Text
0.30 Wh per 1,000 text tokens
If you do not know your weekly token total, the calculator estimates it from habits: 1,200 tokens per quick chat and 12,000 tokens per long work session.
Calculator methodology
Transparent assumptions, not fake precision.
The calculator is meant to give people a defensible order-of-magnitude estimate of personal AI electricity use. It is not a provider-grade meter. The assumptions below are public, revisable, and intentionally easy to inspect.
Current formula
Images
3.0 Wh per generated image
This is a rounded central placeholder based on published image-generation benchmarks and retained because image models still vary dramatically.
Video
12 Wh per second of generated video
This is the least certain assumption in the calculator: a provisional public heuristic, not a settled universal constant.
How we derived it
Central estimates were chosen to be explainable, not flattering.
1
Text is anchored to public prompt-level disclosures.
UNESCO and UCL cited about 0.34 Wh per prompt as a broad public benchmark in 2025. Google later published a more detailed estimate of 0.24 Wh for a median Gemini Apps text prompt. We use 0.30 Wh per 1,000 tokens as a simple central heuristic between those public signals, while acknowledging that real prompts differ in length, model size, and hardware path.
2
Image generation still has huge model spread.
Published image-generation research in 2025 found very large variation between models, including up to a 46× difference across tested systems. That uncertainty argues against pretending we know a single exact per-image number. We keep 3 Wh per image as a rounded midpoint placeholder rather than a precision claim.
3
Video is currently the shakiest part of public estimation.
Public reporting and open-model measurements suggest video can swing wildly depending on model family, resolution, frame count, inference steps, and infrastructure. That is why the calculator explicitly labels video as the weakest assumption, and why video is the first thing we re-check when the methodology is reviewed.
4
The range is there because the field is not stable yet.
The site shows a plausible range around the central estimate because public provider disclosures are incomplete, workloads differ, and usage patterns can change electricity demand substantially. The calculator is a transparency tool, not a false meter.
Current source logic
What this version is leaning on
No calculator factor change. The public brackets are still the same.
Re-read the calculator against current public evidence. Google’s median Gemini Apps text prompt is still ~0.24 Wh (May 2025 measurement; no newer official median found). That still sits under the site’s 0.30 Wh per 1,000 tokens midpoint, with the older UNESCO/UCL ~0.34 Wh-per-prompt benchmark as the upper public bracket. Image generation still shows a large model spread, including the 2025 Hidden Cost paper’s up to 46× difference; 3 Wh per image remains a rounded midpoint, not a meter. Video is still the weakest public assumption: no new universal per-second factor replaced 12 Wh/s. Factors stay 0.30 / 3.0 / 12. Review cadence is monthly, not daily.
Core sources: UNESCO / UCL report summary; Google Cloud methodology post; Google-scale Gemini inference measurement (arXiv); The Hidden Cost of an Image; Hugging Face CogVideoX documentation.
The public factors remain order-of-magnitude estimates.
Re-checked the public evidence base for text, image, and video inference energy. Google’s median Gemini Apps text-prompt figure (~0.24 Wh) and the broader UNESCO/UCL prompt-level benchmark still bracket the site’s 0.30 Wh per 1,000 tokens midpoint. Image generation remains highly model-dependent (large published spreads across systems), so 3 Wh per image stays a rounded placeholder. Video remains the weakest public assumption: measurements are still sparse, model-specific, and sensitive to duration/resolution/steps, so 12 Wh per second stays provisional. No calculator factor changed in this review.
Core sources: UNESCO / UCL report summary; Google Cloud methodology post; Google-scale Gemini inference measurement (arXiv); The Hidden Cost of an Image; Hugging Face CogVideoX documentation.
Public prompt-level disclosures still support a rough midpoint.
UNESCO and UCL cited about 0.34 Wh per prompt as a broad benchmark, while Google later disclosed 0.24 Wh for a median Gemini Apps text prompt. The site’s 0.30 Wh per 1,000 tokens remains a simple central heuristic, not a provider-grade meter.
Image generation still needs visible uncertainty.
Published image-generation research found very large efficiency differences across tested systems. A rounded 3 Wh per image stays defensible only as a midpoint placeholder with explicit uncertainty.
Video remains provisional.
Public video-generation measurements are still too sparse and model-specific for a clean universal factor. The 12 Wh-per-second assumption remains the calculator’s weakest estimate.
Review policy
The calculator assumptions are reviewed about monthly, and whenever public evidence changes enough to move a factor. When the underlying evidence changes meaningfully, the site changes too. That is a feature, not a flaw.
Last reviewed: September 9, 2026