AI Ledger Methodology: Tracking AI Assets Under Conservation Constraints

Date: November 2025 Parent Framework: AI ROI Extension

Overview

The AI Ledger is a sub-ledger within Property, Plant & Equipment (PP&E) and Intangible Assets that tracks AI-specific invested capital (IC(^{AI})) from public filings. By embedding AI asset tracking within the existing 15-constraint conservation matrix, we ensure AI capex analysis respects double-entry accounting physics.

1. AI Invested Capital Definition

[ ]

where:

Net Fixed Assets (AI): [ NFA^{AI}_t = ^{AI}_t + ^{AI}_t ]

Components: 1. GPU Servers: NVIDIA H100/A100 clusters, AMD MI300, Google TPUs 2. Data Center Infrastructure: Buildings, cooling systems, power distribution 3. Networking Equipment: InfiniBand, Ethernet switches for GPU interconnect 4. Capitalized AI Software: Model training frameworks, inference engines, proprietary algorithms 5. Power Infrastructure: On-site generation, PPA commitments (capitalized portion)

Net Working Capital (AI) (typically small): [ NWC^{AI}_t = _t + _t - _t ]

2. Roll-Forward Constraint

2.1 Conservation Equation

[ IC^{AI}t = IC^{AI}{t-1} + ^{AI}_t - ^{AI}_t - ^{AI}_t + _t ]

This extends the existing PP&E roll-forward constraint in the accounting core.

2.2 Matrix Representation

Augment the constraint matrix (A) with AI-specific rows:

[
$$\begin{bmatrix} \vdots \\ 1 & -1 & +1 & -1 & -1 & 0 & \cdots \\ \vdots \end{bmatrix} \begin{bmatrix} \vdots \\ IC^{AI}_t \\ IC^{AI}_{t-1} \\ \text{Additions}^{AI}_t \\ \text{D\&A}^{AI}_t \\ \text{Disposals}^{AI}_t \\ \vdots \end{bmatrix}$$

= 0 ]

Key Benefit: If reported roll-forward doesn’t balance, the feasibility LP will correct it, and (Q_t) will flag the inconsistency.

3. Data Extraction from 10-K/10-Q

3.1 Direct Disclosures

Best Case: Separate AI capex disclosure (rare)

Example (Microsoft 2024 Annual Report, hypothetical): > “Capital expenditures for fiscal year 2024 were $48.4 billion, of which approximately $35 billion related to cloud and AI infrastructure.”

Extract: - Total Capex: $48.4B - AI Capex (estimated): $35B (72% of total)

3.2 Footnote Roll-Forwards

PP&E Schedule (typical 10-K format):

Property, Plant and Equipment (in millions)
                                2024      2023
Buildings                      $ 80,000  $ 72,000
Computer equipment              195,000   145,000
  (of which: AI/ML servers       85,000    45,000)  ← Footnote disclosure
Furniture and fixtures            5,000     4,800
Construction in progress         25,000    18,000
  Total                         305,000   239,800
Less: Accumulated D&A          (125,000) (105,000)
  Net PP&E                     $180,000  $134,800

Extraction Logic: 1. AI Servers: “AI/ML servers” line item = $85B (2024), $45B (2023) 2. Additions: $85B - $45B = $40B gross additions 3. D&A (AI): Estimate as (AI % of Computer Equipment) × (Total D&A on Computer Equipment)

If AI servers are 44% of computer equipment ($85B / $195B), and computer equipment D&A is $18B, then: [ ^{AI}_t ,000 = $7,920M ]

3.3 Intangible Assets (Capitalized Software)

Intangibles Schedule:

Intangible Assets (in millions)
                                2024      2023
Developed technology           $ 12,000  $ 10,500
Customer relationships           8,500     8,200
Patents                          3,200     3,100
Capitalized software            18,000    14,000
  (of which: AI models/platform  6,500     3,000)  ← AI-specific
  Total                          41,700    35,800
Less: Accumulated amortization (22,000)  (18,500)
  Net intangibles              $ 19,700  $ 17,300

Extraction: - AI Intangibles (2024): $6,500M - AI Intangibles (2023): $3,000M - Additions: $6,500M - $3,000M = $3,500M

3.4 Construction in Progress

Often, data centers under construction are in CIP (Construction in Progress):

Construction in progress         $25,000   $18,000
  (of which: AI data centers      15,000     8,000)

Treatment: Include in IC(^{AI}_t) if tagged as AI-related. Upon completion, reclassify from CIP to Buildings/Computer Equipment.

3.5 Narrative Disclosures (MD&A)

If quantitative footnotes are vague, extract from Management Discussion & Analysis:

“We invested $48 billion in capital expenditures in fiscal 2024, primarily for cloud and AI infrastructure to support Azure growth and AI product development. AI-related capital spending accounted for approximately 70% of total capex.”

Extraction: AI Capex ≈ 0.70 × $48B = $33.6B

4. Mapping to XBRL Tags

4.1 US-GAAP Taxonomy

AI Ledger Component Primary US-GAAP Tag Alternatives
Computer Equipment (AI Servers) PropertyPlantAndEquipmentGross ComputerEquipmentGross
Capitalized Software (AI) CapitalizedComputerSoftwareGross IntangibleAssetsGrossExcludingGoodwill
D&A (AI-attributed) Depreciation DepreciationDepletionAndAmortization
Capex (AI-specific) PaymentsToAcquirePropertyPlantAndEquipment CapitalExpendituresIncurredButNotYetPaid

4.2 Extensible Taxonomy

Some companies create custom tags for AI assets:

<us-gaap:PropertyPlantAndEquipmentGross contextRef="FY2024">
  305000000000
</us-gaap:PropertyPlantAndEquipmentGross>

<MSFT:AIMLServersGross contextRef="FY2024">
  85000000000
</MSFT:AIMLServersGross>

Our extractor scans for: - Custom tags containing “AI”, “ML”, “MachineLearning”, “GPU”, “Accelerated” - Footnote anchors referencing “artificial intelligence”, “generative AI”, “large language models”

5. Estimation When Data Is Sparse

5.1 Proxy Method: Capex-to-Revenue Ratio

If AI-specific capex is not disclosed:

[ _t _t ]

Assumption: Capital intensity scales with revenue contribution.

Example (Meta FY2024): - Total Capex: $38B - AI-attributed revenue growth: +$5B (from ad targeting improvements) - Total revenue growth: +$20B - AI Capex (estimated): $38B × (5/20) = $9.5B

5.2 Industry Benchmark

Use peer average when company-specific data unavailable:

Peer AI Capex as % of Total (2024)
Microsoft 72%
Amazon (AWS) 65%
Google 68%
Meta 55%
Median 66.5%

For unknown Company X with Total Capex = $10B: [ _t^X $10B = $6.65B ]

5.3 Power Consumption Back-Calculation

If power usage is disclosed:

[ _t ]

Typical Ratio: $1M capex per MW of data center power capacity (conservative; ranges $0.5M–$2M depending on PUE and GPU density).

Example (AWS: “3.8 GW added”): [ 3,800 $1M/ = $3.8B ]

6. Reconciliation with Cash Flow Statement

6.1 Capex Validation

Cash Flow from Investing:

Payments to acquire PP&E          $(48,400)
Capitalized software costs          (2,100)
  Total capex                      $(50,500)

Check: Does PP&E roll-forward “Additions” match cash outflow?

[ ^{} + ^{} ]

Discrepancy: Accrued but unpaid capex, finance leases, or timing differences.

If discrepancy > 5%, flag for quality adjustment ((Q_t) increases).

6.2 Finance Lease Treatment

IFRS 16 / ASC 842: Finance leases capitalized on balance sheet.

Cash Flow: Lease payments split between interest (operating) and principal (financing).

AI Ledger: Include right-of-use (ROU) assets for leased data centers:

[ IC^{AI}_t = IC^{AI}_t + ^{AI}_t ]

Extract from Lease Footnote:

Operating lease ROU assets        $ 8,500
Finance lease ROU assets            4,200
  (of which: Data centers)          3,800  ← AI-specific

7. Depreciation and Amortization

7.1 Useful Life Assumptions

Typical D&A Schedules:

Asset Type Useful Life Annual D&A Rate
Buildings 30–40 years 2.5%–3.3%
GPU Servers 3–5 years 20%–33%
Networking Equipment 5–7 years 14%–20%
Capitalized Software 3–5 years 20%–33%

AI-Heavy Impact: GPUs depreciate fast (3–5 yrs) → high D&A burden → pressure on ROIC.

7.2 D&A Calculation

If IC(^{AI}_t) = $111B and weighted average life = 4 years:

[ ^{AI}_t = $27.8B/ ]

Check against reported: Does this match the D&A component attributable to AI assets in the income statement?

If reported D&A is lower, either: 1. Useful lives are longer than assumed 2. Not all AI assets fully deployed (CIP not yet depreciating) 3. Residual values assumed (reducing depreciable base)

8. Power Purchase Agreements (PPAs)

8.1 Capitalization Rules

US GAAP / IFRS: Long-term PPAs may be capitalized if they convey right to use specific assets.

Example: Microsoft’s 20-year nuclear PPA with Constellation - If structured as operating lease: Off-balance-sheet (pre-ASC 842) - If structured as finance lease / right-to-use: On-balance-sheet as ROU asset

8.2 Extraction from Footnotes

Commitments & Contingencies Section:

Purchase Commitments:
  Power purchase agreements         $18,500  (over 15 years)
  Annual commitment (avg):            1,233
  Capitalized portion (ROU):          8,200  ← Add to IC^AI

8.3 Power as Constraint

If total PPA capacity < Implied AI power demand:

[ = ]

Example: IC(^{AI}) = $111B, ratio = 0.75 MW/$M: [ = 111,000 / 1,000 = 83.3 ]

If disclosed PPA = 4.2 GW, power-constrained (70× shortfall).

9. Quality Metrics

9.1 Roll-Forward Consistency

Test: [ | IC^{AI}t - ( IC^{AI}{t-1} + - - ) | < ]

Typical Tolerance: 1% of IC(^{AI}_{t-1})

If violated, contributes to (Q_t): [ Q_t^{AI} = ]

9.2 Capex-to-Depreciation Ratio

Steady State: Capex ≈ D&A (replacement-level investment)

Growth Phase: Capex >> D&A (expanding capacity)

AI Reality (2024): Capex / D&A ≈ 3–5× (massive build-out)

Metric: [ = ]

If ratio > 10×, flag as aggressive expansion (higher risk).

10. Implementation Notes

10.1 Data Structures

Python Class:

@dataclass
class AILedger:
    cik: str
    fiscal_year: int
    ic_ai_beginning: float
    additions: float
    depreciation: float
    disposals: float
    fx_other: float
    ic_ai_ending: float
    rou_assets: float  # Finance leases
    ppa_capacity_gw: float
    quality_score: float  # Roll-forward consistency

10.2 Extraction Pipeline

  1. Parse 10-K XBRL: Extract PP&E schedule, intangibles, capex
  2. Tag Matching: Identify AI-related line items (regex + ML classifier)
  3. Roll-Forward Validation: Check conservation
  4. Gap Filling: Apply proxy methods if sparse
  5. PPA Augmentation: Add capitalized ROU assets
  6. Output Ledger: IC(^{AI}_t) time series

10.3 Test Coverage

Unit Tests: - Roll-forward arithmetic (conservation holds) - XBRL tag extraction (no false positives/negatives) - Proxy method accuracy (within ±10% of disclosed values when available)

Integration Tests: - Full pipeline on MSFT, GOOGL, META, AMZN 10-Ks (FY2022–2024) - Validate against analyst estimates (consensus benchmarks)

Target: 90% accuracy on disclosed AI capex; 70% accuracy on estimated AI capex (vs. analyst consensus).

11. Citations

Amazon.com, Inc. (2024). Form 10-K for fiscal year ended December 31, 2024. U.S. Securities and Exchange Commission.

Bloom Energy. (2025). 2025 Data Center Power Report.

Citigroup Research. (2025). Big Tech AI Spending Forecast.

Financial Accounting Standards Board. (2016). ASC 842: Leases.

Microsoft Corporation. (2024). 2024 Annual Report. https://www.microsoft.com/investor/reports/ar24/

Module: src/ai_roi/ai_ledger.py Test Module: tests/ai_roi/test_ai_ledger.py Oracle Fixtures: tests/ai_roi/oracles/msft_2024_ledger.yaml