AI Footprint
What does an AI workload actually cost? Plug in your tokens, model class, and region — get an honest estimate of dollars, grid energy, CO₂e, cooling water, and the resulting permanent contribution to global mean temperature (IPCC AR6 TCRE). Comparisons to cars, flights, households, trees, and the remaining 1.5 °C budget keep the numbers human-scale. Built for both "is my prototype's bill reasonable" and "what if someone ran four quintillion tokens on frontier models?"
Advanced (optional)
How it works
You describe a workload (tokens × model class × period × region). The
AI model that powers the rest of this site computes five parallel
estimates: cost (USD, from public list pricing),
energy (kWh, from per-token measurements published by
Google, Meta, and academic studies), carbon (CO₂e,
using region-specific grid intensity from the IEA and EPA),
water (litres, including both direct cooling and
indirect thermoelectric power-plant cooling), and
global temperature contribution (°C, applying the
IPCC AR6 Transient Climate Response to Cumulative Emissions of
0.45 °C per 1000 GtCO₂). It then maps the result to visceral
comparisons — cars, flights, households, trees — plus reduction
ideas with effort-vs-impact ratings. Schema
ai-footprint.v1, storage aiapp.aifootprint.v1
— no account, no cloud storage.
Grounding numbers (2024–2026 public data) baked into the system prompt: frontier pricing $3–15 input / $15–75 output per million tokens; workhorse $0.20–1 / $0.50–3; inference energy ~0.001–0.003 kWh per 1k tokens (frontier), ~0.0001–0.0005 (edge); grid CO₂e 50 g/kWh (Nordic) → 700 g/kWh (India); cooling water 0.5–3 L/kWh depending on datacenter type and power-plant mix. The model also amortizes training and embodied hardware carbon — usually trivial per-token at scale, but it's surfaced honestly.
Honest caveat: these are educational estimates, not an audit. Real costs depend on volume discounts, prompt caching, and your contract; real energy and carbon depend on datacenter PUE, GPU mix, model architecture, and the exact regional grid hour-by-hour. Uncertainty ranges of ±50–200% are normal for this kind of bottom-up estimate. Use this to build intuition, not to compute your sustainability report.