GHG Protocol Compliance¶

Guide for reporting AI emissions under the GHG Protocol Corporate Standard.

What is GHG Protocol?¶

The Greenhouse Gas Protocol is the world's most widely used greenhouse gas accounting standard. It provides frameworks for organizations to measure and report their emissions.

Classification¶

AI inference emissions fall under:

• Scope 3: Indirect emissions from value chain
• Category 1: Purchased Goods and Services

This classification applies because: - You don't own the data centers (not Scope 1) - You don't purchase the electricity directly (not Scope 2) - The API service is a purchased good (Scope 3, Cat 1)

Calculation Methods¶

GHG Protocol defines four methods for Scope 3, Category 1:

1. Supplier-Specific Method¶

Uses emissions data provided by the supplier (API provider).

Status: Not yet available — providers don't publish per-request emissions.

2. Hybrid Method¶

Combines supplier data with secondary data.

Status: Partially available — some providers publish aggregate data.

3. Average-Data Method¶

Uses industry-average emission factors.

Status: Primary method — OpenClaw uses academic research factors.

4. Spend-Based Method¶

Uses emissions per dollar spent.

Status: Available as fallback, but less accurate.

Data Quality¶

GHG Protocol requires reporting data quality. OpenClaw maps confidence to the 5-point scale:

DQS	Description	OpenClaw Confidence
1	Primary data from suppliers	≥80%
2	Published secondary data	≥60%
3	Average secondary data	≥40%
4	Estimated data	≥20%
5	Highly uncertain	<20%

Most AI inference emissions are DQS 3-4 (average-data method with research-based factors).

Export Format¶

Generate GHG Protocol-compliant export:

openclaw green export --format ghg-protocol --period 2025-Q1

Output Structure:

{
  "reportingPeriod": "2025-Q1",
  "organizationalBoundary": "Operational control - AI inference API usage",
  "scope3Category1": {
    "emissions_tCO2eq": 0.01245,
    "calculationMethod": "Average-data method using per-model emission factors",
    "dataQuality": "Good",
    "uncertainty_percent": 30,
    "emissionFactorSources": [
      "ML CO2 Impact Calculator (Lacoste et al., 2019)",
      "Cloud Carbon Footprint methodology",
      "CodeCarbon hardware measurements"
    ]
  }
}

Required Disclosures¶

When reporting under GHG Protocol, disclose:

1. Organizational Boundary¶

Operational control approach. Emissions from AI inference
API calls made by [Organization] using third-party providers.

2. Operational Boundary¶

Scope 3, Category 1: Purchased Goods and Services
Sub-category: Cloud computing services (AI inference)

3. Calculation Methodology¶

Emissions calculated using average-data method with per-model
emission factors derived from academic research (Lacoste et al. 2019,
Patterson et al. 2022). Factors account for GPU power consumption
during inference with 3:1 output-to-input energy ratio.

4. Emission Factors¶

Per-token emission factors (gCO₂eq per million tokens):
- Input tokens: 30-400 depending on model size
- Output tokens: 90-1200 (3x input due to iterative generation)
- Cache reads: ~10% of input

Grid carbon intensity: [configured value] gCO₂/kWh
PUE factor: 1.2 (cloud data center assumption)

5. Data Quality Assessment¶

Data Quality Score: 3 (Average secondary data)
Confidence: [X]%
Uncertainty: ±[Y]%

Limitations:
- No supplier-specific data available
- Model architecture details not disclosed
- Data center locations unknown

Verification¶

For third-party verification, provide:

1. Raw trace data: openclaw green export --format json
2. Summary by period: openclaw green export --format ghg-protocol
3. Methodology documentation: Link to this guide
4. Factor sources: Academic papers cited

Example Report Section¶

## Scope 3 Emissions - Category 1

### AI Inference Services

| Metric | Value | Unit |
|--------|-------|------|
| Total emissions | 12.45 | kg CO₂eq |
| API calls | 1,847 | count |
| Avg per call | 6.74 | g CO₂eq |
| Data quality | 3 | DQS (1-5) |
| Uncertainty | ±30% | |

**Methodology**: Average-data method using per-model emission
factors from academic research. Factors applied to token counts
from API responses.

**Emission factor sources**:
- Lacoste et al. (2019) "Quantifying the Carbon Emissions of Machine Learning"
- Patterson et al. (2022) "Carbon Emissions and Large Neural Network Training"
- CodeCarbon project hardware measurements

**Limitations**: No supplier-specific data. Estimates based on
model size heuristics and published research on similar architectures.

Best Practices¶

1. Report quarterly — More granular than annual, catches trends
2. Track by provider — Different providers have different footprints
3. Document assumptions — Grid carbon, PUE, model mappings
4. Update factors — Check for new research/provider data annually
5. Set targets — Use SBTi framework for reduction commitments

Related Standards¶

• CDP Climate — Uses GHG Protocol as foundation
• TCFD — Requires GHG Protocol-aligned disclosure
• ISO 14064 — Compatible with GHG Protocol
• SBTi — Targets based on GHG Protocol inventory