Every estimate is produced by a versioned, deterministic pipeline built on published research — never a black box. The engine can never emit a single number: every output is a range with a confidence score and its reasons.
Low bounds multiply with low bounds and high with high, so the final range honestly compounds every layer of uncertainty.
Everything we assume, why, and where it comes from. When an assumption changes, the version number changes with it.
| Assumption | Value | Why |
|---|---|---|
| Measurement boundary | device × PUE | Host and idle-reserve overhead are not separately counted. Google's fleet measurement shows the comprehensive boundary is ~2.4× the narrow one (0.24 vs 0.10 Wh per median prompt) — so we may undercount, and we say so. |
| Energy / 1k tokens — standard | 0.15–1.2 Wh | Three independent methods converge on ~0.2–0.4 Wh per median chat interaction: Google's fleet measurement, Epoch AI's estimate, and measured academic benchmarks (Luccioni et al., FAccT 2024). |
| Energy / 1k tokens — light | 0.03–0.3 Wh | Nano/mini models measure ~70× below frontier reasoning models in the same benchmark. |
| Energy / 1k tokens — reasoning | 0.4–3.5 Wh | Hidden thinking tokens are invisible to any client. Upper anchor reproduces the highest published long-prompt estimates (o3 ≈ 39 Wh at ~10k tokens). Those figures are inferred and contested — this class is capped at low confidence. |
| Datacenter overhead (PUE) | 1.09–1.6× | Which facility serves a request is unknowable; band spans the best published hyperscaler fleet (Google, 1.09) to the industry average that has been flat for six years (Uptime Institute 2025, 1.54). |
| Tokens per turn (unmeasured) | 150–900 | Used only when no visible-text estimate exists; spans short Q&A to long-form turns. |
| Token estimate from text | ≈ chars ÷ 4 | Standard English heuristic. Hidden reasoning, system, and tool tokens are invisible — real totals may be higher, and confidence is capped accordingly. |
| Carbon display | deferred | Grid factors are stored (41–700 gCO₂/kWh by region) but the serving region is unknowable — a ~17× spread. No per-prompt carbon number is shown until it can be shown honestly. |
| Water display | deferred | Published per-prompt water figures differ ~100× purely on measurement boundary (0.26 mL onsite-only vs 10–50 mL full-cycle). No agreed standard exists; deferring is the honest choice. |
| Review cadence | quarterly | One provider published a 33× efficiency improvement in a single year. Static numbers decay in months; staleness is itself a confidence input. |
| Source | Organization | Year |
|---|---|---|
| Measuring the environmental impact of delivering AI at Google Scale Fleet measurement: 0.24 Wh median prompt; the boundary-problem evidence; 33× one-year efficiency gain. | Google / DeepMind | 2025 |
| How much energy does ChatGPT use? Independent bottom-up estimate: ~0.3 Wh per typical query. | Epoch AI | 2025 |
| Power Hungry Processing Measured open-model inference; the measured floor of our bands. | Luccioni, Jernite, Strubell (FAccT) | 2024 |
| How Hungry is AI? Inferred estimates for 30 closed models; contested; used as upper-bound anchor at low confidence. | Jegham et al. | 2025 |
| AI Energy Score Reproducible measured benchmark for open models. | Hugging Face / Salesforce | 2025 |
| Mistral Large 2 life-cycle assessment First provider LCA; shows how much boundary choice moves numbers. | Mistral AI · Carbone 4 · ADEME | 2025 |
| Global Data Center Survey 2025 Industry-average PUE 1.54, flat for six years. | Uptime Institute | 2025 |
| Fleet PUE disclosure Trailing-twelve-month fleet PUE 1.09. | 2025 | |
| Electricity 2025 — Emissions Global grid intensity ~445 gCO₂/kWh, trending down. | IEA | 2025 |
| Yearly Electricity Data Open per-country grid factors (France ~41 → coal-heavy ~700). | Ember | 2025 |
| Making AI Less Thirsty Foundational onsite + off-site water methodology. | Li, Yang, Islam, Ren (UC Riverside) | 2023 |
| H100 / H200 / B200 specifications TDP anchors; hardware absorbed into model-class bands because serving hardware is unobservable. | NVIDIA | 2024 |
| Tokenizer documentation Basis for the chars ÷ 4 token heuristic. | OpenAI / Anthropic docs | 2024 |
Prototype note: current bands are placeholders within the ranges reported by the cited literature, re-reviewed quarterly. They must pass external research review before any public launch — honesty about that is part of the methodology too.