Energy
About nine tenths of the energy of an hour of streaming is drawn in the home, by the screen and the router: 89% of the 188 Wh of the Carbon Trust's European average hour, and about 89% of the 113 Wh of Ofcom's UK one (Carbon Trust, 2021; Ofcom, 2022, as 3GPP sums it up). The network beyond the home draws nearly the same power whether video flows through it or not, so on a fixed line fewer bits save almost no energy: about 0.05 W for each Mbit/s (Carbon Trust; Mytton, Lundén and Malmodin, 2024).
What does move with a stream is the mobile radio, about 1.5 W for each Mbit/s (Carbon Trust), the decoder on your device, where a hardware one uses less than 9% of the energy of a software one (Kränzler, Herglotz and Kaup, 2024), and above all the screen. Fewer bits do cost less to send, since delivery is billed by the gigabyte (as Amazon CloudFront's prices are, for one): that, and mobile data, is their honest claim, not energy on a fixed line.
Below, each hop of one viewing hour, what it draws and how each figure is known, in the classes of WAI's energy measurement; then what all viewing on these servers drew; then what the player does.
How a figure is known
- CalibratedInstrument
- read from a calibrated instrument.
- OnChipCounter
- read from the energy counter the chip keeps about itself.
- ModelBased
- computed from a model calibrated on the very machine it describes.
- Estimator
- estimated from published measurements or models for a kind of device or network, not calibrated on yours.
- Unmetered
- no figure: none is shown, and none is counted as zero.
The network's figure is an estimate too: its power model was calibrated on operators' networks, not on the path to you, and WAI counts a coefficient not calibrated where it is used as an estimate, however it is multiplied. No figure here is ModelBased. Figures of two classes are never added together: a reading added to an estimate is neither. So the hour has a total per class, not one number.
One hour of one stream
This site's reference stream: His Girl Friday, 1080p AV1, 4.3 Mbit/s. Change what it is played on below.
Origin: this site's serverThis server's process, its share for the media it sent
0.00059 Wh (0.00026–2.6) in the hour
OnChipCounter read from the energy counter the chip keeps about itself, on this server.
Declared: ±0.023% (standard uncertainty, k = 1, of the part that can be evaluated) over 649 hours metered; counter filtering Unknown; the counter's own error is not stated by its maker and is not in this.
Read from the kernel's own count of this server process's CPU energy (macOS, per task, no root), every second, a minute counted only where the counter was live; 100% of the video sent was sent in such minutes. Split to the media by CPU time: low, the media requests' own work over the process's; centre, over all requests' own; high, with every CPU second outside a request's own work (connections, file reads, background jobs) charged to the media too. The CPU only, on the release build: not memory, disks, the network card, the machine's idle draw or anything beyond this host. Not an idle-bracketed figure, so WAI's bracket ratio does not apply.
Sources: WAI
NetworkA fixed line (Wi-Fi or cable)
0.37 Wh (0.32–4.2) in the hour
Estimator estimated from published measurements or models for a kind of device or network, not calibrated on yours.
The range the cited studies give: 0.32–4.2 Wh.
The router's added draw while streaming, and the access and core network's slope of about 0.05 W per Mbit/s, at 4.3 Mbit/s: the Carbon Trust's power model with Malmodin's coefficients, calibrated on operators' networks, not on the path to you, so an estimate. The router and the line draw 13 W and 2.8 W whether or not you stream: your share of the hour, not your change.
Sources: Jens Malmodin 2023; Greening of Streaming 2026; The Carbon Trust 2021; Mytton 2024
Decoding on your deviceA hardware decoder, for scale
0.72 Wh in the hour, part of another figure, not added
Estimator estimated from published measurements or models for a kind of device or network, not calibrated on yours.
One published measurement; no range is published.
0.72 W decoding HD at 30 frames a second on an evaluation platform: part of the screen's figure below, which was measured with the decoder inside it, so not added again. Hardware decoders use less than 9% of the energy of software ones.
Sources: Herglotz 2022; Kränzler 2024
Your screenA TV, size not known
100 Wh (28–182) in the hour
Estimator estimated from published measurements or models for a kind of device or network, not calibrated on yours.
The range the cited studies give: 28–182 Wh.
A TV of a size the browser cannot tell: the Carbon Trust's assumed 100 W, between the field averages of small and large sets. Its whole draw while playing, the decoder in hardware included, at the default picture: figures measured or assumed for that kind of screen, not for yours. A stream's picture size or bitrate moved a TV's draw by 1-4%, not significantly; its eco mode saved about 37% on average.
Sources: The Carbon Trust 2021; European Commission; Fraunhofer FOKUS and Deutsche Telekom 2025; Greening of Streaming 2026; Fraunhofer FOKUS 2024
The hour, per class
- OnChipCounter 0.00059 Wh (0.00026–2.6)
- Estimator 100 Wh (28–186)
Eco mode
One switch, in the player's settings and on Settings, as Arcom recommends of video services: it plays at most 720 lines and in SDR, and suggests lowering your screen's brightness, which only you can do. What each part saves in an hour of the reference stream, from the same figures as above:
A TV, size not known
- The picture at most 720 lines, on the screen
- Nothing measurable (a stream's picture size moved a TV's draw by 1-4%, not significantly)
- The same, in the network on a fixed line
- 0.14 Wh (0.14–0.16) less
- The same, on 4G
- 4.2 Wh (2.9–5.6) less
- SDR instead of HDR
- 42 Wh (8.1–142) less
- Lowering the screen's brightness, or its own eco mode (your doing)
- 37 Wh (16–67) less
A phone
- The picture at most 720 lines, on the screen
- 0.18 Wh (0–0.18) less
- The same, in the network on a fixed line
- 0.14 Wh (0.14–0.16) less
- The same, on 4G
- 4.2 Wh (2.9–5.6) less
- SDR instead of HDR
- No figure: HDR on a phone's or laptop's screen was not measured
- Lowering the screen's brightness, or its own eco mode (your doing)
- 0.43 Wh (0.19–0.58) less
A computer (a laptop's figure)
- The picture at most 720 lines, on the screen
- No figure: not measured
- The same, in the network on a fixed line
- 0.14 Wh (0.14–0.16) less
- The same, on 4G
- 4.2 Wh (2.9–5.6) less
- SDR instead of HDR
- No figure: HDR on a phone's or laptop's screen was not measured
- Lowering the screen's brightness, or its own eco mode (your doing)
- No figure: not measured
Estimator figures throughout. The screen's brightness is the lever that moves the hour; the picture's size barely moves a TV (Greening of Streaming, 2026; Fraunhofer FOKUS, 2024).
All viewing on these servers
2026-11-01 to 2026-11-14 (UTC days). Both servers: this one (server 1) and server 3's counts as of 2026-11-14 17:05 UTC. Their counts are added before anything is hidden.
- This site's server, its share for the media it sent
- OnChipCounter 0.00059 Wh (0.00026–2.6) per viewing hour; ±0.023% (standard uncertainty, k = 1, of the part that can be evaluated) over 649 hours metered; 100% of the video was sent while metered
- The network, if every hour went over a fixed line
- Estimator 0.22 Wh (0.17–4.1) per viewing hour added, at the 1.4 Mbit/s sent for each hour of video
- The network, if every hour went over 4G
- Estimator 2.1 Wh (1.4–2.8) per viewing hour added, at the 1.4 Mbit/s sent for each hour of video
- Screens, by the kinds the players counted
- Estimator 14 Wh (8.9–23) per viewing hour
- Software decoding added, where no power-efficient decoder was found
- Estimator 0.85 Wh (0.53–0.86) per viewing hour
- Streams begun, by screen
- Count TVs: 5.9%; Phones: 43%; Computers: 51%
- Streams begun, by decoder
- Count A power-efficient decoder, nothing set aside: 24%; A power-efficient decoder, less efficient renditions set aside: 3.1%; No power-efficient decoder for any rendition: 21%; The browser could not say: 16%; The browser's own player chose: 36%
- Stopped because the page was hidden or the screen locked
- Count 8.5% of the streams begun (at most one a day for a page and a source)
- Stopped because the player was scrolled out of view
- Count 29% of the streams begun (at most one a day for a page and a source)
- Played in eco mode
- Count 4.6% of the streams begun
- HDR pictures shown, and withheld
- Count 27 shown; 60 withheld (the display could not show HDR, its decoder was not power-efficient, or eco mode was on)
1,744 streams begun were counted by the players, at most one note a day for each page and source. Figures need 50 plays behind them; only ratios and shares are shown, so nothing hidden can be had by subtraction. The screens and decoders are weighted by streams begun, not by hours watched, as a play's length is not counted.
The servers' energy, fabric by fabric
Per viewing hour, on each kind of processor in the servers. A fabric no counter reads is said to be unmetered: no figure is shown for it, and none is counted as zero.
- CPU: this site's server, its share for the media it sent
- OnChipCounter 0.00059 Wh (0.00026–2.6) per viewing hour; ±0.023% (standard uncertainty, k = 1, of the part that can be evaluated) over 649 hours metered; 100% of the video was sent while metered
- CPU: the library channel's encoder
- Unmetered Not metered. One server encodes INV Live: Library with x264 (libx264) on its CPU around the clock, watched or not, in a process of its own that the counter does not read.
- GPU
- Unmetered Not metered. The servers do none of the work of sending video on it, and no counter they can read covers it.
- Media engine
- Unmetered Not metered. The servers send each title's files as they were encoded ahead of time, decoding and encoding nothing to send them, and no counter they can read covers it.
- Neural engine
- Unmetered Not metered. The servers do none of the work of sending video on it, and no counter they can read covers it.
What a visit costs before you press anything
0 ad-minutes inserted in an hour. The site inserts no advertising, in its player or on its pages. Some live channels' streams carry their broadcasters' own ads, and films that play in YouTube's player show YouTube's ads; the site does not insert, choose or count them.
One page of each of 56 kinds of page on https://invisiblenetwork.ai, each opened in a fresh browser and left for 30 seconds with nothing pressed, measured on 2026-11-05 at 05:25 UTC. They drew from 57 KB to 1,594 KB each; 0% of all their bytes came from other hosts, no other domain.
On none of them did anything play or sound by itself, and none asked for sound or video. Nothing on them moved for more than five seconds by itself, apart from the site's own signs that pulse or spin while they say something (the "Live" badges and dots), which stop when your system asks for reduced motion.
Each kind of page, its first 30 seconds with nothing pressed
| Page | Bytes | Requests | Script | From other hosts | Played or moved by itself |
|---|
| Home/ | 1,024 KB | 52 | 34 KB | None | Nothing |
|---|
| Moments/moments | 778 KB | 20 | 50 KB | None | Nothing |
|---|
| The guide/tv | 282 KB | 25 | 41 KB | None | Nothing; pulsing: the "Live" badges and dots |
|---|
| Tonight on the channels/tv/tonight | 69 KB | 16 | 35 KB | None | Nothing; pulsing: the "Live" badges and dots |
|---|
| A channel/watch/live-library | 60 KB | 14 | 34 KB | None | Nothing; pulsing: the "Live" badges and dots |
|---|
| A channel's schedule/channel/ch_0001 | 57 KB | 14 | 34 KB | None | Nothing |
|---|
| Live TV around the world/tv/world | 256 KB | 31 | 34 KB | None | Nothing; pulsing: the "Live" badges and dots |
|---|
| A world channel/live/10TV.in | 306 KB | 25 | 55 KB | None | Nothing; pulsing: the "Live" badges and dots |
|---|
| The wall/wall | 104 KB | 15 | 59 KB | None | Nothing |
|---|
| Stations/stations | 60 KB | 14 | 34 KB | None | Nothing |
|---|
| A station/stations/stn_film_noir | 330 KB | 44 | 34 KB | None | Nothing |
|---|
| The library/library | 1,063 KB | 48 | 34 KB | None | Nothing |
|---|
| A film/title/title_000002 | 305 KB | 22 | 37 KB | None | Nothing |
|---|
| A film from the archives/watch/archive/ia/AboutBan1935 | 177 KB | 28 | 37 KB | None | Nothing |
|---|
| Rights holders' films/library/rights-holders | 580 KB | 42 | 36 KB | None | Nothing |
|---|
| A rights holder's film/watch/rights-holders/ge07CzDMqfw | 340 KB | 16 | 36 KB | None | Nothing |
|---|
| News/news | 556 KB | 37 | 34 KB | None | Nothing; pulsing: the "Live" badges and dots |
|---|
| Comedy/comedy | 741 KB | 49 | 36 KB | None | Nothing |
|---|
| Anime/anime | 657 KB | 43 | 36 KB | None | Nothing |
|---|
| Sports/sports | 57 KB | 15 | 34 KB | None | Nothing |
|---|
| A sport/sports/archery | 247 KB | 32 | 36 KB | None | Nothing |
|---|
| Music/music | 82 KB | 15 | 41 KB | None | Nothing |
|---|
| Radio/radio | 232 KB | 36 | 38 KB | None | Nothing |
|---|
| Podcasts/podcasts | 836 KB | 91 | 34 KB | None | Nothing |
|---|
| A podcast/podcast/1200361736 | 98 KB | 16 | 47 KB | None | Nothing |
|---|
| A podcast's episode/podcast/1200361736/0a23f8a3ae27 | 125 KB | 17 | 47 KB | None | Nothing |
|---|
| Free music/music/free | 821 KB | 81 | 34 KB | None | Nothing |
|---|
| An album/music/free/album/ia-NS050 | 224 KB | 28 | 47 KB | None | Nothing |
|---|
| Radio drama/radio-drama | 1,168 KB | 75 | 34 KB | None | Nothing |
|---|
| A radio series/radio-drama/series/yours-truly-johnny-dollar | 96 KB | 17 | 47 KB | None | Nothing |
|---|
| A radio episode/radio-drama/series/yours-truly-johnny-dollar/e16b889c9480 | 83 KB | 18 | 47 KB | None | Nothing |
|---|
| Books/books | 1,111 KB | 57 | 36 KB | None | Nothing |
|---|
| A book/books/gutenberg/11 | 916 KB | 49 | 37 KB | None | Nothing |
|---|
| A book, to read/books/gutenberg/11/read | 227 KB | 19 | 44 KB | None | Nothing |
|---|
| An audiobook/books/librivox/119 | 188 KB | 21 | 43 KB | None | Nothing |
|---|
| Print/print | 1,328 KB | 62 | 33 KB | None | Nothing |
|---|
| A comic, magazine or paper/print/ia/americas-greatest-comics-1-fawcett-publications-1941 | 672 KB | 36 | 33 KB | None | Nothing |
|---|
| Art/art | 1,594 KB | 45 | 32 KB | None | Nothing |
|---|
| A work of art/art/work/rijks/SK-A-4054 | 1,273 KB | 38 | 41 KB | None | Nothing |
|---|
| Gallery mode/art/gallery | 341 KB | 19 | 41 KB | None | Nothing |
|---|
| An object in 3D/art/objects/nasa/1999-rq36-asteroid | 95 KB | 17 | 41 KB | None | Nothing |
|---|
| Games/games | 62 KB | 14 | 34 KB | None | Nothing |
|---|
| A game/games/2048 | 58 KB | 14 | 33 KB | None | Nothing |
|---|
| Karaoke/karaoke | 257 KB | 17 | 39 KB | None | Nothing |
|---|
| A song/karaoke/song/openscore-6583477 | 65 KB | 15 | 37 KB | None | Nothing |
|---|
| Learn/learn | 360 KB | 33 | 34 KB | None | Nothing |
|---|
| A course/learn/ia/Philosophy_185_Fall_2007_UC_Berkeley | 169 KB | 26 | 34 KB | None | Nothing |
|---|
| A lecture/learn/ia/Philosophy_185_Fall_2007_UC_Berkeley/2f2bb8c86e43 | 65 KB | 17 | 37 KB | None | Nothing |
|---|
| Search/search?q=night | 64 KB | 14 | 34 KB | None | Nothing |
|---|
| Ask/ask | 58 KB | 15 | 33 KB | None | Nothing |
|---|
| The studio/studio | 91 KB | 17 | 59 KB | None | Nothing |
|---|
| Yours/mine | 82 KB | 19 | 56 KB | None | Nothing |
|---|
| Settings/settings | 96 KB | 21 | 58 KB | None | Nothing |
|---|
| Energy/energy | 69 KB | 14 | 34 KB | None | Nothing |
|---|
| About/about | 77 KB | 14 | 34 KB | None | Nothing |
|---|
| A page not found/no-such-page | 106 KB | 14 | 42 KB | None | Nothing |
|---|
Bytes are what arrived on the wire, headers and compressed bodies, read from the browser's network layer: the page, its scripts, styles, pictures and the data it asked for, and its service worker's. A browser that has visited before draws less. The browser: Chromium 156.0.8078.4 (Playwright 1.64.0), headless, 1280×900, autoplay allowed without a press.
To measure it again, from the site's source: cd e2e && node scripts/meters.mjs https://invisiblenetwork.ai. It only asks for pages, sends nothing that would be counted or kept, and writes the same table as JSON. In any browser, the same can be read in its network panel: open a page in a new private window, press nothing, and wait 30 seconds.
What the player does
- It stops what nobody can see. When the page is hidden, the screen locks, or an in-page player is scrolled out of view, the player pauses and stops fetching; it carries on where it was when you come back. A picture-in-picture window, a cast, a headset or a watch-together room keeps playing.
- It plays nothing you did not start. No previews on hover, and no channel or film starts by itself when a page opens; a card or a Play button starts it. In Moments, one press plays a set of up to 20 moments in a row, each giving way to the next at its end, and then the set stops and asks; you can choose instead to have each moment wait for your press (in Moments' Your feed, or in Settings). On Comedy, one press plays its lineup of records and radio comedy in the music bar, one after another, and a separate press plays up to ten sets in YouTube's player, one after another; each stops after its last. On Anime, one press plays a series' first episodes offered, or a series' episodes from the one pressed, up to ten, in order, in YouTube's player, and stops after the last; on the anime channel, one press plays the episode on the air (or, where that one does not play here, the next that does) from its beginning and the next ones on its schedule that play here, up to ten in all, in YouTube's player, and stops after the last. On Karaoke, one press of Sing plays a song, or a lineup of up to 20 songs one after another, on your device, and stops after the last. On the site's channels, one press plays the title on air from where the clock is, and each next title at its start, until you stop it or leave the page; after four hours with no press, it stops at the next title and asks. On your channel, one press plays the title picked first and each next one at its start, in the same way, until you stop it or leave the page; after four hours with no press, it stops at the next title and asks. On a marathon (a radio series played on demand, not on a schedule), one press of Start plays up to 20 episodes one after another in the music bar, in broadcast order, from where you left off, and stops after the last. On the wall, one press starts every tile on it, with sound from one tile only, and nothing on it plays before that press; where the wall has a film from a rights holder's YouTube channel, that press also loads YouTube's player script and player, and YouTube's player chooses that tile's picture size itself, so the wall's rule of the smallest watchable picture does not reach it. Turning autoplay off cut daily viewing by 21 minutes on average in a ten- to seventeen-day trial with 76 people (Schaffner and others, 2024).
- It never sends more picture than the screen shows. The rendition is capped at the player's size times the screen's pixel density, and HDR is sent only where the display can show it and a power-efficient decoder takes it.
- The decoder that draws least, strictly. The player asks the browser (Media Capabilities) which renditions it decodes power-efficiently, in practice in hardware, and plays only those wherever there is one; it says which it chose, as a count above.
- Eco mode, one switch. At most 720 lines and SDR, with what that saves said plainly above.
- The same picture in fewer bits, for cost. Films are encoded in AV1, HEVC and H.264. Tears of Steel at 1080p takes 1.9 Mbit/s in AV1 and scores 93 out of 100 against its master (VMAF), where the H.264 copy the site carried before took 4.5 Mbit/s. On a fixed line that saves almost no energy; it saves data on a mobile plan, the radio's draw, and what delivery costs.
- Only what you watch. The player reads 30 seconds ahead, never more than 60, so video you never reach is never sent.
Counted on this server since it started
18.0 hours ago. The site is served by two servers, so each sees part of the traffic.
Requests served
1,073,367
Video segments players asked for
116
From the players' own reports (CMCD).
Average bitrate players chose
0.8 Mbit/s
Capped at the player's size on the screen.
Bitrate on screen
0.3 Mbit/s
What was shown, not only what was fetched (CMCD version 2).
From Play to the picture
25.7 s
The mean, each viewing once (CMCD version 2).
Stalls
359.6 an hour, 86.96% of the time
After the start, per hour of video fetched (CMCD version 2).
Dropped frames
13.5 a minute
Per minute of video fetched (CMCD version 2).
Watching now
0
Playback sessions in the last ten minutes.
Live figures are at /metrics; what the servers sent, day by day, and the levers of a whole hour, at Telemetry.
Sources
- The Carbon Trust (DIMPACT), Carbon impact of video streaming, June 2021.
- Carnstone for Ofcom, Carbon emissions of streaming and digital terrestrial television, Oct 2022, revised Jul 2024.
- 3GPP TR 26.942, section 4.2.7, summarising Ofcom's study, Energy estimates for broadcast and streaming in the UK (viewing devices 58.0, home equipment 21.5, peripherals 21.5, network 12.0 Wh), read 22 Oct 2026.
- Mytton, Lundén and Malmodin, Journal of Industrial Ecology 28(4), Network energy use not directly proportional to data volume: the power model approach, Aug 2024.
- Kränzler, Herglotz and Kaup, PCS 2024, Decoding energy of AVC, HEVC, VP9, AV1 and VVC, in software and hardware, 14 Feb 2024.
- Amazon Web Services, Amazon CloudFront pay-as-you-go pricing (data transfer out to the internet, per GB), read 22 Oct 2026.
- WAI, Transaction Science, Inc., WAI extension: energy measurement (acquisition classes, declared uncertainty, attribution), read 22 Oct 2026.
- Jens Malmodin (Ericsson), Just measure IT! (one Swedish home, measured continuously from November 2022), 2023.
- Greening of Streaming, Hackathon data analysis (TV and router power at the plug), Feb 2026.
- Herglotz, Kränzler, Schober and Kaup, IEEE Circuits and Systems Magazine (preprint), Sweet streams are made of this: the system engineer's view on energy efficiency in video communications (decoding HD at 30 fps: 0.719 W in hardware), 30 Sep 2022.
- European Commission, EPREL product database, Energy label records of electronic displays (SDR and HDR on-mode power), read 21 Oct 2026.
- Fraunhofer FOKUS and Deutsche Telekom, Putting sustainable streaming to the test (field test in homes, per-second smart plugs), 9 Sep 2025.
- Fraunhofer FOKUS, Green streaming white paper: intensity and power models side by side, Oct 2024.
- Accedo, Report on the energy use of video streaming (smart TV picture settings), 31 Aug 2023.
- Soltani Nezhad et al., TNO and Radboud University, Quantifying the energy of video on a smartphone (poster), 30 Jun 2025.
- Northwest Energy Efficiency Alliance (report E16-298), Revising the TV energy use test procedure: incorporating HDR, 16 Nov 2016.
- Robert Seeliger, Fraunhofer FOKUS, HbbTV Symposium, Green streaming decoded: myths and reality (a 55-inch OLED TV, 540p at 5 Mbit/s against 2160p60 at 25 Mbit/s), 14 Nov 2024.
- Schaffner, Ulloa, Sahni, Li, Cohen, Messier, Gao and Chetty (University of Chicago), PACM HCI (CSCW), An experimental study of Netflix use and the effects of autoplay on watching behaviors (76 people, 10-17 days), 20 Dec 2024.
- Arcom, Recommendation under article 26 of the law to reduce the environmental footprint of digital technology (a one- or two-click energy-sobriety setting: picture quality, autoplay off), 26 Jul 2023.
- Arcom, Arcep and ADEME, Study of the environmental impact of audiovisual uses in France (lower picture quality among the sobriety measures), Oct 2024.