Environmental Impact Tracking — Sources & Methodology

Infrastructure Multipliers

Source: Jegham et al. Table 1, supplemented with provider sustainability reports (Google, Microsoft, AWS, Scaleway) and independent verification.

PUE Cross-Validation

WUE Cross-Validation

Primary Energy & Abiotic Depletion Lifecycle Factors

Source: EcoLogits / ADEME Base Empreinte® / ecoinvent electricity lifecycle data.

These factors convert electricity consumption (Wh) into lifecycle impact metrics, accounting for the full supply chain of electricity generation including fuel extraction, processing, transport, and infrastructure.

Primary Energy (PE)

Measures the total energy extracted from natural resources (fossil fuels, nuclear, renewables) required to produce the electricity consumed. Includes extraction, refining, transport, and generation losses. A factor of ~0.0097 MJ/Wh means roughly 2.7× the direct electricity is consumed as primary energy from nature.

Abiotic Depletion Potential for Elements (ADPe)

Measures the depletion of non-renewable mineral and metal resources (lithium, copper, gold, rare earths) required for the electricity generation infrastructure. Expressed in kg antimony equivalent (kg Sb eq) per the CML-IA characterization method. France has lower ADPe than the US grid due to nuclear-dominated generation requiring less diverse mineral extraction.

Per-Model Energy (Wh per 1,000 Tokens)

Cross-Validation of Energy Estimates

Multiple independent sources converge on a 0.2–0.4 Wh range for a standard text query to a frontier model, with reasoning models consuming 5–70× more depending on complexity.

Convergence Table: Standard Text Queries

Convergence Table: Reasoning Models

Jegham et al. Full Model Benchmark (Short Prompt: 100 input / 300 output tokens)

Energy Breakdown by Component (Google 2025)

Google's technical paper (arXiv:2508.15734) provides the only published component-level energy breakdown for production AI inference:

Output vs Input Token Energy Ratio

Multiple sources confirm that output tokens are significantly more energy-intensive than input tokens:

The 5:1 ratio used as our baseline (from Couch/API pricing) is a conservative estimate — direct measurement suggests the true ratio may be higher.

Real-World Comparison References

Used in the tooltip to provide human-understandable context for each metric.

Energy Comparisons

Logic: < 120s → "≈ Xs streaming video"; < 50 searches → "≈ X Google searches"; else → "≈ X% phone charge"

Water Comparisons

Logic: < 100 drops → "≈ X water drops"; else → "≈ X teaspoons"

CO₂ Comparisons

Note on car CO₂: The 130 g/km value is an estimate for the current European on-road fleet (all vehicle ages). EEA reported the 2024 new-car fleet average at 108 g CO₂/km, down from 160–170 g/km in 2015–2019 due to EV adoption and EU emissions standards (Regulation 2019/631). The total on-road fleet average is higher because older, less efficient vehicles remain in service (average European car age ~12 years). EU 2025 target: 93.6 g/km; 2030: 49.5 g/km; 2035: 0 g/km (zero tailpipe). The EPA US fleet average is ~249 g/km — if a US-only comparison were needed, the value should be 0.249 gCO₂/m.

Logic: < 100m → "≈ car driving Xm"; else → "≈ X Google searches"

Primary Energy Comparisons

Logic: < 50 matches → "≈ X matches burned"; < 100 Cal → "≈ X food Calories"; else → "≈ boiling X cups of water"

ADPe Comparisons

ADPe is expressed in kg Sb eq (antimony equivalent), an abstract characterization factor. To make it tangible, we convert to the equivalent mass of copper that would need to be mined to cause the same mineral depletion, using the CML-IA characterization factor for copper (1.4 × 10⁻³ kg Sb eq/kg Cu).

Logic: Convert kg Sb eq → equivalent copper mass: copperKg = adpKgSb / 1.4e‑3. Display as mg or g of copper mined (mg minimum for readability).

Context: Manufacturing one smartphone (1.25 × 10⁻³ kg Sb eq, Fairphone 5 LCA) causes the same mineral depletion as generating ~3,000 kWh of electricity — roughly one year of average European household electricity. A typical LLM query's ADPe equates to mining 0.001–0.05 mg of copper.

Additional Reference Values (not used in tooltip)

Uncertainty Methodology

There is currently no single standardized uncertainty methodology for LLM energy estimation, though progress is being made. The SCI for AI specification (ISO/IEC 21031:2024, ratified December 2025 by the Green Software Foundation) provides standardized functional units, system boundaries, and reporting requirements for AI carbon accounting — a critical step toward comparable, reproducible environmental impact reporting. Different sources currently use different approaches:

Our Uncertainty Ranges

Each model in our system defines uncertaintyMin and uncertaintyMax multipliers (e.g., 0.5–1.5 means the true value could be 50%–150% of the nominal estimate). These ranges are applied uniformly to all 5 metrics.

Rationale for ±30–50% default range:

• Jegham et al. error bars range from ±15% (well-known models like GPT‑4o: 0.42 ± 0.13 = ±31%) to ±52% (o3: 7.03 ± 3.66)
• Oviedo/Microsoft IQR spans roughly ±50% of median (0.18–0.67 on 0.34)
• Caravaca et al. found batch size alone causes 36× variation (Llama 405B: 21.7 Wh single vs 0.6 Wh batched), but production systems always batch
• Nature Scientific Reports 2024 identifies hardware, geography, and utilization as the primary uncertainty sources
• Non-production estimates overstate by 4–20× (Oviedo/Microsoft), suggesting our academic-derived estimates may be conservatively high

Models with more data points (GPT‑4o, Claude 3.7 Sonnet) have tighter ranges; models extrapolated from architectural reasoning (Claude Opus 4, Gemini 3 Pro) have wider ranges.

Key Caveats & Limitations

1. Limited official data. Only Google has published official per-query energy data (0.24 Wh for Gemini, August 2025). Microsoft Research (Oviedo et al.) provides the next most rigorous estimate (0.34 Wh median for frontier models). OpenAI's single data point (0.34 Wh for ChatGPT) came from a CEO statement without methodology. All other per-token figures are academic estimates with ~20–50% margin of error.
2. Hidden reasoning tokens. Reasoning models (o3, o4-mini) generate hidden chain-of-thought tokens that dramatically increase actual energy consumption. Jegham et al. measured o3 at 7.03 Wh (short) to 39.2 Wh (long) — up to 70× more than efficient models. Oviedo/Microsoft found test-time compute causes a 13× energy increase. Kumar et al. (OverThink) showed adversarial attacks can inflate reasoning chains by 18–46×, proportionally increasing energy. The internal reasoning tokens are not visible via the API but consume compute. This makes per-visible-token metrics potentially misleading for these models.
3. "100% renewable" claims. AWS and Microsoft use annual Renewable Energy Certificate (REC) matching — purchasing certificates equal to annual consumption from any location and time period. This is an accounting solution, not an engineering one: a REC from a solar farm in Arizona at noon can "offset" fossil-fueled consumption in New York at midnight. Peer-reviewed research found REC-based claims lead to "an inflated estimate of the effectiveness of mitigation efforts." Google's 66% CFE (carbon-free energy, measured hourly on the same regional grid) is more conservative and transparent. Greenpeace found AWS meeting only 12% of its renewable commitment physically in Virginia; grid actual renewable is <5%. Dominion Energy (Virginia) mix: ~33% nuclear, ~33% gas, ~25% coal, ~4–6% renewable.
4. GreenPT transparency gap. GreenPT has not published per-token energy figures despite marketing as "green AI." Their primary environmental advantage comes from running on the French nuclear grid (21.7 gCO₂/kWh direct in 2024, ~27 gCO₂eq/kWh lifecycle in 2025 per RTE France — among the lowest in the world) and Scaleway's efficient data centers (PUE 1.25, adiabatic cooling), rather than from demonstrated model-level efficiency.
5. Perplexity: no sustainability data. No environmental reports, no energy consumption figures, no climate commitments. Their infrastructure multipliers are assumed from AWS (their primary cloud provider). The additional energy cost of web search + retrieval in each query is estimated, not measured.
6. Model version drift. Jegham et al. benchmarked Claude 3.7 Sonnet and o3, not Claude Opus 4.6 / Sonnet 4.6 or GPT‑4.1. Our per-token estimates for current models are derived from the closest benchmarked equivalents and architectural reasoning. As of February 2026, no published paper provides direct energy measurements for GPT‑4.1, Claude Opus 4.6, Claude Sonnet 4.6, or Gemini 3 Pro. Sonnet 4.6 inherits Sonnet 4.5's energy factors based on identical API pricing ($3/$15 per MTok) and similar output throughput (~57 tok/s vs ~63 tok/s).
7. Batch size and utilization. Energy per query varies dramatically with server utilization. Caravaca et al. found 36× reduction from single-prompt to batch-100 for Llama 405B (21.7 → 0.6 Wh). Jegham et al. assumes batch size 8; real production loads vary. Oviedo/Microsoft warns non-production estimates overstate energy by 4–20×.
8. Water accounting. WUE figures include both Scope 1 (on-site cooling: evaporative towers, adiabatic systems) and Scope 2 (off-site: water consumed by electricity generation). Per IEA 2023, two-thirds of total data center water is indirect/off-site. Provider-reported WUE values: AWS 0.15 L/kWh (2024, best-in-class), Microsoft 0.30 L/kWh (FY2024), Google ~1.0 L/kWh (global on-site). Li et al. (CACM 2025) provide the most comprehensive framework for total water footprint estimation.
9. Input/output split. The per-1k-token split between input and output is estimated for most models. Caravaca et al. found output tokens have ~11× greater energy impact than input tokens (direct measurement). Couch 2026 uses a 5:1 ratio based on API pricing. SemiAnalysis suggests ~15× at smaller scales. The true ratio varies by model architecture and is not published by any provider.
10. PE and ADPe precision. Lifecycle factors are grid-level averages from ADEME Base Empreinte / ecoinvent databases, as confirmed by EcoLogits' JOSS paper and methodology documentation. They do not capture provider-specific optimizations (e.g., Google's on-site solar) or time-of-day variations in grid composition. EcoLogits confirms they use ADEME specifically due to "a lack of open, up-to-date alternatives" for PE and ADPe. The CML-IA characterization method (Leiden University) underpins all ADPe calculations; latest revision: van Oers et al. 2020.
11. Jevons Paradox. (Jegham v6) As AI becomes more efficient and cheaper, total resource consumption may actually increase. De Vries (Patterns/Cell Press, 2025) projects AI systems produced 32.6–79.7 million tonnes CO₂ and consumed 312.5–764.6 billion liters of water in 2025 alone. Google's total emissions rose 11% to 11.5M tonnes CO₂ in 2024 (+51% from 2019), mostly Scope 3.
12. France carbon intensity validated by Scaleway. Our CIF of 0.065 gCO₂/Wh (65 gCO₂/kWh) for Scaleway matches their own Environmental Footprint Calculator published value for PAR-2 (DC5). Scaleway uses a location-based methodology per ADEME PCR guidelines with EMBER electricity mix data, deliberately excluding their 100% renewable Guarantees of Origin. This figure is higher than RTE France 2024 grid intensity (21.7 gCO₂eq/kWh direct, 30.2 lifecycle) because ADEME's regulatory average for France (52 gCO₂e/kWh, multi-year) is structurally higher than single-year RTE figures, and the PUE multiplier (1.25) further increases effective carbon per useful kWh.
13. Embodied emissions not included. Our calculations cover only operational energy (use phase). Hardware manufacturing (embodied) emissions are significant — per Boavizta and EcoLogits methodology, the manufacturing phase of server hardware contributes additional GWP, PE, and ADPe that are amortized over the equipment's useful life. Fairphone LCA data shows manufacturing accounts for ~100% of ADPe in consumer electronics; similar dominance applies to data center hardware. NVIDIA's HGX H100 Product Carbon Footprint reports 1,312 kg CO₂e cradle-to-gate per system. TechInsights (2026) projects GPU manufacturing emissions to grow ~16× from 2024 to 2030 due to AI chip demand, underscoring the growing importance of embodied carbon in total AI lifecycle assessments.
14. No provider publishes per-token energy. Google provides per-query data (but not per-token breakdown). All per-token values are derived from per-query benchmarks divided by estimated token counts. Couch 2026 provides the most explicit derivation methodology.

Calculation Formula

For a given model with inputTokens input tokens and outputTokens output tokens:

All multipliers (water, CO₂, PE, ADPe) are infrastructure-level factors applied uniformly to the total energy consumed. Water, CO₂, PE, and ADPe depend on the electricity grid and provider infrastructure, not on the model architecture.

Each metric also has min/max uncertainty estimates derived from the model's uncertaintyMin and uncertaintyMax multipliers (e.g., 0.5–1.5 means the true value could be 50%–150% of the nominal estimate).

Displayed Metrics

The environmental impact tooltip shows 5 metrics per message:

All 5 metrics include real-world comparisons: Energy (streaming video / Google searches / phone charge), GHG (car driving / Google searches), Water (water drops / teaspoons), Primary Energy (matches / food Calories / boiling water), and Abiotic Resources (copper mining equivalent).

The inline footer below each message shows Energy, GHG (CO₂), and Water for compactness (matching tooltip order).