Energy and data in media

One hour of watching or listening

Modelled, not measured: a web page cannot read your device's power, so it is modelled from the device and its settings. Every value is a range from published measurements; the bitrates are this site's own.

What

Device

Screen

HDR

Rendition

Codec, size and average bitrate, as this title's master playlist states them (sound included).

Decoding

Network

Segments come

This site today: Cloudflare passes video segments through to our servers and keeps none (measured, see below).

Encode shared over

Grid

Title Any country's grid Any hour, UK time

The hour matters only for Great Britain here: other grids are yearly averages.

Playing: His Girl Friday: 1080p AV1, 4.3 Mbit/s. The renditions are read from this server's copy of the title.

Your change

120 Wh range 82–190

45 g CO₂ range 31–72

What this hour adds to what is drawn anyway: the power model. Use it to see what moving a lever does.

140 Wh range 91–220

52 g CO₂ range 35–84

This hour's share of everything that served it, the always-on router and network included: an average, as a yearly footprint counts it.

Grid of United States: 380 g CO2e per kWh, 2025, yearly average (life-cycle emissions of generation). Data read 2026-10-21. Ember: Yearly electricity data (CC BY 4.0).

The hour, stage by stage (modelled, in watt-hours)
Stage	Your change	Your share
Device and screen	120 Wh82–180	120 Wh82–180
measured Fraunhofer FOKUS and Deutsche Telekom (9 Sep 2025); Dom Robinson, Greening of Streaming (May 2023); Jens Malmodin (Ericsson) (2023); Greening of Streaming (Feb 2026); Fraunhofer FOKUS (Oct 2024)
Home network	0.15 Wh0.10–4.0	13 Wh8.0–18
measured Jens Malmodin (Ericsson) (2023); Greening of Streaming (Feb 2026); Prysmian (n.d.)The router is on whether or not you stream: your share of the hour is its whole draw, your change only what streaming adds.
Access and core network	0.22 Wh0.22–0.25	3.0 Wh0.97–7.3
power model, measured The Carbon Trust (DIMPACT) (June 2021); Mytton, Lundén and Malmodin, Journal of Industrial Ecology 28(4) (Aug 2024); Jens Malmodin (Ericsson) (2023); Prysmian (n.d.); Lundén et al., Sustainability 14(5) (2022)Wi-Fi and a cable differ only at hundreds of Mbit/s (2.5-5.5 W more over Wi-Fi above 400 Mbit/s); at video bitrates they are the same.
CDN and data centre	0.028 Wh0–1.3	1.3 Wh0.028–1.3
derived, measured Netflix (28 Apr 2023); The Carbon Trust (DIMPACT) (June 2021)Most of a CDN's power is drawn at rest, so a cache hit saves origin work and backbone transit but little energy now; over years, capacity is built for the evening peak, which is where caching pays.
Encoding, shared over viewings	0 Wh0	0.80 Wh0.29–10
derived Arunruangsirilert and Katto (24 Nov 2025); Silveira, Porto and Bampi, EUSIPCO 2017 (2017)26 renditions at 23.976 frames a second, encoded once and shared over 1,000 viewings. Already spent, so watching adds nothing to it.

What moves this hour most, from where it is set

Each lever moved across its options with everything else held, in grams of CO2, relative to now. Context is what a viewer rarely changes for one hour.

Country's grid (yearly averages) context −94% to +240% Central African Republic … Turkmenistan
Device context −98% to 0% Phone … TV, over 50 inches
Screen brightness or picture mode −37% to +48% Eco or dim … Vivid or full
HDR 0% to +42% HDR off … HDR on
Network 0% to +5.3% Fixed line, cable to the router … Mobile, 5G
Resolution and bitrate −0.16% to +0.41% 360p AV1, 0.46 Mbit/s … 2160p AV1, 14 Mbit/s
Codec, same picture size −0.019% to 0% 1080p HEVC, 3.9 Mbit/s … 1080p AV1, 4.3 Mbit/s
Edge cache or our servers 0% to 0% From the edge cache … From our servers

A bytes-based model (Sustainable Web Design v4, about 0.30 kWh per GB) would put 580 Wh on this hour's 1.9 GB: 4.9 times the power model's 120 Wh. Network power is mostly drawn at rest, so bytes are not joules at the margin.

This stream sends 4.3 Mbit a second, about 430,000 times the 10 or so bits a second a person takes in. The picture's fidelity sets the bits; attention does not.

A link to this hour

Why bytes are not joules

Attention is fixed. A person takes in about 10 bits a second, whatever arrives (Zheng and Meister, 2024). A video stream sends hundreds of thousands of times that; what sets its bitrate is the picture's fidelity, not what anyone can attend to.
Energy per bit keeps falling. The average electricity per gigabyte carried by the internet halved about every two years from 2000 to 2015 (Aslan, Mayers, Koomey and France, 2018). That average falls because traffic grows over equipment whose power changes little.
Network power is mostly fixed. A home's modem and router drew 12.7 W whatever the use, and streaming added 0.1 to 0.2 W (Malmodin, 2023); wired networks' power is almost independent of traffic, and capacity follows the evening peak (Schien and others, 2025). So a model that multiplies bytes by kilowatt-hours per gigabyte answers a yearly-footprint question, not "what if I change this" (Mytton, Lundén and Malmodin, 2024).
Byte-based calculators overstate video. Sustainable Web Design v4, CO2.js's default, puts about 0.30 kWh on every gigabyte: for an hour of video at 2 to 7 GB, 0.6 to 2.1 kWh, three to ten times the end-to-end totals the studies above find. It is shown beside each hour above for comparison, never used.
The screen is where the energy is. On a fixed line with a TV, the screen and the always-on home equipment take most of an hour's footprint; data centres and CDNs take 1% in the Carbon Trust's European figures (2021) and 5% in the IEA's global ones (2020). On a phone over 4G the radio network takes most of it instead.

Measured here: what our servers sent

Counted by the servers as they answer, per day (UTC). Both servers: this one (server 1) and server 3's counts as of 2026-10-27 18:36 UTC. Their counts are added before anything is hidden.

2026-10-21 to 2026-10-27 (UTC days). A figure with fewer than 50 plays behind it is not shown, nor a total that would reveal one by subtraction ("withheld").

Each day: plays, what was sent, and how
Day	Title plays	Sent	Hours of media	Average bitrate	Through Cloudflare	Text compressed ahead	Pages' LCP	Players' reports
2026-10-27	251	7.5 GB	9.6	1.7 Mbit/s	100%	73% saved	96% within 1.5 s (99% within 2.5 s)	3.4 per 1,000 segments; 2.6 Mbit/s asked for, 2.7 Mbit/s on screen; started in 0.57 s; stalled 0.012% of the time; 0.59 dropped frames a minute
2026-10-26	1,471	16 GB	25	1.5 Mbit/s	100%	73% saved	99% within 1.5 s (99% within 2.5 s); swapped in place, 96% within 1.5 s	2.0 per 1,000 segments; 2.8 Mbit/s asked for, 2.8 Mbit/s on screen; started in 0.63 s; stalled 0.015% of the time; 0.95 dropped frames a minute
2026-10-25	124	0.43 GB	1.5	0.62 Mbit/s	100%	75% saved	96% within 1.5 s (99% within 2.5 s)	2.4 per 1,000 segments; 2.0 Mbit/s asked for
2026-10-24	fewer than 50	fewer than 50	fewer than 50	fewer than 50	fewer than 50	77% saved	82% within 1.5 s (91% within 2.5 s)	fewer than 50
2026-10-23	fewer than 50	fewer than 50	fewer than 50	fewer than 50	fewer than 50	77% saved	53% within 1.5 s (76% within 2.5 s)	fewer than 50
2026-10-22	821	3.2 GB	12	0.59 Mbit/s	100%	74% saved	98% within 1.5 s (99% within 2.5 s)	2.2 per 1,000 segments; 2.2 Mbit/s asked for
2026-10-21	157	0.62 GB	2.6	0.53 Mbit/s	100%	75% saved	99% within 1.5 s (100% within 2.5 s)	2.0 per 1,000 segments; 1.9 Mbit/s asked for

By media family and picture height, over the same days
Family	Plays	Sent	Hours	Average bitrate	Per play
AV1	1,335	22 GB	21	2.3 Mbit/s	16 MB
2160 pixels high	fewer than 50	fewer than 50		fewer than 50
1920 pixels high	fewer than 50	fewer than 50		fewer than 50
1440 pixels high	fewer than 50	fewer than 50		fewer than 50
1080 pixels high	fewer than 50	fewer than 50		fewer than 50
1078 pixels high	fewer than 50	fewer than 50		fewer than 50
1012 pixels high	fewer than 50	fewer than 50		fewer than 50
960 pixels high	fewer than 50	fewer than 50		fewer than 50
856 pixels high	fewer than 50	fewer than 50		fewer than 50
720 pixels high	212	1.9 GB		1.5 Mbit/s
692 pixels high	fewer than 50	fewer than 50		fewer than 50
676 pixels high	fewer than 50	fewer than 50		fewer than 50
640 pixels high	fewer than 50	fewer than 50		fewer than 50
572 pixels high	fewer than 50	fewer than 50		fewer than 50
546 pixels high	fewer than 50	fewer than 50		fewer than 50
540 pixels high	172	0.46 GB		0.54 Mbit/s
536 pixels high	53	1.3 GB		0.81 Mbit/s
520 pixels high	fewer than 50	fewer than 50		fewer than 50
506 pixels high	70	0.88 GB		2.3 Mbit/s
480 pixels high	fewer than 50	fewer than 50		fewer than 50
428 pixels high	fewer than 50	fewer than 50		fewer than 50
408 pixels high	98	0.37 GB		0.61 Mbit/s
402 pixels high	258	0.97 GB		0.52 Mbit/s
360 pixels high	fewer than 50	fewer than 50		fewer than 50
346 pixels high	fewer than 50	fewer than 50		fewer than 50
338 pixels high	110	0.13 GB		0.68 Mbit/s
286 pixels high	fewer than 50	fewer than 50		fewer than 50
272 pixels high	fewer than 50	fewer than 50		fewer than 50
268 pixels high	fewer than 50	fewer than 50		fewer than 50
AV1 HDR	fewer than 50	fewer than 50	fewer than 50	fewer than 50	fewer than 50
720 pixels high	fewer than 50	fewer than 50		fewer than 50
540 pixels high	fewer than 50	fewer than 50		fewer than 50
506 pixels high	fewer than 50	fewer than 50		fewer than 50
338 pixels high	fewer than 50	fewer than 50		fewer than 50
HEVC	131	3.2 GB	3.4	2.1 Mbit/s	24 MB
1440 pixels high	fewer than 50	fewer than 50		fewer than 50
1080 pixels high	fewer than 50	fewer than 50		fewer than 50
1078 pixels high	fewer than 50	fewer than 50		fewer than 50
1038 pixels high	fewer than 50	fewer than 50		fewer than 50
818 pixels high	fewer than 50	fewer than 50		fewer than 50
720 pixels high	fewer than 50	fewer than 50		fewer than 50
640 pixels high	fewer than 50	fewer than 50		fewer than 50
540 pixels high	fewer than 50	fewer than 50		fewer than 50
480 pixels high	fewer than 50	fewer than 50		fewer than 50
360 pixels high	61	0.14 GB		0.27 Mbit/s
HEVC HDR	fewer than 50	fewer than 50	fewer than 50	fewer than 50	fewer than 50
H.264	52	0.51 GB	0.68	1.7 Mbit/s	9.9 MB
720 pixels high	fewer than 50	fewer than 50		fewer than 50
692 pixels high	fewer than 50	fewer than 50		fewer than 50
676 pixels high	fewer than 50	fewer than 50		fewer than 50
572 pixels high	fewer than 50	fewer than 50		fewer than 50
536 pixels high	fewer than 50	fewer than 50		fewer than 50
480 pixels high	fewer than 50	fewer than 50		fewer than 50
360 pixels high	fewer than 50	fewer than 50		fewer than 50
346 pixels high	fewer than 50	fewer than 50		fewer than 50
338 pixels high	fewer than 50	fewer than 50		fewer than 50
286 pixels high	fewer than 50	fewer than 50		fewer than 50
272 pixels high	fewer than 50	fewer than 50		fewer than 50
268 pixels high	fewer than 50	fewer than 50		fewer than 50
H.264 in MPEG-TS (one rendition)	501	0.83 GB	2.5	0.74 Mbit/s	1.7 MB
Stereo sound (AAC)	930	1.2 GB	19	0.13 Mbit/s	1.2 MB
Spatial sound (IAMF, Opus)	302	0.59 GB	4.1	0.32 Mbit/s	1.9 MB

Plays: Requests for a title's master playlist (title plays), or for a rendition's own playlist, which a player fetches once per play that uses that rendition. A browser that keeps a playlist (five minutes) does not ask again.
Sent, hours, bitrate: Body bytes of the segments and playlists the servers sent, and the minutes of media in those segments (each at its playlist's mean segment length). Initialisation segments are not counted (see the edge, next).
The edge: Checked on 21 October 2026 by requesting through Cloudflare and reading its cf-cache-status: playlists (.m3u8) and video and sound segments (.m4s) answered DYNAMIC, not kept at the edge; initialisation segments (init.mp4) answered HIT, kept by their .mp4 extension. So every playlist and segment a player fetched reached our servers and is counted here, less what browsers kept. The servers cannot see the edge's cache status themselves: Cloudflare adds that header on its way out. "Through Cloudflare" is the share of media requests that arrived with Cloudflare's cf-ray; the rest came straight to a server (tests, the other server).
Text compressed ahead: Scripts, styles and data sent in Brotli made once at its highest setting, or as a delta against the copy a returning visitor already has (compression dictionaries): the share of bytes saved against the files uncompressed. Pages themselves are compressed as they stream and are not counted.
Pages' LCP: The share of page views whose Largest Contentful Paint, measured in visitors' browsers, was within the site's budget of 1.5 s, and within 2.5 s (Core Web Vitals' "good").
Players' reports: From the CMCD headers players send with each segment request: requests made after the buffer ran dry, per thousand segments, and the mean bitrate asked for. Where players send CMCD version 2 (the site's own player does from 26 October 2026): the mean bitrate on screen, the mean time from pressing play to the picture moving, the share of the time stalled after starting (against the video in the segments those players fetched, so a little under the share of time watched), and frames dropped per minute of that video. Sums only: the counts a player keeps for its stream are subtracted from its previous ones as they come, and nothing of them is kept but the day's totals.
What is kept: Sums per day, per media family and picture height: no address, no identifier, no cookie, no title, no time finer than the day. Kept 400 days, then deleted. (Plays per work, next, are kept per day the same way.)

Measured here: how widely plays spread

How many of each catalogue's works the plays of the last 30 days reached, and how evenly: plays per work, counted per day with nothing about who. Both servers' plays, added together before any figure is shown.

2026-09-28 to 2026-10-27 UTC. A medium with fewer than 50 plays is not shown, nor a total that would reveal one by subtraction ("withheld"). No work is named.

Each medium: its plays, how many of its works they reached, and how evenly
Medium	Plays	Works played	In the catalogue	Tail coverage	Exposure Gini	Entropy	Effective catalogue size	Following the catalogue's own counts
Films and shows	212	26	35	74%	0.62	4.1 bits	13	no honest baseline
Films from the archives	fewer than 50 plays		609,183
Lectures	fewer than 50 plays		35,551
Radio drama	fewer than 50 plays		17,707
Podcasts	fewer than 50 plays		71,478
Music	fewer than 50 plays		17,594
Books	fewer than 50 plays		78,333
Audiobooks	fewer than 50 plays		21,822
Art	112	89	1,468,171	0.0061%	1.00	6.0 bits	71	no honest baseline
Live channels	67	40	4,611	0.87%	0.99	4.9 bits	26	no honest baseline

All media together: 497 plays.

A play: A work began to play in a visitor's browser, or, for a picture or a book, its page was on screen for a second; a moment in Moments counts for its work. The browser sends the work's page address and nothing else, and the server counts it only for a work in its catalogue. Crawlers, pages from other sites and prerendered pages never shown are not counted.
In the catalogue: The works the site lists for an adult profile, as its own pages list them: not withdrawn, not found broken and not failing their checks; the same definition for every medium. A play counts only for such a work, so what is played and what it is a share of are the same works.
Tail coverage: The share of the catalogue with at least one play in these days.
Exposure Gini: The Gini coefficient of plays per work over the whole catalogue, works never played included: 0 when every work has as many plays, close to 1 when one work has them all.
Entropy: The Shannon entropy of the plays over works, in bits: how many yes-or-no questions it takes, on average, to name the work a play went to. It is at most log₂ of the works played.
Effective catalogue size: 2 Σ p_i·i − 1 over the works' shares of plays, largest first (Gomez-Uribe and Hunt, ACM TMIS 2015, appendix A): 1 when one work takes every play, n when n works share them alike.
Following the catalogue's own counts: The effective catalogue size the same number of plays would have if each went to a work in proportion to the catalogue's own crowd counts: a popularity-only baseline. Shown for Project Gutenberg's books (its downloads) and LibriVox's recordings (the Internet Archive's downloads), where one source counts every work in one unit; the other catalogues count in mixed units across their sources, or not at all, so no honest baseline can be made for them. Each is the mean of five seeded draws.
Why these: The site spends what people attend to on breadth and keeps crowd counts from deciding what is met (laws 2 and 3 of its plan); these are the measures those laws name.
What is kept: For each day, how many plays each work had, by its catalogue id; never who played it, from where, or when within the day. Kept 400 days, like the rest, then deleted.

Energy on our servers

Measured on these servers (OnChipCounter, the kernel's own count of the server process's CPU energy): 0.001 Wh (0.00038–0.35) per viewing hour, its share for the media sent, ±0.19% (standard uncertainty, k = 1, of the part that can be evaluated) over 114 hours metered. For comparison, modelled, not measured: the Carbon Trust's 1.3 Wh per viewing hour for data centres and CDN together (DIMPACT members, 2020; Netflix's delivery-only 0.028 Wh at the low end). The method and its limits are on the site's Energy page.

How it is measured

The servers are Apple Mac Studios. Each reads, every second and without root, the kernel's own count of its server process's CPU energy (proc_pid_rusage's ri_energy_nj), the per-task counter of WAI's energy measurement; a minute counts only where the counter was live in it. That energy is split to the media sent 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 charged to the media too. It covers the CPU only, not memory, disks, the network card, the machine's idle draw or the path beyond the servers (Caddy, the tunnel, Cloudflare), which is not metered here and never counted as zero. Where no counter reads, there is no figure.

Grid carbon

Energy becomes grams of CO₂ through the grid it is drawn from. Average intensities only, from sources whose terms allow showing them.

Great Britain over the last 24 hours: from 52 to 120 g CO2 per kWh, half hour by half hour (49 half hours; lighter bars are forecasts). Fetched 2026-10-27 18:26 UTC. Source: Carbon Intensity API, National Energy System Operator, with EDF, the University of Oxford and WWF, under CC BY 4.0. Direct emissions of generation: wind, solar and nuclear count as zero.

United States: 380 g CO2e per kWh in 2025, 500 in 2015 (Ember, life-cycle emissions of generation, yearly). Data read 2026-10-21. Source: Ember, yearly electricity data, under CC BY 4.0: the emissions intensity of each country's total generation, shown as Ember publishes it.

Not used: Electricity Maps' free API (non-commercial use only) and WattTime (no redistribution without a licence), whose terms do not allow this page; the US Energy Information Administration's hourly data is public domain, but its API needs an account and its files give generation by fuel, not emissions, so the United States is shown by Ember's yearly figure. Marginal grid signals are contested and are not shown.

Our sources: answering, and lasting

Most of the library plays from other archives' servers. On 21 October 2026 the Internet Archive was down for hours and this site had no way to say so. Now each source is asked about one of its own works every few minutes, a page's headers or a file's first two bytes; nothing about anyone who visits is counted. Each of the site's two servers asks on its own: these are this server's answers.

Each square is a UTC day, 28 September 2026 to 27 October 2026: every probe answered, some failed (at least 90% answered), many failed, or none made. Hover a square for its counts; each row says them in words too.

Each source: its state now, and the last 30 days
Source	State	The last 30 days
Internet Archivefilms, radio drama, print, audiobooks' sound, free music, lectures and card pictures	Answering since 22 October 2026, 22:55 UTCLast asked 27 October 2026, 18:37 UTC: answered in 997 ms.	Answered 1,627 of 1,639 probes over 30 days (99.3%). Last outage: 22 October 2026, 17:11 UTC to 22 October 2026, 17:13 UTC (2 min).
Wayback Machinearchived copies of works gone from their sources	Answering since 25 October 2026, 22:51 UTCLast asked 27 October 2026, 18:26 UTC: answered in 2,403 ms.	Answered 551 of 569 probes over 30 days (97%). Last outage: 22 October 2026, 19:15 UTC to 22 October 2026, 19:17 UTC (2 min).
Wikimedia Commonsfilms, pictures, print, music and historic broadcasts	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:36 UTC: answered in 184 ms.	Answered 1,636 of 1,636 probes over 30 days (100%).
Library of Congressfilms, newspapers and spoken-word recordings	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:33 UTC: answered in 78 ms.	Answered 822 of 822 probes over 30 days (100%).
NASA Image and Video Libraryfilms	Answering since 22 October 2026, 01:56 UTCLast asked 27 October 2026, 18:39 UTC: answered in 188 ms.	Answered 822 of 822 probes over 30 days (100%).
Open Imagesfilms (the Netherlands Institute for Sound and Vision)	Not answering since 27 October 2026, 09:04 UTCLast asked 27 October 2026, 18:39 UTC: robots.txt: no answer in 8 s.	Answered 525 of 3,356 probes over 30 days (16%). Not answering since 27 October 2026, 09:03 UTC.
Project Gutenbergbooks	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:38 UTC: answered in 208 ms.	Answered 826 of 826 probes over 30 days (100%).
Gutenberg mirror (PGLAF)books' texts, read from the mirror first as Gutenberg asks	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:32 UTC: answered in 250 ms.	Answered 822 of 823 probes over 30 days (99.9%).
LibriVoxaudiobooks' catalogue and chapters (their sound is at the Internet Archive)	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:38 UTC: answered in 421 ms.	Answered 827 of 827 probes over 30 days (100%).
Standard Ebooksbooks	Answering since 27 October 2026, 06:13 UTCLast asked 27 October 2026, 18:36 UTC: answered in 192 ms.	Answered 822 of 857 probes over 30 days (96%). Last outage: 27 October 2026, 06:11 UTC to 27 October 2026, 06:13 UTC (2 min).
Wikisourcebooks	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:37 UTC: answered in 138 ms.	Answered 823 of 823 probes over 30 days (100%).
Wikibooksopen textbooks	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:38 UTC: answered in 136 ms.	Answered 819 of 820 probes over 30 days (99.9%).
LibreTextsopen textbooks	Answering since 27 October 2026, 17:53 UTCLast asked 27 October 2026, 18:33 UTC: answered in 2,919 ms.	Answered 807 of 841 probes over 30 days (96%). Last outage: 27 October 2026, 17:38 UTC to 27 October 2026, 17:53 UTC (16 min).
OpenStaxopen textbooks	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:34 UTC: answered in 403 ms.	Answered 820 of 820 probes over 30 days (100%).
Open Textbook Libraryopen textbooks	Answering since 24 October 2026, 23:31 UTCLast asked 27 October 2026, 18:40 UTC: answered in 311 ms.	Answered 815 of 871 probes over 30 days (94%). Last outage: 24 October 2026, 22:37 UTC to 24 October 2026, 23:31 UTC (55 min).
Rijksmuseumart (its pictures through Micrio's image server)	Answering since 22 October 2026, 02:43 UTCLast asked 27 October 2026, 18:38 UTC: answered in 277 ms.	Answered 817 of 820 probes over 30 days (99.6%).
Art Institute of Chicagoart	Answering since 27 October 2026, 13:58 UTCLast asked 27 October 2026, 18:40 UTC: answered in 214 ms.	Answered 827 of 830 probes over 30 days (99.6%). Last outage: 27 October 2026, 13:57 UTC to 27 October 2026, 13:58 UTC (2 min).
National Gallery of Artart	Answering since 26 October 2026, 03:35 UTCLast asked 27 October 2026, 18:34 UTC: answered in 231 ms.	Answered 824 of 834 probes over 30 days (99%). Last outage: 26 October 2026, 03:32 UTC to 26 October 2026, 03:35 UTC (4 min).
SMKart	Answering since 26 October 2026, 23:35 UTCLast asked 27 October 2026, 18:33 UTC: answered in 713 ms.	Answered 783 of 1,179 probes over 30 days (66%). Last outage: 26 October 2026, 23:33 UTC to 26 October 2026, 23:35 UTC (2 min).
Gettyart	Not askedNot asked: media.getty.edu's robots.txt answers HTTP 503, which RFC 9309 reads as keep away while it does.	Answered 0 of 40 probes over 30 days (0%). Last outage: 22 October 2026, 01:57 UTC to 22 October 2026, 02:43 UTC (47 min).
Metropolitan Museum of Artart	Answering since 23 October 2026, 07:54 UTCLast asked 27 October 2026, 18:32 UTC: answered in 206 ms.	Answered 822 of 828 probes over 30 days (99.3%). Last outage: 23 October 2026, 07:53 UTC to 23 October 2026, 07:54 UTC (2 min).
Cleveland Museum of Artart	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:39 UTC: answered in 166 ms.	Answered 825 of 825 probes over 30 days (100%).
CERNlectures	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:32 UTC: answered in 672 ms.	Answered 821 of 821 probes over 30 days (100%).
TEDtalks	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:38 UTC: answered in 101 ms.	Answered 822 of 822 probes over 30 days (100%).
MIT OpenCourseWarecourses	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:37 UTC: answered in 85 ms.	Answered 827 of 827 probes over 30 days (100%).
NPTELcourses	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:38 UTC: answered in 1,375 ms.	Answered 819 of 819 probes over 30 days (100%).
Gresham Collegelectures	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:34 UTC: answered in 820 ms.	Answered 823 of 823 probes over 30 days (100%).
Open Yale Coursescourses	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:31 UTC: answered in 132 ms.	Answered 823 of 823 probes over 30 days (100%).
CS50courses	Answering since 22 October 2026, 02:43 UTCLast asked 27 October 2026, 18:36 UTC: answered in 206 ms.	Answered 821 of 821 probes over 30 days (100%).
Stanford Engineering Everywherecourses	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:34 UTC: answered in 151 ms.	Answered 827 of 827 probes over 30 days (100%).
YouTubethe embedded players of lectures published there	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:35 UTC: answered in 161 ms.	Answered 822 of 822 probes over 30 days (100%).
Apple Podcasts directorypodcasts' charts, records and search	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:38 UTC: answered in 218 ms.	Answered 553 of 553 probes over 30 days (100%).
Radio Browserthe world radio directory	Answering since 22 October 2026, 02:43 UTCLast asked 27 October 2026, 18:33 UTC: answered in 227 ms.	Answered 549 of 549 probes over 30 days (100%).
iptv-orgthe world TV directory	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:26 UTC: answered in 82 ms.	Answered 555 of 555 probes over 30 days (100%).
Carbon Intensity APIGreat Britain's grid, on the telemetry page	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:30 UTC: answered in 530 ms.	Answered 552 of 552 probes over 30 days (100%).
Emberevery country's yearly grid, on the telemetry page	Answering since 22 October 2026, 01:57 UTCLast asked 27 October 2026, 18:29 UTC: answered in 225 ms.	Answered 554 of 555 probes over 30 days (99.8%).

How long works last at their sources

Of the pages on the web in 2013, 38% could not be reached in 2023 (Pew Research Center, 2024); of the American feature films of 1912 to 1929, 14% survive as they were made (Pierce, for the Library of Congress, 2013). Here, each catalogue's works from each source are followed in a fixed sample of 200, the same works every week: a work whose page answers "not found" is asked again over a week before it is counted gone, and one that does not answer is unknown that week, never gone. A half-life is estimated only once the checks have run eight weeks, and as a lower bound until five works have gone.

Wikimedia Commons, all 5 catalogues together. 851 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Internet Archive, all 5 catalogues together. 1,000 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Library of Congress, all 3 catalogues together. 600 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Art, Art Institute of Chicago. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Art, Cleveland Museum of Art. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Art, Getty. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Art, National Gallery of Art. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Art, Rijksmuseum. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Art, SMK. 191 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Art, Wikimedia Commons. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Books, LibriVox. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Books, Project Gutenberg. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Books, Standard Ebooks. 17 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Films, Internet Archive. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Films, Library of Congress. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Films, NASA Image and Video Library. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Films, Open Images. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Films, Wikimedia Commons. 198 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Free music, Internet Archive. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Free music, Wikimedia Commons. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Lectures, CERN. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Lectures, CS50. 13 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Lectures, Gresham College. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Lectures, Internet Archive. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Lectures, MIT OpenCourseWare. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Lectures, NPTEL. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Lectures, Open Yale Courses. 39 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Lectures, Stanford Engineering Everywhere. 9 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Lectures, TED. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Open textbooks, LibreTexts. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Open textbooks, Open Textbook Library. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Open textbooks, OpenStax. 84 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Open textbooks, Wikibooks. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Podcasts, Publishers' feeds. 195 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Print, Internet Archive. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Print, Library of Congress. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Print, Wikimedia Commons. 53 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Radio drama, Internet Archive. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Radio drama, Library of Congress. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.
Radio drama, Wikimedia Commons. 200 works checked weekly since 22 October 2026 (1 check); 0 confirmed gone. Too little history to estimate a half-life: an estimate needs 8 weeks of checks, so it can begin to mean something from 17 December 2026.

Works missing at their sources now: 5 suspected, 0 confirmed gone. A gone work keeps its page and its address here, says since when it has not been available, and links an archived copy where one is found (the Wayback Machine, or the source's own mirror), named as a copy.

1 January

On 1 January 2026, works published in 1930 entered the public domain in the United States, and sound recordings first published in 1925. On 1 January 2027 (in 66 days), works published in 1931 enter, and sound recordings first published in 1926. At midnight UTC the site's catalogues that state public-domain status from a work's year (Print's magazines, periodicals and newspapers) read their listings again, and the days until the year's works are listed are measured here. The rules, as the law documents them: the Copyright Office's Circular 15A and 17 U.S.C. § 304 for published works (95 years from publication, to the end of the year), § 1401 and the Office's summary for sound recordings, and Duke's Center for the Study of the Public Domain.

As JSON: /api/sources; the same counts are at the main site's /metrics.

Every coefficient

What the model multiplies, with the range each source gives and how it was read. "Derived" means our arithmetic from the cited figures; no source states it.

Coefficients of the model
Coefficient	Value (range)	Basis	How it was read, and sources
TV over 50 inches, whole set, SDR, default picture	117.4 W (85–175)	measured	Field average 117.4 W for sets over 50 inches. Low: a 70-inch set at 85 W on SD content. High: 175 W, the top of a 55-inch set measured at 117-175 W. Fraunhofer FOKUS and Deutsche Telekom (9 Sep 2025); Dom Robinson, Greening of Streaming (May 2023); Jens Malmodin (Ericsson) (2023)
TV of 50 inches or smaller, whole set, SDR, default picture	55.9 W (29–100)	measured	Field average 55.9 W for sets of 50 inches or smaller. Low: about 29 W, the EU label's SDR on-mode power of a 40-inch set. High: the Carbon Trust's assumed 100 W. Fraunhofer FOKUS and Deutsche Telekom (9 Sep 2025); European Commission, EPREL product database (read 21 Oct 2026); The Carbon Trust (DIMPACT) (June 2021)
TV of a size not known, whole set, SDR, default picture	100 W (29–175)	assumed by the source	A browser cannot tell a TV's size. Central: the Carbon Trust's assumed 100 W for a TV. Low and high: the ends of the two size classes, 29 W (the EU label's SDR on-mode power of a 40-inch set) and 175 W (the top of a 55-inch set measured at 117-175 W). The Carbon Trust (DIMPACT) (June 2021); European Commission, EPREL product database (read 21 Oct 2026); Fraunhofer FOKUS and Deutsche Telekom (9 Sep 2025)
TV decoding HDR in software, added	50 W (50–60)	measured	A 70-inch set drew 75-85 W with its built-in decoder and 135 W when HDR fell back to its CPU, which that decoder could not take: 50 W more (one set, a handheld meter; high: 135 W against 75 W). Phil Kurz, TV Tech, reporting Dom Robinson's measurements (6 Mar 2023)
Hardware video decoder, HD at 30 fps	0.719 W	measured	0.719 W decoding HD at 30 frames a second and 2 Mbit/s in hardware, on an evaluation platform; HEVC in software on the same platform drew 24.45 W. Across six codecs, hardware decoders used less than 9% of the energy of software ones. Herglotz, Kränzler, Schober and Kaup, IEEE Circuits and Systems Magazine (preprint) (30 Sep 2022); Kränzler, Herglotz and Kaup, PCS 2024 (14 Feb 2024)
Laptop, whole device, hardware decoding	13.5 W (13.5–22)	measured	13.5 W measured at the wall on a Surface Book 2 playing H.264 in hardware. High: the Carbon Trust's assumed 22 W. Herglotz et al., QoMEX 2022 (11 Oct 2022); The Carbon Trust (DIMPACT) (June 2021)
Phone, whole device, a typical setting	2.22 W (1–2.76)	measured	One phone measured over an hour of YouTube: 2.22 Wh at a typical setting (75% brightness, 1080p, half volume) and 2.76 Wh at its highest (full brightness, 1440p, full volume). Low: the Carbon Trust's 1 W (battery capacity over hours of video). Fraunhofer FOKUS assumes 2 W. Soltani Nezhad et al., TNO and Radboud University (30 Jun 2025); The Carbon Trust (DIMPACT) (June 2021); Fraunhofer FOKUS (Oct 2024)
TV eco mode or low backlight, against the default picture	0.63× (0.41–0.63)	measured	Eco mode saved about 37% on average in the field test, and was off by default on nearly every set. Low: 51 W at 20% backlight against 120-130 W at the default (derived). Fraunhofer FOKUS and Deutsche Telekom (9 Sep 2025); Accedo (31 Aug 2023)
TV with energy saving off (vivid), against the default picture	1.48× (1.31–1.67)	derived	170-200 W with energy saving off against 120-130 W at the default, on the same smart TVs (ratios derived). Accedo (31 Aug 2023)
Phone at 50% brightness, a step down	0.79×	measured	From 100% to 50% brightness the phone drew 21% less (0.58 Wh in an hour, its other settings at their highest). One phone and one measured step, applied once downward from the typical setting; from 75% the saving may be smaller, and no measurement separates it. Soltani Nezhad et al., TNO and Radboud University (30 Jun 2025)
Phone at full brightness, a step up	1.243×	derived	The phone's highest setting against its typical one: 2.76 Wh against 2.22 Wh in an hour (derived: 1.243 times). Soltani Nezhad et al., TNO and Radboud University (30 Jun 2025)
TV showing HDR, against SDR	1.42× (1.29–1.78)	measured	HDR content drew 29-78% more than the SDR test clip on four TVs in the lab, and 42% more in retail tests reported in the same study; sets raised their backlight by themselves. Across EU label records the HDR premium runs from 0% to 131% by model. Northwest Energy Efficiency Alliance (report E16-298) (16 Nov 2016); European Commission, EPREL product database (read 21 Oct 2026)
TV power across stream resolutions and bitrates	1× (0.96–1.04)	measured	Changing a stream's resolution or bitrate moved total TV power by 1-4%, not significant at p<0.05. An HD stream upscaled on some 4K sets drew 4-10 W more, so lower is not always less. Greening of Streaming (Feb 2026); Fraunhofer FOKUS (Oct 2024)
Phone playing 720 lines or fewer, against 1440	0.92× (0.92–1)	measured	From 1440p to 720p the phone drew about 8% less (0.23 Wh in an hour). Applied to renditions of 720 lines or fewer; no saving is claimed between 720 and 1440. Soltani Nezhad et al., TNO and Radboud University (30 Jun 2025)
Laptop decoding AV1 in software, added	4.78 W (2.97–4.85)	derived	VP9 in software drew 17.6 W against 13.5 W for H.264 in hardware (+4.1 W); AV1's software decoder uses 16.6% more energy than VP9's (derived: 4.8 W). The range is the spread across the three codecs. Herglotz et al., QoMEX 2022 (11 Oct 2022); Kränzler, Herglotz and Kaup, PCS 2024 (14 Feb 2024)
Laptop decoding HEVC in software, added	4.85 W (2.97–4.85)	derived	+4.1 W for VP9 in software over H.264 in hardware; HEVC's software decoder uses 18.3% more than VP9's (derived: 4.9 W). Herglotz et al., QoMEX 2022 (11 Oct 2022); Kränzler, Herglotz and Kaup, PCS 2024 (14 Feb 2024)
Laptop decoding H.264 in software, added	2.97 W (2.97–4.85)	derived	+4.1 W for VP9 in software over H.264 in hardware; H.264's software decoder uses 27.6% less than VP9's (derived: 3.0 W). Herglotz et al., QoMEX 2022 (11 Oct 2022); Kränzler, Herglotz and Kaup, PCS 2024 (14 Feb 2024)
Home modem and Wi-Fi router, always on	12.7 W (8–18)	measured	A fibre modem and Wi-Fi router drew 12.7 W, varying about 1% whatever the use. Low: 8 W, a router while streaming in the hackathon. High: 18 W, a cable line's home equipment per subscriber. Jens Malmodin (Ericsson) (2023); Greening of Streaming (Feb 2026); Prysmian (n.d.)
What streaming adds to the home equipment	0.15 W (0.1–4)	measured	Streaming added 0.1-0.2 W in the Swedish home. One router in the hackathon went from 4 W idle to 8 W streaming, the high end. Jens Malmodin (Ericsson) (2023); Greening of Streaming (Feb 2026)
Fixed access and core network, per Mbit/s	0.05 W per Mbit/s (0.05–0.057)	power model	Malmodin's 2020 coefficients as the Carbon Trust reproduces them: about 18 W per line at rest, 23 W at 100 Mbit/s. About 0.1 Wh per GB (derived). The Carbon Trust (DIMPACT) (June 2021); Mytton, Lundén and Malmodin, Journal of Industrial Ecology 28(4) (Aug 2024)
Fixed network outside the home, per household, always on	2.8 W (0.75–7.1)	measured	Per household: street cabinet about 2 W, metro 0.5 W, core 0.3 W (measured). Low: a fibre line's access share, 0.75 W. High: 7.1 W, European operators' fixed-network electricity per subscription (62 kWh a year, derived). Jens Malmodin (Ericsson) (2023); Prysmian (n.d.); Lundén et al., Sustainability 14(5) (2022)
4G radio and core network, per Mbit/s	1.53 W per Mbit/s (1.03–2.03)	power model	Access 1.5 W and core 0.03 W per Mbit/s; the radio's slope is 1-2 W per Mbit/s on 4G-era equipment. About 3.4 Wh per GB (derived). The Carbon Trust (DIMPACT) (June 2021)
4G network per subscriber, always on	1.2 W (1.2–4.2)	power model	Access 1 W and core 0.2 W per subscriber in the power model. High: about 4.2 W per subscriber at a base station (Malmodin 2023). The Carbon Trust (DIMPACT) (June 2021); Jens Malmodin (Ericsson) (2023)
CDN and data centres, per viewing hour	1.3 Wh (0.028–1.3)	measured	1.3 Wh per viewing hour: the DIMPACT members' data centres and CDN in 2020, encoding and storage included (about 1% of the hour). Low: 0.028 W per stream on Netflix's Open Connect appliances, delivery alone. The Carbon Trust (DIMPACT) (June 2021); Netflix (28 Apr 2023)
One more stream from an edge cache	0.028 Wh (0–0.028)	derived	A CDN server's draw is mostly fixed (one measured at about 450 W, plus about 50 W at the evening peak), so one more stream adds less than its average share: at most 0.028 Wh in an hour. Netflix (28 Apr 2023); Fraunhofer FOKUS (Oct 2024)
One more stream that the edge does not keep	0.028 Wh (0–1.3)	derived	A request the edge passes on also uses the origin server and the backbone. No published figure separates that part; the high end is the whole data-centre and CDN share of a viewing hour. Netflix (28 Apr 2023); The Carbon Trust (DIMPACT) (June 2021)
Encoding one frame of one rendition	1.29 J (0.47–16.67)	derived	x264 at its medium preset (one of this site's encoders) drew 77.4 W encoding 1080p60 in real time: 1.29 J a frame. Low: 0.47 J with a hardware (Quick Sync) AVC encoder. High: 16.67 J, x265 medium per 1080p frame on a 2017 Xeon (that paper's absolute figures run high for its CPU). Each rendition is counted as a 1080p encode, which overstates the small ones. Arunruangsirilert and Katto (24 Nov 2025); Silveira, Porto and Bampi, EUSIPCO 2017 (2017)
A bytes-based model: Sustainable Web Design v4, per GB	0.3 kWh per GB	derived	Data centre 0.055 + 0.012, network 0.059 + 0.013, device 0.080 + 0.081 kWh per GB (operational + embodied): about 0.30 (derived). Shown only to compare: it scales with bytes, which network power does not. Sustainable Web Design (n.d.)

Cited works

Putting sustainable streaming to the test (field test in homes, per-second smart plugs), Fraunhofer FOKUS and Deutsche Telekom, 9 Sep 2025.
Just measure IT! (one Swedish home, measured continuously from November 2022), Jens Malmodin (Ericsson), 2023.
Energy label records of electronic displays (SDR and HDR on-mode power), European Commission, EPREL product database, read 21 Oct 2026.
Revising the TV energy use test procedure: incorporating HDR, Northwest Energy Efficiency Alliance (report E16-298), 16 Nov 2016.
Hackathon data analysis (TV and router power at the plug), Greening of Streaming, Feb 2026.
The LESS Accord (home test of a 70-inch TV), Dom Robinson, Greening of Streaming, May 2023.
Quantifying the energy of video on a smartphone (poster), Soltani Nezhad et al., TNO and Radboud University, 30 Jun 2025.
Carbon impact of video streaming, The Carbon Trust (DIMPACT), June 2021.
Green streaming white paper: intensity and power models side by side, Fraunhofer FOKUS, Oct 2024.
Power of YouTube playback on a laptop, measured at the wall, Herglotz et al., QoMEX 2022, 11 Oct 2022.
Decoding energy of AVC, HEVC, VP9, AV1 and VVC, in software and hardware, Kränzler, Herglotz and Kaup, PCS 2024, 14 Feb 2024.
Report on the energy use of video streaming (smart TV picture settings), Accedo, 31 Aug 2023.
Energy consumption white paper (access and home equipment per subscriber), Prysmian, n.d..
Electricity consumption and operational carbon emissions of European telecom network operators, Lundén et al., Sustainability 14(5), 2022.
Energy of real-time 1080p60 encoding in hardware and software, Arunruangsirilert and Katto, 24 Nov 2025.
x265 encoding energy per frame, by preset, Silveira, Porto and Bampi, EUSIPCO 2017, 2017.
Energy efficiency in streaming (Open Connect appliances), Netflix, 28 Apr 2023.
Electricity intensity of internet data transmission, Aslan, Mayers, Koomey and France, Journal of Industrial Ecology 22(4), Aug 2018.
Network energy use not directly proportional to data volume: the power model approach, Mytton, Lundén and Malmodin, Journal of Industrial Ecology 28(4), Aug 2024.
Network power is almost independent of traffic; peak traffic drives capacity, Schien, Shabajee, Krug, McSorley and Preist, Journal of Industrial Ecology 29(5), 2025.
Estimating digital emissions (model v4, the default in CO2.js since v0.18), Sustainable Web Design, n.d..
Factcheck: what is the carbon footprint of streaming video on Netflix?, Carbon Brief, with the IEA's George Kamiya, 25 Feb 2020, updated 25 Nov 2020.
The unbearable slowness of being: why do we live at 10 bits/s?, Jieyu Zheng and Markus Meister (accepted at Neuron), 3 Aug 2024, revised 15 Nov 2024.
Carbon emissions of streaming and digital terrestrial television, Carnstone for Ofcom, Oct 2022, revised Jul 2024.
Energy estimates for broadcast and streaming in the UK (viewing devices 58.0, home equipment 21.5, peripherals 21.5, network 12.0 Wh), 3GPP TR 26.942, section 4.2.7, summarising Ofcom's study, read 22 Oct 2026.
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), Herglotz, Kränzler, Schober and Kaup, IEEE Circuits and Systems Magazine (preprint), 30 Sep 2022.
The greening of streaming: is less really more? (a 70-inch TV at 75-85 W, 135 W decoding HDR in software), Phil Kurz, TV Tech, reporting Dom Robinson's measurements, 6 Mar 2023.
Green streaming decoded: myths and reality (a 55-inch OLED TV, 540p at 5 Mbit/s against 2160p60 at 25 Mbit/s), Robert Seeliger, Fraunhofer FOKUS, HbbTV Symposium, 14 Nov 2024.
An experimental study of Netflix use and the effects of autoplay on watching behaviors (76 people, 10-17 days), Schaffner, Ulloa, Sahni, Li, Cohen, Messier, Gao and Chetty (University of Chicago), PACM HCI (CSCW), 20 Dec 2024.
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), Arcom, 26 Jul 2023.
Study of the environmental impact of audiovisual uses in France (lower picture quality among the sobriety measures), Arcom, Arcep and ADEME, Oct 2024.
Amazon CloudFront pay-as-you-go pricing (data transfer out to the internet, per GB), Amazon Web Services, read 22 Oct 2026.
WAI extension: energy measurement (acquisition classes, declared uncertainty, attribution), WAI, Transaction Science, Inc., read 22 Oct 2026.
Carbon Intensity API (Great Britain, half-hourly; CC BY 4.0), National Energy System Operator, with EDF, the University of Oxford and WWF, live.
Yearly electricity data (CC BY 4.0), Ember, yearly.

The research behind this page, with the disputes between models and what each figure is fit for, is in the site's repository (docs/research/telemetry-energy.md and telemetry-delivery.md). This page as Markdown: /index.md; the same hour as JSON: /api/model, and for agents the get_energy_model tool at https://invisiblenetwork.ai/mcp.