impact.py

"""
This estimate will evolve in time. Cloud computing is not transparent about inference impacts
and a lot of factors can influence C02 emissions. We found this article to be insightful:
    https://towardsdatascience.com/is-googles-reveal-of-geminis-impact-progress-or-greenwashing/

We provide multiple estimate of C)2 emissions and energy usage. If you come across a more accurate 
way of estimating LLM inference impacts, specifically for Google Cloud infrastructre, we'd love 
to hear about it and incorporate it into these estimates.

Here we only include estimate of Google Cloud because that is the infrastructure powering 
the VoucherVisionGO API. 

Assumptions:
- For a typical herbarium specimen
        OCR: 
            Total  = 2500 tokens
            Input  = 2000 tokens
            Output =  500 tokens
        LLM Parsing:
            Total  = 2500 tokens
            Input  = 2000 tokens
            Output =  500 tokens
        Overall:
            Total  = 5000 tokens

- For Notebook transcription the results vary significantly.
- Water use is extremely complex and varies signifcantly by data center. We rely on Google's estimate.
- We assume the "cost" of input and output tokens are equivalent, though "thinking" models may differ.

Estimates:
- source_google
    Original: https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference
    Archive: https://web.archive.org/web/20251001194210/https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference
    Google researchers, not peer reviewed

- source_oviedo
    Original: https://arxiv.org/pdf/2509.20241
    Microsoft researchers, not peer reviewed


https://web.archive.org/web/20251109020239/https://assets.publishing.service.gov.uk/media/62aee1fbe90e0765d523ca33/2022-ghg-cf-methodology-paper.pdf?utm_source=chatgpt.com
"""
import json

# Derived from Google disclosure: 0.24 Wh → 0.03 g CO₂e, 0.26 mL water
EMISSIONS_INTENSITY_G_PER_WH = 0.03 / 0.24      # 0.125 g CO₂e per Wh
WATER_INTENSITY_ML_PER_WH    = 0.26 / 0.24      # 1.083 mL water per Wh

def estimate_impact(tokens: float | None) -> dict:
    """
    Estimate AI inference environmental impact (energy, CO₂, water)
    combining:
      - Google (Aug 2025): active vs. total (fixed per prompt)
      - Oviedo et al. 2025: token-scaled linear interpolation
    Defaults to 5000 tokens if None, 0, or negative.
    """
    # Default token fallback
    if tokens is None or tokens <= 0:
        tokens = 5000

    # --- GOOGLE (fixed per median Gemini prompt) ---
    google_active = {"watt_hours": 0.10, 
                     "grams_CO2": 0.02, 
                     "mL_water": 0.12}
    google_total  = {"watt_hours": 0.24, 
                     "grams_CO2": 0.03, 
                     "mL_water": 0.26}

    source_google = {
        "estimate_low":  google_active,
        "estimate_high": google_total,
        "notes": ("Low - Active compute only (TPU/GPU). High - Full system CPU, RAM, idle, and PUE of 1.09. From Google Cloud Blog Aug 2025."),
        "url_source": ("https://web.archive.org/web/20251001194210/https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference"),
    }

    # --- OVIEDO et al. (2025): token-scaled estimates ---
    t_low, t_high = 300.0, 5000.0
    low_wh_300,  high_wh_300  = 0.18, 0.67
    low_wh_5000, high_wh_5000 = 2.38, 7.38

    def lerp(x1, y1, x2, y2, x):
        return y1 + (y2 - y1) * (x - x1) / (x2 - x1)

    low_wh  = lerp(t_low, low_wh_300,  t_high, low_wh_5000,  tokens)
    high_wh = lerp(t_low, high_wh_300, t_high, high_wh_5000, tokens)

    source_oviedo = {
        "estimate_low": {
            "watt_hours": low_wh,
            "grams_CO2": low_wh * EMISSIONS_INTENSITY_G_PER_WH,
            "mL_water": low_wh * WATER_INTENSITY_ML_PER_WH,
        },
        "estimate_high": {
            "watt_hours": high_wh,
            "grams_CO2": high_wh * EMISSIONS_INTENSITY_G_PER_WH,
            "mL_water": high_wh * WATER_INTENSITY_ML_PER_WH,
        },
        "notes": ("Linear interpolation between 300 and 5000 tokens per Oviedo et al. (2025). Water scaled using Google est. 1.083 mL per Wh ratio."),
        "url_source": ("https://arxiv.org/pdf/2509.20241"),
    }

    # --- Overall means from the 4 estimates (Google low/high + Oviedo low/high) ---
    all_wh = [
        source_google["estimate_low"]["watt_hours"],
        source_google["estimate_high"]["watt_hours"],
        source_oviedo["estimate_low"]["watt_hours"],
        source_oviedo["estimate_high"]["watt_hours"],
    ]
    all_co2 = [
        source_google["estimate_low"]["grams_CO2"],
        source_google["estimate_high"]["grams_CO2"],
        source_oviedo["estimate_low"]["grams_CO2"],
        source_oviedo["estimate_high"]["grams_CO2"],
    ]
    all_water = [
        source_google["estimate_low"]["mL_water"],
        source_google["estimate_high"]["mL_water"],
        source_oviedo["estimate_low"]["mL_water"],
        source_oviedo["estimate_high"]["mL_water"],
    ]

    est_wH = sum(all_wh) / len(all_wh)
    est_CO2 = sum(all_co2) / len(all_co2)
    est_w = sum(all_water) / len(all_water)

    return {
        "total_tokens_all": tokens,
        "estimate_grams_CO2": est_CO2,   # grams CO2e (mean)
        "estimate_watt_hours": est_wH,     # watt-hours (mean)
        "estimate_mL_water": est_w,       # milliliters water (mean)
        "source_google": source_google,
        "source_oviedo": source_oviedo,
    }

if __name__ == "__main__":
    for n in [5000]:
        print(json.dumps(estimate_impact(n), indent=2))

    # Example usage
    # for n in [5000, None, 0, -500, 100, 2000, 10000]:
    #     print(json.dumps(estimate_impact(n), indent=2))