ABCDEFGHIJKLMNOPQRSTUVWXYZAAAB
1
OutputEnergy (watt-hours)ScopeScale exponentRationale
2
📝 Paragraph of AI text3AI1Elsworth 2025 gives .24 Wh for Gemini "median text generation," but if this includes code this could be underestimating by an order of magnitude, since writing code requires far less energy than natural language (per ML.ENERGY 2025) and code completion is more incremental than writing paragraphs, which would have an outsized influence on the number of user prompts that go into finding the median. On the trustworthy side, Google does claim to have reduced energy by 33x since 2024, and hopefully they've considered code in that calculation as well.

Earlier studies have concluded:

IEA 2025 gives .3 Wh for "text generation" with a small model, 5 Wh for a medium-sized model.

ML.ENERGY 2025 gives .075 Wh = (900 J * 1 Wh/3600 J * 150 tokens per paragraph/500) for Mistral Large, but that's open weights.

Epoch AI (via O'Donnell 2025) gives .3 Wh per query.
3
🖥️ 100 lines of AI code0.5AI1ML.ENERGY 2025 gives .06 Wh = (200 J * 1 Wh/3600 J) per 10 lines of code (80 output tokens) for CodeLlama; that's open weights, but it seems consistent with other studies.
4
🖼 1024x1024 AI image2AI1IEA 2025 gives 1.7 Wh.

O'Donnell 2025 gives 1.2 = (4400 J * 1 Wh/3600 J) for Stable Diffusion 3 with 50 denoising steps.

ML.ENERGY 2025 gives .42 Wh = (1500 J * 1 Wh/3600 J) with Stable Diffusion 3 at 15 denoising steps.
5
🎬 3 seconds of AI video from text130AI2Delavande 2025 gives 128 Wh = 415 Wh * (3^2/5.4^2) using WAN2.1-T2V-14B at 720×1280 for 81 frames at 15 fps. (I'm converting from 5.4 to 3 seconds, using the finding that the temporal dimension scales quadratically due to attention layers).

O'Donnell 2025 gives 570 Wh = (3400000 J * 1 Wh/3600 J * 3/5) using CogVideoX at "high quality" (converting from 5 to 3 seconds).

ML.ENERGY 2025 gives 1.1 Wh = (4000 J * 1 Wh/3600 J) for 16 frames at 512px with Animatediff, so 3 seconds would be at least triple that and a better resolution/model could easily be another 10 times that.
6
🎬 3 seconds of AI video from image170AI2ML.ENERGY 2025 gives 4.7 Wh = (17000 J * 1 Wh/3600 J) Image to video (Stable Video Diffusion XT, 25 frames at 1024x576). This is 4.7/1.1 = 4.3x the energy; accounting for doubled resolution brings this to 2.1x, and accounting for the difference in frames brings it to 1.3x the energy. So I'm multiplying that by the Delavande estimate for text-to-video.
7
🔍 Searching Google without AI0.3data1Google 2009
8
🪫 Charging a smartphone20electric1EPA 2024
9
☁️ Storing 5 GB in the cloud for 1 month8data1Greenly 2024 gives 8 Wh = (100,000 Wh/1000 GB-year)*(1 year/12 months). Note that Greenly rejects much higher values from Stanford 2017 due to improved efficiency. "The electricity required to save a gigabyte (GB) of data to the cloud is around 0.0078 kWh per month, totaling to almost 0.1 kWh per year. [This average was derived by Greenly via evaluating various server LCAs, such as the DELL PowerEdge R730, various declared Power Usage Efficiencies by cloud services providers such as AWS, and average storage use rates suggested by the Natural Resources Defense Council.] "
10
🍿 Watching 1-hour Netflix movie200data1My calculations at Ippolito 2025c result in 150-200 Wh.

Greenly 2024 gives 66 Wh/GB, but I'm increasing this on the assumption it's viewed on a big TV.

Rejected Kamiya's "100x" claim in favor of the clearer graph at a new Kamiya source (Carbonbrief.org) corroborated by a later European study (Carbon Trust 2021). Revised results suggest about 200 Wh.
11
🎦 Zooming for 1 hour with 10 people2000data1My calculations at Ippolito 2025c result in 1500-2000 Wh.

Mytton 2020 gives up to 300 Wh for data transfer only
12
🐈 Scrolling TikTok for 30 minutes over 5G120data1Ippolito 2025d. This analysis is preliminary; I need to corroborate all sources and calculations as well as determine if water usage is comparable to AI rate.
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100