{ "comparators": [ { "author": "Jon Ippolito", "calculatorUrl": "https://what-uses-more.com/", "evidenceUrl": "https://docs.google.com/spreadsheets/d/1xgvGhd6rmnBAEk9t2lesKlJeUy4TRQHrW6uWDSEVDFY/edit?gid=0#gid=0", "id": "jon-ippolito-what-uses-more", "name": "What Uses More?", "sourceId": "ippolito-what-uses-more" } ], "compatibility": { "calculatorInput": false, "releaseArtifact": true, "websiteInput": false }, "datasetVersion": "1.0.0", "overlayVersion": "1.0.0", "records": [ { "accessedOn": "2026-07-14", "comparatorClaim": { "boundary": "Ippolito synthesis across several model benchmarks", "derivation": "Selected by Ippolito after comparing IEA 1.7 Wh, O'Donnell 1.2 Wh, and ML.ENERGY 0.42 Wh estimates.", "displayValue": "2", "unit": "Wh per 1024x1024 image" }, "comparatorId": "jon-ippolito-what-uses-more", "decisionSummary": "Replace the unreconstructable 0.48 Wh image coefficient with the IEA's pinned 1.7 Wh benchmark, while crediting Ippolito for the comparison that led to the source review.", "disposition": "adapt", "evidenceLocator": "Measurements sheet, Image generation row and its IEA, O'Donnell, and ML.ENERGY comparison notes", "evidenceSourceIds": [ "iea-energy-and-ai", "ippolito-what-uses-more", "ml-energy-leaderboard", "odonnell-ai-energy-footprint", "ifri-ai-dc-energy" ], "evidenceUrl": "https://docs.google.com/spreadsheets/d/1xgvGhd6rmnBAEk9t2lesKlJeUy4TRQHrW6uWDSEVDFY/edit?gid=0#gid=0", "id": "ippolito-image-generation-2026-07", "lastReviewed": "2026-07-14", "projectClaimAfter": { "boundary": "Controlled H100 experimental testing; GPU electricity only", "derivation": "Adopts the underlying IEA Figure 1.16 benchmark surfaced by the Ippolito comparison rather than Ippolito's rounded 2 Wh synthesis.", "displayValue": "1.7", "unit": "Wh per SD-XL 1.0-base image" }, "projectClaimBefore": { "boundary": "Legacy row with no pinned model, hardware, image settings, or exact derivation", "derivation": "Stored coefficient associated with a secondary Ifri report that states a rounded 0.5 Wh figure.", "displayValue": "0.48", "unit": "Wh per generated image" }, "rationale": [ "The IEA identifies the model, H100 hardware, controlled test context, and GPU-only measurement boundary.", "Ippolito's 2 Wh selection is useful synthesis but is not itself a measured value.", "The previous 0.48 Wh coefficient cannot be reconstructed from the included source chain." ], "targetField": "serverWhPerUnit", "targetId": "image-generation", "targetType": "activity", "topic": "Image-generation energy", "unresolvedQuestions": [ "The IEA report does not state the image dimensions in the Figure 1.16 note.", "A future production-serving benchmark could justify adding facility overhead or replacing the GPU-only coefficient." ] }, { "accessedOn": "2026-07-14", "comparatorClaim": { "boundary": "Universal conversion applied to task energy in What Uses More?", "derivation": "Midpoint between about 1 mL/Wh derived from Google and 13 mL/Wh derived from Mistral through a carbon-intensity conversion.", "displayValue": "7", "unit": "mL per Wh" }, "comparatorId": "jon-ippolito-what-uses-more", "decisionSummary": "Retain the existing split water method, add Ippolito's 7 mL/Wh result as an explicit comparator, and document that the stated Mistral bridge evaluates to 11.84 rather than 13 mL/Wh when 0.3 gCO2/Wh is used.", "disposition": "retain", "evidenceLocator": "Conversions sheet, water-conversion section comparing Google, Mistral, and the selected 7 mL/Wh midpoint", "evidenceSourceIds": [ "schneider-google-scale", "mistral-environmental-standard", "lbnl-data-center-report", "ippolito-what-uses-more", "scenario-methods" ], "evidenceUrl": "https://docs.google.com/spreadsheets/d/1xgvGhd6rmnBAEk9t2lesKlJeUy4TRQHrW6uWDSEVDFY/edit?gid=1688264219#gid=1688264219", "id": "ippolito-water-conversion-2026-07", "lastReviewed": "2026-07-14", "projectClaimAfter": { "boundary": "Rounded project bridge separating direct cooling water from broader electricity water", "derivation": "Retained while the method record now exposes Google, Mistral, LBNL, Ippolito, and the corrected 11.84 mL/Wh cross-metric arithmetic.", "displayValue": "1 direct / 6 broader", "unit": "mL per cloud-side Wh" }, "projectClaimBefore": { "boundary": "Rounded project bridge separating direct cooling water from broader electricity water", "derivation": "Internal teaching heuristic with prompt-specific exceptions.", "displayValue": "1 direct / 6 broader", "unit": "mL per cloud-side Wh" }, "rationale": [ "Ippolito's 7 mL/Wh is a midpoint heuristic rather than an activity-specific measurement.", "Mistral reports water and carbon per response, not water per Wh.", "Keeping direct and broader water separate preserves a distinction that the comparator collapses." ], "targetId": "water-method", "targetType": "method", "topic": "Generic electricity-to-water conversion", "unresolvedQuestions": [ "A boundary-matched direct-plus-indirect water coefficient should replace the generic 6 mL/Wh heuristic when available.", "Future activity-specific provider measurements may replace generic conversions row by row." ] }, { "accessedOn": "2026-09-05", "comparatorClaim": { "boundary": "What Uses More? teaching-calculator configuration", "derivation": "Ippolito's coding and paragraph examples separate the action being requested from the depth of AI work. They are distinct examples rather than evidence for a generic coding-request average.", "displayValue": "Configured code example; no universal request coefficient adopted", "unit": "configured action/depth comparison" }, "comparatorId": "jon-ippolito-what-uses-more", "decisionSummary": "Use Ippolito's action/depth framework to clarify the counted interaction, while adopting Hausfather's primary modeled logs for an explicitly intensive 150 Wh request example.", "disposition": "adapt", "evidenceLocator": "Measurements sheet, configured code example and separate paragraph example; Hausfather, Measuring my own footprint and footnote 1", "evidenceSourceIds": [ "hausfather-agentic-ai", "ippolito-what-uses-more", "scenario-methods" ], "evidenceUrl": "https://docs.google.com/spreadsheets/d/1xgvGhd6rmnBAEk9t2lesKlJeUy4TRQHrW6uWDSEVDFY/edit?gid=0#gid=0", "id": "ippolito-coding-request-energy-2026-09", "lastReviewed": "2026-09-05", "projectClaimAfter": { "boundary": "Intensive-use example; modeled data-center electricity triggered by one human message, with automatic agent work included", "derivation": "170,000 Wh / 1,138 typed human prompts = 149.38 Wh, rounded to 150 Wh; repeated model-call usage records were deduplicated by API message ID.", "displayValue": "150", "unit": "Wh per human coding-agent request" }, "projectClaimBefore": { "boundary": "Historical internally normalized hourly row with a 30 Wh laptop allowance", "derivation": "The retired row divided Couch's approximate 1,300 Wh-per-day estimate by an unsupported assumption of four active coding hours and added an internal 30 Wh laptop allowance.", "displayValue": "325 cloud/source; 355 total-system", "unit": "Wh per coding-agent hour" }, "rationale": [ "The old hourly coefficient depended on an undocumented duration normalization.", "Hausfather reports the human-prompt denominator and a deduplication rule, so the 149.38 Wh arithmetic is reproducible.", "Ippolito's configured examples help explain interaction depth but do not establish the physical coefficient." ], "targetField": "serverWhPerUnit", "targetId": "coding-agent-requests-intensive", "targetType": "activity", "topic": "Coding-agent request energy and functional unit", "unresolvedQuestions": [ "Exact Claude model versions and mix are not fully disclosed.", "One intensive user's modeled workload is not a generic request average.", "Future provider measurements may justify separate workload tiers." ] }, { "accessedOn": "2026-09-05", "comparatorClaim": { "boundary": "Universal water conversion in What Uses More?", "derivation": "Ippolito selects a midpoint heuristic across unlike water evidence.", "displayValue": "7", "unit": "mL per Wh" }, "comparatorId": "jon-ippolito-what-uses-more", "decisionSummary": "Apply the project's existing 1 mL/Wh direct and 6 mL/Wh broader scenario to the new energy value, while keeping Ippolito's 7 mL/Wh choice as comparison rather than attribution.", "disposition": "adapt", "evidenceLocator": "Conversions sheet water comparison; data/methods/water-method.json shared AI conversion", "evidenceSourceIds": [ "hausfather-agentic-ai", "ippolito-what-uses-more", "schneider-google-scale", "lbnl-data-center-report", "scenario-methods" ], "evidenceUrl": "https://docs.google.com/spreadsheets/d/1xgvGhd6rmnBAEk9t2lesKlJeUy4TRQHrW6uWDSEVDFY/edit?gid=1688264219#gid=1688264219", "id": "ippolito-coding-request-water-2026-09", "lastReviewed": "2026-09-05", "projectClaimAfter": { "boundary": "Project teaching scenario transferred to the 150 Wh intensive-use energy example; broader water includes direct water", "derivation": "150 Wh x 1 mL/Wh = 150 mL direct; 150 Wh x 6 mL/Wh = 900 mL broader.", "displayValue": "900", "unit": "mL broader water per human coding-agent request" }, "projectClaimBefore": { "boundary": "Historical project 1:6 conversion applied to the retired 325 Wh hourly row", "derivation": "325 Wh x 1 mL/Wh direct and x 6 mL/Wh broader.", "displayValue": "325 direct / 1,950 broader", "unit": "mL per coding-agent hour" }, "rationale": [ "Joel selected the existing shared AI water scenario for consistency with the reasoning row.", "The broader 900 mL value already includes the direct 150 mL and is not summed with it.", "Neither Hausfather nor Ippolito measures provider-specific coding-request water." ], "targetField": "totalWaterMlPerUnit", "targetId": "coding-agent-requests-intensive", "targetType": "activity", "topic": "Coding-agent request water scenario", "unresolvedQuestions": [ "Cooling technology, serving location, electricity supply, and time are unknown.", "The shared 6 mL/Wh broader scenario remains an internal assumption needing boundary-matched evidence." ] } ], "reviewedOn": "2026-09-05", "status": "review-overlay" }