hr, 8×A100 = 2.5 PFLOPS ≈ $0.013/GFLOP-hr, 385× cheaper per FLOP) - Power: A100 draws 400W; industrial electricity $0.07/kWh
6.2 Empirical Case Studies
Commonwealth Bank of Australia (Nvidia case study): - 640× performance boost (RAPIDS Accelerator) - 80% cost reduction vs. CPU
AT&T: - 3.3× faster data processing - 60% lower cost
IRS: - 20× speed improvements - 50% cost reduction
We extract disclosed metrics from footnotes and test consistency with Nvidia’s unit-economics claims.
6.3 Feasibility Check
If savings < (D&A + Power), ROI fails despite performance gains. Embed as contra-COGS item:
[ _t = _t - _t ]
Then recompute margins on conservation-adjusted financials (_t).
7. Power Constraints and Utilization
7.1 Capacity Requirement
Citi Estimate (2025): 55 GW of new power needed by 2030 Cost: ~$50B per GW ($2.8T total infrastructure)
AWS Disclosure (2024): Added 3.8 GW capacity in past 12 months
For a given AI capex trajectory, implied power demand:
[ _t^{} = ]
Typical Ratio: ~0.5–1 MW per $1M in data center capex (varies by PUE and GPU density)
7.2 Utilization Rate Estimation
Industry Reality (multiple sources): - 85%+ of GPU capacity sits idle - 10–15% typical utilization in hybrid research/production systems
Traditional metrics (nvidia-smi) overstate utilization by measuring kernel execution time, not computational throughput. Advanced metrics (SM efficiency) reveal true waste.
Financial Impact: At 15% utilization, effective cost per useful GPU-hour is 6.7× nominal cost.
Adjustment Factor: [ ^{AI}_t = ]
Recalculate ROIC(^{AI}) using effective IC to reflect capital efficiency.
7.3 Power Purchase Agreement (PPA) Tracking
Extract from 10-K risk factor disclosures: - Microsoft + Constellation: 20-year PPA for Three Mile Island reactor (PA) - AWS + Talen Energy: 960 MW from Susquehanna nuclear plant (PA)
Constraint: If disclosed PPA capacity < implied power demand, flag power-constrained scenario.
8. Quality-Adjusted Cash Flows
8.1 Haircut Formula
For companies with high (Q_t) (noisy accounting):
[ _t = FCFF_t (1 - Q_t) ]
where () is a severity parameter (default: 0.5).
Rationale: Mechanically inconsistent reporters don’t get full credit in DCF valuation.
8.2 Application to AI Capex
If AI-attributed NOPAT estimates rely on management commentary (Method 2), apply quality adjustment:
[ NOPAT^{AI, }_t = NOPAT^{AI}_t (1 - Q_t) ]
This penalizes companies with loose accounting discipline when making aggressive AI ROI claims.
9. Comps by Conservation Signature
9.1 Signature Vector
For each company, compute:
[ =]
from conservation-adjusted financials (_t).
9.2 Distance Metric
Choose comps by minimizing Mahalanobis distance:
[ d_{ij} = ]
where () is the covariance matrix of signatures across the universe.
Advantage: Accounting-consistent multiples for private valuations, ignoring buzzwords.
10. Debt Capacity and Rating Impact
10.1 Coverage Ratio
As AI capex scales, interest coverage may decline:
[ _t = ]
Rating Thresholds (S&P): - AAA: Coverage > 20× - AA: Coverage 10–20× - A: Coverage 6–10× - BBB: Coverage 3–6×
10.2 WACC Re-Pricing
If AI-driven capex pushes coverage below threshold: 1. Credit rating downgrades (e.g., AA → A) 2. Cost of debt (r_d) increases (e.g., 4.0% → 4.5%) 3. WACC rises: (WACC = w_e r_e + w_d r_d (1-)) 4. Hurdle rate increases, making AI capex less attractive
This feedback loop is embedded in the solver to prevent circular logic.
11. Practical Workflow
11.1 Input Data
From 10-K/10-Q: - Total capex (quarterly, annual) - Segment revenue and operating income - PP&E roll-forwards (additions, D&A, disposals) - Footnote disclosures for AI assets (GPUs, data centers, capitalized software) - Risk factor mentions of power, AI competition
From external sources: - AI revenue estimates (analyst reports, management commentary) - GPU utilization benchmarks (industry surveys, academic papers) - Power PPA details (press releases, utility filings)
11.2 Computation Steps
- Build AI Ledger: Extract IC(^{AI}_t) from roll-forwards + footnotes
- Estimate (NOPAT^{AI}): Apply Methods 1–3, report bounds
- Calculate ROIC(^{AI}): Compare to WACC
- Lag Model: Fit ARDL with (K=4) to (8) quarters
- Terminal Multiple: Check physics-consistency with growth assumptions
- Power Constraint: Validate capacity vs. implied demand
- Quality Adjust: Apply (Q_t) haircut if accounting noisy
- Generate Report: Bands, feasibility flags, sensitivity tables
11.3 Output Format
Company: MSFT (FY2024)
──────────────────────────────────────
AI Invested Capital: $111.0B (ending), $98.0B (average)
Δ NOPAT (AI, estimated): $5.3B – $8.9B
ROIC^AI: 5.4% – 9.1%
WACC: 8.5%
Status: ⚠ AMBIGUOUS (lower bound < WACC < upper bound)
──────────────────────────────────────
Terminal Multiple (15.0×) implies:
g = 6.5%, ROIC = 10%, WACC = 8.5%
Physics-consistent range: 11.7×
→ EXIT MULTIPLE TOO HIGH BY 3.3×
──────────────────────────────────────
Power Constraint:
Disclosed PPA: 0.8 GW
Implied Demand: 83.3 GW
→ POWER-CONSTRAINED (104× gap)
──────────────────────────────────────
Recommendation: Expand PPA capacity or
reduce terminal growth12. Measurement Uncertainty
12.1 Data Availability
Available in 10-K/10-Q (high quality): - Total capex by segment - Segment revenue and operating income - PP&E roll-forwards - Risk factor disclosures (qualitative)
Requires Estimation (moderate to high uncertainty): - AI-specific capex (% of total) - AI-attributed NOPAT - GPU utilization rates - Distributed lag parameters
Not Disclosed (must infer): - Model training/inference costs per unit - PPA pricing terms ($/MWh) - Detailed ROI by AI product line
12.2 Sensitivity Analysis
For each key input, compute elasticity:
[ ]
High-Impact Parameters (from literature): 1. AI-attributed NOPAT margin (±5 pp changes ROIC(^{AI}) by ±2–3 pp) 2. Utilization rate (15% → 30% halves effective capex, doubles ROIC(^{AI})) 3. Distributed lag (K) (4 qtrs → 8 qtrs delays payback, reduces NPV by 10–15%)
Generate tornado charts and scenario tables for transparency.
13. Citations
Acemoglu, D. (2024). The Simple Macroeconomics of AI. MIT Economics. https://economics.mit.edu/sites/default/files/2024-05/The%20Simple%20Macroeconomics%20of%20AI.pdf
Bank of England. (2024, October). Financial Stability Report. https://www.bankofengland.co.uk/
Bloom Energy. (2025). 2025 Data Center Power Report. https://www.bloomenergy.com/
Bom, P. R., & Ligthart, J. E. (2014). What Have We Learned From Three Decades Of Research On The Productivity Of Public Capital? Journal of Economic Surveys, 28(5), 889-916.
Brynjolfsson, E., Li, D., & Raymond, L. R. (2023). Generative AI at Work (NBER Working Paper No. 31161). https://www.nber.org/papers/w31161
Citigroup Research. (2025, September). Big Tech AI Spending Forecast.
Goldman Sachs Research. (2024, June). Gen AI: Too Much Spend, Too Little Benefit? Top of Mind, Issue 129. https://www.goldmansachs.com/insights/top-of-mind/gen-ai-too-much-spend-too-little-benefit
McElheran, K., Yang, M., Brynjolfsson, E., & Kroff, Z. (2025). The Rise of Industrial AI in America: Microfoundations of the Productivity J-curve(s) (Census Working Paper CES-WP-25-27). https://www2.census.gov/library/working-papers/2025/adrm/ces/CES-WP-25-27.pdf
Morgan Stanley Research. (2024). GenAI Revenue Growth and Profitability. https://www.morganstanley.com/insights/articles/genai-revenue-growth-and-profitability
Framework Version: 1.0 (November 2025)
Module: src/ai_roi/ Test Coverage
Target: 85%+ Publication Status: Draft for
peer review