Sources & Method

Sources and Method

This page shows the current row values used in the calculator, the main sources behind them, and the method notes that explain how the totals are being built. If you want a quick read, start with the activity reference, then use the source list and method sections below to see where the page separates cloud-side from total-system energy, direct from total water, and direct benchmarks from site-built estimates. The point is not exact personal accounting. It is to make it easier to see what the calculator is assuming and how those assumptions shape the comparisons on the site.

Calculator methods

Live row values

Calculator activity reference

Activity	Unit	Status	Server + network	Total-system	Direct water	Total water	Sources
AI text prompts Uses an everyday text-prompt estimate of about 0.3 Wh while keeping a separate prompt-specific water estimate of 0.26 mL direct and 1.3 mL total.	prompts Basic text questions	Estimated	0.3 Wh	0.3 Wh	0.26 mL	1.3 mL	Epoch AI (2025) Jegham et al. (2025)
High-reasoning prompts Uses the long high-reasoning 33.8 Wh benchmark rather than the much smaller medium-query comparison value, so this row stays an intentional outlier.	prompts Long, high-reasoning queries	Inferred	33.8 Wh	33.8 Wh	33.8 mL	202.8 mL	Jegham et al. (2025) Internal synthesis (2026)
AI image generation Uses an open-model image-generation average and the same water-conversion rule used across the broader comparison set.	images One generated image	Inferred	0.48 Wh	0.48 Wh	0.48 mL	2.88 mL	Roucy-Rochegonde et al. (2025) Internal synthesis (2026)
Coding-agent use Uses the current one-hour coding-agent benchmark and adds a 30 Wh device allowance in the total-system column.	hours Hours of sustained agent use, not prompts	Estimated	325 Wh	355 Wh	325 mL	1950 mL	Couch (2026) Internal synthesis (2026)
TikTok or short-video scrolling Built from device-side social-video measurement plus streaming-style infrastructure assumptions; keep this row estimated until a platform-specific cloud-side benchmark exists.	hours Video-heavy social feed	Estimated	7.5 Wh	36 Wh	7.5 mL	45 mL	Greenspector (2021) Holter (2025) Kamiya (2020) Internal synthesis (2026)
Instagram scrolling This remains an analogy-based estimate built from social-media measurement plus streaming comparisons, useful for relative scale rather than platform accounting.	hours Mixed image and short-video feed	Estimated	5 Wh	25 Wh	5 mL	30 mL	Greenspector (2021) Kamiya (2020) Internal synthesis (2026)
Snapchat Another low-confidence social-media estimate built from broader comparison logic rather than a platform disclosure.	hours Short-form messaging and video	Estimated	4 Wh	20 Wh	4 mL	24 mL	Greenspector (2021) Internal synthesis (2026)
YouTube watching Uses the same CDN-style video benchmark as Netflix so the calculator keeps an apples-to-apples streaming anchor for the cloud-side and total-system columns.	hours Streaming video hour	Inferred	22 Wh	26 Wh	22 mL	132 mL	Kamiya (2020) Internal synthesis (2026)
Netflix streaming Uses the IEA/Kamiya 22 Wh cloud-side and 77 Wh total benchmark used throughout the site's streaming comparisons.	hours One hour streamed	Inferred	22 Wh	37 Wh	22 mL	132 mL	Kamiya (2020) Internal synthesis (2026)
Email, cloud docs, and browsing This is a teaching bundle for light browsing, email, and cloud documents rather than a per-click benchmark.	daily blocks A light mixed-work block	Estimated	4 Wh	30 Wh	4 mL	24 mL	Internal synthesis (2026)
Zoom as participant Uses a one-hour participant meeting estimate in which network plus platform routing form the cloud-side figure and participant device energy is added back for the total-system column.	hours Joining a class or meeting	Estimated	54 Wh	74 Wh	54 mL	324 mL	Mytton (2023) Verdecchia et al. (2022) Internal synthesis (2026)
Zoom as host Uses a hosted Zoom estimate that counts all participant devices in the total-system column.	meetings One 50-minute, 10-person hosted session	Estimated	192 Wh	358 Wh	192 mL	1152 mL	Mytton (2023) Verdecchia et al. (2022) Internal synthesis (2026)

Source list

Calculator sources

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

Source	Used for	Notes
How Much Energy Does ChatGPT Use? Epoch AI (2025) · Tier 3 · Expert analysis	0.3 Wh AI text-prompt row prompt-specific comparison language in the boundary and water method sections	Used for the 0.3 Wh everyday text-prompt estimate and related prompt-volume framing.
How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference Jegham et al. (2025) · Tier 1 · Peer-reviewed or preprint benchmark	33.8 Wh high-reasoning prompt row direct and total water framing for prompt-heavy rows maintenance note preserving the long high-reasoning benchmark	Primary inference benchmark for reasoning-heavy prompt classes and environmental multipliers.
Digital inventory calculator and scenario method Internal synthesis (2026) · Tier 3 · Internal synthesis	scenario-style 1 mL direct to 6 mL total water bridge browsing bundle row preset mix logic total-system add-ons for non-AI rows	Site-built method model that combines prompt benchmarks, streaming comparisons, meeting estimates, and water-conversion rules into the live calculator.
AI, Data Centers and Energy Demand: Reassessing and Exploring the Trends Roucy-Rochegonde et al. (2025) · Tier 2 · Policy report	image-generation benchmark row comparison logic in the cloud-side versus total-system bridge	Useful for prompt comparison tables and the Jevons framing, but should be paired with Tier 1 sources where possible.
Electricity use of AI coding agents Couch (2026) · Tier 3 · Expert analysis	coding-agent use row Gen Z vibe-coder preset cloud-side versus total-system bridge for high-intensity AI use	Useful for the vibe-coding scenario but should remain marked as estimated and napkin-math-derived.
What is the environmental footprint for social media applications? 2021 Edition Greenspector (2021) · Tier 2 · Industry measurement	TikTok, Instagram, and Snapchat directional rows cloud-side versus total-system bridge for social-video estimates maintenance note on low-confidence social rows	Used for device-side social-media energy measurements, especially TikTok.
Why Your ChatGPT Prompt Uses Half the Energy of a TikTok Video Holter (2025) · Tier 4 · Opinion or blog analysis	TikTok total-system estimate cross-check	Directional comparison used only as a low-confidence reference for TikTok total-system estimates.
The carbon footprint of streaming video: fact-checking the headlines Kamiya (2020) · Tier 1 · Institutional commentary	Netflix row YouTube row streaming-side values used to anchor TikTok and Instagram estimates	Used for the 77 Wh total streaming baseline and the ~22 Wh server-plus-network split adopted for Netflix and YouTube comparisons.
Zoom, video conferencing, energy, and emissions Mytton (2023) · Tier 3 · Expert analysis	Zoom participant row Zoom host row hosted Zoom derivation table worker-hosted-Zoom preset	Used for Zoom scenario modeling; should stay labeled as an estimate rather than a direct measurement of every meeting type.
Video Calls in Mobile Applications: Energy Consumption and Performance Analysis Verdecchia et al. (2022) · Tier 1 · Peer-reviewed paper	participant device add-on for Zoom total-system device add-on in the hosted Zoom derivation	Used for device-side video-call energy and camera-on versus camera-off impacts in the Zoom estimates.
Measuring the environmental impact of delivering AI at Google Scale Schneider (2025) · Tier 1 · Research paper	water methodology context on direct cooling versus broader electricity water	Primary source for WUE framing, cooling tradeoffs, and provider-scoped production prompt metrics.
2024 United States Data Center Energy Usage Report Lawrence Berkeley National Laboratory (2024) · Tier 1 · National laboratory report	water methodology context on electricity-side water intensity	Primary U.S. demand and water-intensity context for the macro-risk pages.

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, reasoning, and image rows currently keep server and total-system energy equal because the live calculator does not model a separate end-user device add-on for those actions.
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.

Current bridge rules used in the calculator

Row family	Cloud-side basis	Total-system basis	Current calculator result	Maintenance note
AI text / reasoning / image prompts	Inference benchmarks used for everyday prompting	Kept equal to cloud-side in the live calculator	0.3/0.3, 33.8/33.8, 0.48/0.48	No separate end-user device model is currently applied
Coding-agent use	High-intensity coding-agent benchmark	Adds a 30 Wh device allowance	325/355 Wh per hour	Keep this tied to the current coding-agent estimate, not to prompt counts
Netflix / YouTube	22 Wh per hour server + network	77 Wh per hour total-system benchmark	22/77 Wh per hour	Uses the IEA/Kamiya streaming split already adopted on the site
TikTok	Estimated 7.5 Wh per hour cloud-side	Estimated 36 Wh per hour total-system	7.5/36 Wh per hour	Built from social-video measurement plus streaming-style inference
Instagram	Estimated 5 Wh per hour cloud-side	Estimated 25 Wh per hour total-system	5/25 Wh per hour	Analogy-based row; keep estimated status
Snapchat	Estimated 4 Wh per hour cloud-side	Estimated 20 Wh per hour total-system	4/20 Wh per hour	Lowest-confidence social-media row in the calculator
Email / cloud docs / browsing	Internal light-work bundle	Bundle with device add-on	4/30 Wh per daily block	Not a per-click measurement
Zoom as participant	One-hour participant estimate	Adds participant device energy	54/74 Wh per hour	One-hour meeting framing
Zoom as host	Hosted meeting derivation used on the site	Adds all participant devices	192/358 Wh per meeting	One 50-minute hosted session with 10 people

Sources Epoch AI (2025) Jegham et al. (2025) Roucy-Rochegonde et al. (2025) Kamiya (2020) Greenspector (2021) Mytton (2023) Verdecchia et al. (2022) Couch (2026) Internal synthesis (2026)

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.

Water conversion rules used by the live calculator

Row family	Direct water	Total water	Basis	Maintenance note
Text prompts	0.26 mL per prompt	1.3 mL per prompt	Prompt-specific water estimate	Keep this special case unless the underlying prompt benchmark changes
Reasoning prompts	33.8 mL per prompt	202.8 mL per prompt	33.8 Wh energy row with a 1:6 water bridge	33.8 x 6 = 202.8
Image generation	0.48 mL per image	2.88 mL per image	0.48 Wh image benchmark with a 1:6 water bridge	0.48 x 6 = 2.88
Social media / streaming / browsing rows	1 mL per Wh	6 mL per Wh	Rounded direct-plus-indirect water bridge	Used for TikTok, Instagram, Snapchat, YouTube, Netflix, and browsing
Zoom participant / host	1 mL per Wh	6 mL per Wh	Rounded direct-plus-indirect water bridge	Used for both meeting rows
Coding-agent use	325 mL per hour	1,950 mL per hour	1:6 water bridge	325 x 6 = 1,950

Sources Schneider (2025) Lawrence Berkeley National Laboratory (2024) Epoch AI (2025) Jegham et al. (2025) Roucy-Rochegonde et al. (2025) Internal synthesis (2026)

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.

One hosted Zoom meeting with 10 participants over 50 minutes

Energy component	Per-meeting energy (Wh)	Basis
Zoom server routing	7	SFU-style routing estimate
Network across all participants	185	Mytton energy-per-GB framing
Devices across all participants	166	Laptop-energy assumption used for participant devices
Server + network subtotal	192	Cloud-side figure used in the calculator
Total-system energy	358	Hosted meeting figure used in the calculator