Carbon Factors Methodology¶

How emission factors are derived for AI inference.

Overview¶

Carbon factors define emissions per unit of activity:

CO₂ (grams) = tokens × factor (gCO₂/token)

OpenClaw provides per-model factors for major providers.

Factor Structure¶

Each factor includes:

Field	Unit	Description
inputCo2PerMillionTokens	gCO₂eq	Input processing emissions
outputCo2PerMillionTokens	gCO₂eq	Output generation emissions
cacheReadCo2PerMillionTokens	gCO₂eq	Cache read emissions
waterMlPerMillionTokens	mL	Water for cooling
confidence	0-1	Data quality indicator
source	string	Factor derivation method

Derivation Methodology¶

Step 1: Model Size Estimation¶

Estimate parameter count from model name and behavior:

Model Class	Estimated Parameters
Large (Opus, GPT-4)	200B+
Medium (Sonnet, GPT-4o)	50-200B
Small (Haiku, Mini)	7-20B

Step 2: Energy per Token¶

Based on academic research:

Energy (J/token) = f(parameters, architecture)

Key papers: - Lacoste et al. (2019) — ML CO2 Impact methodology - Patterson et al. (2022) — Large model training energy - Luccioni et al. (2024) — Inference energy measurements

Step 3: Output Multiplier¶

Output tokens require ~3x more energy than input:

output_factor = input_factor × 3

Rationale: - Each output token requires full forward pass - Input tokens can be batched/parallelized - Autoregressive generation is inherently sequential

Step 4: Grid Carbon¶

Convert energy to CO₂:

CO₂ = Energy × Grid_Carbon × PUE

Where: - Grid carbon: gCO₂/kWh (default 400) - PUE: Power Usage Effectiveness (default 1.2)

Step 5: Cache Adjustment¶

Cache reads are ~10% of input processing:

cache_factor = input_factor × 0.1

Default Factors¶

Anthropic Models¶

Model	Input	Output	Cache	Confidence
claude-opus-4	400	1200	40	0.25
claude-sonnet-4	150	450	15	0.30
claude-haiku-4	30	90	3	0.35

OpenAI Models¶

Model	Input	Output	Cache	Confidence
gpt-4o	200	600	20	0.30
gpt-4o-mini	40	120	4	0.35
o1	500	1500	50	0.20

Fallback Factors¶

For unknown models:

Size Class	Input	Output	Cache	Confidence
Large	300	900	30	0.15
Small	50	150	5	0.15

Academic Sources¶

Lacoste et al. (2019)¶

"Quantifying the Carbon Emissions of Machine Learning"

• Introduced ML CO2 Impact Calculator
• Established GPU power consumption baselines
• Provided training-to-inference scaling

Patterson et al. (2022)¶

"Carbon Emissions and Large Neural Network Training"

• Measured large model training energy
• Documented efficiency improvements over time
• Provided parameter-to-energy relationships

Luccioni et al. (2024)¶

"Power Hungry Processing: Watts Driving the Cost of AI Deployment"

• Direct inference measurements
• Compared model architectures
• Quantified efficiency variations

Li et al. (2023)¶

"Making AI Less Thirsty"

• Water consumption analysis
• Cooling requirements
• Regional variations

Assumptions¶

Hardware¶

• Modern inference hardware (H100/A100 class)
• Mixed precision (FP16/BF16) inference
• Typical batch sizes

Infrastructure¶

• Cloud data center
• PUE 1.2 (hyperscaler average)
• Shared infrastructure (not dedicated)

Grid Carbon¶

• Default: 400 gCO₂/kWh (world average)
• Configurable per deployment
• Future: Real-time regional data

Limitations¶

No Supplier Data¶

Providers don't publish per-request emissions. Factors are estimates.

Architecture Unknown¶

Exact model architectures aren't disclosed. Estimates use heuristics.

Location Unknown¶

Data center locations (and thus grid carbon) unknown. Uses averages.

Efficiency Changes¶

Hardware and software improvements change factors over time.

Factor Updates¶

Factors should be reviewed:

1. Annually — New research and measurements
2. On model release — New models may have different efficiency
3. On provider disclosure — If providers publish data

Custom Factors¶

Override factors with better data:

{
  "green": {
    "factorOverrides": {
      "anthropic:claude-sonnet-4": {
        "inputCo2PerMillionTokens": 120,
        "outputCo2PerMillionTokens": 360,
        "confidence": 0.6,
        "source": "measured"
      }
    }
  }
}

Future Improvements¶

Supplier-Specific Data¶

If providers publish emission factors: - confidence → 0.8+ - source → "supplier-specific"

Real-Time Grid Carbon¶

With Electricity Maps API: - Regional grid intensity - Time-of-day variations - Renewable energy tracking

Hardware-Specific Factors¶

Different hardware has different efficiency: - TPU vs GPU - Inference-optimized chips - Quantized models