Kaneda Consulting · AI Resource Calculator

How thirsty is
your AI session?

Enter your session token totals to estimate the water and energy your conversation consumed. Energy figures are drawn from peer-reviewed research (2025–2026). Water figures combine independent research with provider self-reported infrastructure data (Amazon, Microsoft, and Google environmental reports). Google is the only major AI provider to have directly measured and published its own production energy data; energy estimates for Claude and ChatGPT are derived from independent research. Since models continue to improve, actual current consumption is likely lower than these figures.

01 — Select AI assistant
Gemini Google Infrastructure Nearest: The Dalles, OR
Claude Anthropic / AWS Nearest: Boardman, OR
ChatGPT OpenAI / Azure Nearest: Quincy, WA
02 — Session token totals
optional — shows per-query average
Session totals

💧 Water consumed

—

⚡ Energy drawn

—

per-query average

💧 Avg water / query

—

⚡ Avg energy / query

—

Parameters in use
Methodology

Gemini directly measured

// Elsworth et al. (2025) — Google production instrumentation Energy = 0.24 Wh × (output_tokens / 300) // Scaled from fleet median (~300 output tokens). No variant breakdown published. // Water uses derivation formula below — Elsworth measured on-site only (0.26 mL/300 tokens)

Claude / ChatGPT

Energy floor Oviedo et al. 2026

// Production-scale H100, continuous batching. Joule peer-reviewed. E_floor = 0.34 Wh × (output_tokens / 300) // Corroborated by Sam Altman disclosure, June 2025

Energy ceiling Jegham et al. 2025

// API latency + TPS benchmarks. Flagship per provider: // Claude → Claude 3.7 Sonnet (j_s 0.950 Wh) | ChatGPT → GPT-4o (j_s 0.423 Wh) // Oviedo argues this overstates production energy by 4–20× // Current models (Claude 4.x, GPT-5) not yet in peer-reviewed literature E_ceil = j_s × (output_tokens / 300) // j_s = short-query (~300 token) measured value, used as rate for session-level scaling

Water derivation — all providers Jegham et al. 2025 Eq. 4

Water (mL) = [(E_kwh / PUE) × WUE_site + E_kwh × WUE_source] × 1000 // Gemini: PUE 1.09 | WUE_site 0.96 (Google 2025 Env. Report, 2024 data) | WUE_source 5.11 (Oregon grid / WRI) // Claude: PUE 1.14 | WUE_site 0.12 (Amazon 2025) | WUE_source 5.11 (Oregon grid / WRI) // ChatGPT: PUE 1.12 | WUE_site 0.27 (Microsoft 2026) | WUE_source 4.35 (WA grid / WRI)
Sources

Gemini (directly measured): Elsworth et al. (2025) "Measuring the environmental impact of delivering AI at Google Scale." arXiv:2508.15734. The only directly measured, published, production-scale figure for any major AI assistant. Fleet-wide — no variant breakdown available. Energy baseline (0.24 Wh / 300 tokens) and on-site water figure (0.26 mL / 300 tokens) from Elsworth et al. WUE_site of 0.96 L/kWh and PUE of 1.09 derived from Google 2025 Environmental Report (2024 data): data center water consumption 7,787M gallons ÷ data center electricity 30,825,600 MWh. Google does not publish WUE directly — this is back-calculated from fleet-wide totals.

Claude / ChatGPT energy floor: Oviedo et al. (2026) "Energy Use of AI Inference: Efficiency Pathways and Test-Time Scaling." Joule. DOI: 10.1016/j.joule.2026.102430. Frontier model production median 0.34 Wh per 300 output tokens. Corroborated by Sam Altman, June 2025.

Claude ceiling — Claude 3.7 Sonnet: Jegham et al. (2025) "How Hungry is AI?" arXiv:2505.09598v6 Table 4. Short: 0.950 Wh | Medium: 2.989 Wh | Long: 5.671 Wh.

ChatGPT ceiling — GPT-4o: Jegham et al. (2025) Table 4. Short: 0.423 Wh | Medium: 1.215 Wh | Long: 2.875 Wh.

Water formula — all providers: Jegham et al. (2025) Eq. 4. Applied to all three providers for a fair comparison. Gemini's published figure (Elsworth et al., 0.26 mL/300 tokens) covers on-site cooling only; upstream electricity generation water is added here using the same method as Claude/ChatGPT. Gemini WUE_site 1.15 L/kWh from Elsworth et al. Gemini PUE 1.09 = Google fleet-wide average. Gemini WUE_source 5.11 L/kWh = Oregon regional grid (WRI methodology), same grid as AWS/Boardman. AWS PUE 1.14, WUE_source 5.11; Azure PUE 1.12, WUE_source 4.35 — all from Jegham et al. Table 1.

AWS WUE_site (Claude): Amazon (2025) "Amazon data center water usage." Self-reported 0.12 L/kWh in 2025, a 52% improvement since 2021.

Azure WUE_site (ChatGPT): Microsoft (2026) "Inside Microsoft's two-decade push to cut water intensity while scaling for growth." blogs.microsoft.com, June 24 2026. Self-reported 0.27 L/kWh for owned datacenter fleet in 2025.

Data gap: Published benchmarks cover models through late 2025. Current models (Claude 4.x, GPT-5, Gemini 2.x) not yet independently measured. Efficiency trends suggest actual figures are lower.