Sources and Method

Live row values

Calculator activity reference

Source list

Calculator sources

This table lists the main sources used in the calculator and the role each one plays in the comparison model.

Cloud-side versus total-system bridge

The calculator keeps the first energy column as a cloud-side comparison view and then adds a second total-system column where the research basis supports it.

• The cloud-side column usually combines server and network demand where the underlying source does not isolate pure data-center energy.
• AI text and reasoning rows currently keep cloud-side and broader-system energy equal because the live calculator does not model a separate end-user device add-on for those actions.
• The image row is narrower: its 1.7 Wh source covers GPU electricity only. That value is carried through both energy columns as a conservative floor and must not be described as measured server-plus-network or total-system energy.
• Media and Zoom rows use a larger bridge from cloud-side to total-system values, which is why those rows show a bigger gap between columns.

Water methodology carried into the calculator

The calculator keeps prompt-specific direct and total water figures where they exist, then uses a broader electricity-to-water bridge for most of the remaining rows.

• The current model rounds direct cooling water to roughly 1 mL per Wh of server energy and total water to roughly 6 mL per Wh once indirect electricity water is added back.
• Text prompts keep their own prompt-specific direct and total water values instead of using the generic 1-to-6 rule.
• Reasoning prompts, image generation, social video, streaming, browsing, and Zoom rows currently follow the broader proportional water rule unless a stronger prompt-specific figure exists.
• The intensive coding-agent request row also uses this shared project scenario: 150 mL direct water and 900 mL broader water per request from its 150 Wh modeled energy value.
• Applying the generic bridge to coding-agent electricity assumes the same average cooling and electricity-water intensities. It is not a provider-specific measurement, does not adopt Jon Ippolito's 7 mL/Wh choice, and adds no new PUE adjustment.
• Jon Ippolito's What Uses More? spreadsheet prompted a review of the generic coefficient. Its 7 mL/Wh selection is retained here as comparative synthesis, not adopted as a measured value.
• Using the spreadsheet's stated 0.3 gCO2/Wh factor, Mistral's 45 mL and 1.14 gCO2e figures imply 11.84 mL/Wh, not 13 mL/Wh. The current 6 mL/Wh broader-water rule therefore remains unchanged pending a boundary-matched source.

Hosted Zoom meeting derivation

The calculator's hosted Zoom row combines a small routing/server term, a larger network term, and participant-device energy.

• The server term is intentionally small because the model assumes SFU-style routing rather than heavy re-encoding.
• The network term dominates the cloud-side figure and the participant devices dominate the total-system add-on.
• This remains an estimate and should stay visibly separate from directly measured prompt benchmarks.

Image-generation evidence and Ippolito crosswalk

The image row now uses a pinned 1.7 Wh IEA benchmark discovered through comparison with Jon Ippolito's What Uses More? source sheet, while keeping Ippolito's 2 Wh figure and other published estimates visible as comparison evidence.

• Jon Ippolito's What Uses More? spreadsheet prompted this review by selecting 2 Wh for a 1024x1024 image and exposing several candidate estimates.
• The calculator does not treat Ippolito's 2 Wh synthesis as a primary measurement. It adopts the underlying IEA report's 1.7 Wh benchmark because the report identifies the model, H100 test hardware, controlled experimental context, and GPU-only boundary.
• The IEA report does not establish a full production-serving or total-system value. The calculator keeps server and total energy equal for this row and labels the result inferred.
• The image water figures are internal 1 mL/Wh direct and 6 mL/Wh broader conversions. None of the image-energy sources measures image-specific water use.

Gaming Console (Xbox Series X + TV) method

The fixed gaming row estimates one active player-hour of a locally rendered 10 to 12-player shooter on Xbox Series X with a reference 54.7-inch television and fixed broadband. It includes an attributional allocation of shared network infrastructure and a low-confidence modeled game-server term.

• This is locally rendered online console play, not Xbox Cloud Gaming. Multiplayer traffic carries game state, voice, authentication, matchmaking, anti-cheat, and telemetry; the console renders the game.
• The central calculator values are rounded because the infrastructure terms are modeled. Exact component arithmetic is retained below for reproducibility.
• The low, central, and high cases are scenario bundles, not statistical confidence intervals.
• The game-server IT coefficient is the weakest component. Activision publishes architectural details but no per-player electricity, host power, utilization, provider share, PUE, or WUE.

Coding-agent request method

The intensive coding-agent row counts human requests and rounds a modeled 149.38 Wh-per-prompt result to 150 Wh, then applies the project's existing inferred direct and broader water conversions.

• Functional unit: one human message sent to a coding agent. Count initial instructions, follow-ups, corrections, and unsuccessful requests. Each human follow-up is a new request. Do not count automatic model or subagent calls separately because they are represented in the usage logs; external tool execution itself is not separately quantified.
• Energy arithmetic: 170,000 Wh / 1,138 typed human prompts = 149.38 Wh per prompt, rounded to 150 Wh per request for the calculator. Repeated model-call usage records were deduplicated by API message ID; the human-prompt denominator is not described as deduplicated.
• The source's roughly 60 to 290 Wh per-prompt sensitivity reflects alternative modeling and cache assumptions. It is not a confidence interval for users and is not used as a generic population range.
• The source workload is one intensive user's Claude Code use. Exact model versions and mix are not fully disclosed, no duration is assigned to a request, and the row does not claim a completed task or comparable outcome.
• Water arithmetic transfers the project's existing AI water scenario: 150 Wh x 1 mL/Wh = 150 mL direct water; 150 Wh x 6 mL/Wh = 900 mL broader water. Broader water includes direct water, so the two values are not added.
• The water transfer assumes the project's average cooling and electricity-water intensities apply to this modeled coding electricity. It is inferred teaching arithmetic, not provider-specific measurement or a coefficient reported by Hausfather or Ippolito.
• Both energy columns remain 150 Wh because no additional end-user device or network energy is quantified. Missing components are outside the boundary rather than zero.