BRENNEN

Industries / Media, arts and recreation / Media and entertainment / Streaming

Encoding and delivery

Encoding and delivery is the work of compressing video and sound, cutting it into streams that adapt to each viewer's connection, and moving it across the internet to screens and speakers.

7

Milestones achieved, from 2009 to 2020.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2009

    Streams that adapt

    On 1 May 2009 Roger Pantos of Apple posted the first draft of HTTP Live Streaming (HLS), "a protocol for transmitting unbounded streams of multimedia data over HTTP," the ordinary protocol of the web. In HLS a video is offered at several quality levels in short pieces, so the player can step up or down as the connection changes, and ordinary web servers can deliver it. The international standards bodies ISO and IEC published a similar method, MPEG-DASH, in April 2012. HLS was published as RFC 8216 in August 2017, which notes that it "has been implemented and deployed" since the first draft.

    datatracker.ietf.orgdatatracker.ietf.orgiso.org

    Credit

    Roger Pantos (Apple) and William May (MLB Advanced Media), authors of RFC 8216; MPEG, the ISO/IEC moving-picture group, for DASH

  2. 2010

    From plugin to browser

    In April 2010 Steve Jobs of Apple published "Thoughts on Flash," explaining why iPhones would not run Adobe's Flash plugin, then the usual way to play video on the web. He wrote that "Flash was created during the PC era – for PCs and mice" and that decoding H.264 video in hardware gave "up to 10 hours" of playback against "less than 5 hours" in software. YouTube made the browser's own HTML5 video player its default on 27 January 2015, once browsers could switch quality levels mid-stream, and said this cut buffering "by more than 50 percent globally." Adobe stopped supporting Flash Player on 31 December 2020.

    LessonThe change took more than ten years, and it came only when open web standards could do what the plugin did, including switching quality levels.

    web.archive.orgweb.archive.orgweb.archive.org

    Credit

    Steve Jobs (Apple); YouTube's engineers; Adobe

  3. 2012

    Servers inside the network

    On 4 June 2012 Netflix announced Open Connect, its own content delivery network. Netflix places cache servers holding copies of its titles inside internet providers' networks "at no cost" to them, or connects at internet exchange points, so that viewing traffic travels a short distance. Netflix said "About 5% of Netflix data is already being served by Open Connect," and noted that YouTube "has long had its own content delivery network." By Netflix's count in December 2022 it had 18,000 servers in 6,000 locations across 175 countries, and it estimated that providers avoided $1.25 billion in costs in 2021.

    web.archive.orgabout.netflix.com

    Credit

    Netflix (Ken Florance, vice president for content delivery)

  4. 2015

    Settings chosen per title

    Netflix had encoded every title with the same ladder of bitrate and resolution steps, topping out at 5,800 kilobits per second (kbps) for 1080p. In December 2015 its engineers described per-title encoding, which picks the ladder for each title from trial encodes. By their figures, Orange Is the New Black needed at most 4,640 kbps, "20% in bitrate savings," and BoJack Horseman could stream 1080p at 1,540 kbps where the fixed ladder had given 480p at 1,750 kbps.

    netflixtechblog.com

    Credit

    Anne Aaron, Zhi Li, Megha Manohara, Jan De Cock and David Ronca (Netflix)

  5. 2016

    A score for picture quality

    In June 2016 Netflix released VMAF (Video Multimethod Assessment Fusion) as open-source software under the Apache 2.0 licence. VMAF predicts how viewers would rate a compressed video by combining several simpler measurements with a model trained on viewers' scores, and Netflix built it into its own encoding checks.

    LessonNetflix's post warns that the score depends on the reference: a standard-definition source scoring 99 "by no means has the same perceptual quality" as a high-definition source scoring 99.

    netflixtechblog.com

    Credit

    Zhi Li, Anne Aaron, Ioannis Katsavounidis, Anush Moorthy and Megha Manohara (Netflix), with Professor C.-C. Jay Kuo's group at the University of Southern California

  6. 2018

    A royalty-free codec

    On 28 March 2018 the Alliance for Open Media (AOMedia), whose members included Amazon, Apple, Google, Microsoft, Mozilla and Netflix, released the specification for AV1. A codec is the method that compresses video and turns it back into pictures, and AOMedia offers AV1 royalty-free. AOMedia said member tests showed "an average of 30 percent greater compression over competing codecs." Netflix began streaming AV1 to Android phones in 2020, TVs in late 2021 and web browsers in 2022, and reported that 88% of the large-screen devices submitted to it for certification from 2021 to 2025 supported AV1.

    aomedia.orgnetflixtechblog.com

    Credit

    Alliance for Open Media

  7. 2020

    Less detail for lockdown

    After a call between Netflix's Reed Hastings and the European Commissioner Thierry Breton, Netflix said on 19 March 2020 that it would cut its bitrates across Europe for 30 days, estimating "around 25%" less traffic. It removed the highest-bitrate stream at each resolution and kept every resolution. BEREC, the body of Europe's telecom regulators, reported on 27 March that "no major congestion issues have occurred," and that Belgium had seen lower peak traffic after measures by Netflix, YouTube and Facebook.

    LessonA quarter of the traffic came off within days with little visible change in the picture, and the regulators' own monitoring found that networks were coping.

    lightreading.comabout.netflix.comberec.europa.eu

    Credit

    Netflix; the European Commission; BEREC

Watch for

  1. AV2, the next open codec

    AOMedia published version 1.0.0 of the AV2 specification, dated 28 May 2026. In a July 2026 post on AOMedia's site, the co-chairs of its codec working group claimed "approximately 30% better compression efficiency compared to AV1" by two objective measures, VMAF and PSNR, and "closer to 40%" when viewers judge by eye. The post does not say who ran the tests. When the tech site CNX Software tried a draft of the reference software in February 2026, it took about 31 seconds to encode each frame and its decoder failed.

    Achieved whenAn outside party publishes AV2's savings over AV1 on public test video, and a consumer service streams AV2 to devices with hardware decoders.

    av2.aomedia.orgaomedia.orgcnx-software.com

    Claimed by

    Alliance for Open Media, May and July 2026

  2. AV1 at Netflix scale

    Netflix said in December 2025 that "AV1 powers approximately 30% of all Netflix viewing." Compared with the older AVC codec, it said AV1 sessions score 4.3 VMAF points higher, use a third less bandwidth and have 45% fewer buffering interruptions. It also said it had put film grain synthesis into use in July 2025: the grain in a film is removed before encoding and the player adds it back. In Netflix's example frame this cut the bitrate from 8,274 kbps to 2,804 kbps, by 66%.

    Achieved whenAn outside measurer, such as an internet traffic analyst, confirms the codec share, or Netflix publishes the data and method behind its figures. A single frame is not an average across a catalog.

    netflixtechblog.com

    Claimed by

    Netflix, December 2025

Open

  1. 01

    Measure streaming's electricity

    A public dataset of the electricity streaming uses in real homes, measured at the wall second by second for the TV or streaming stick and for the router, while idle and while streaming, alongside the resolution and bitrate the player reports.

    The brief
    Credit · unclaimed
  2. 02

    Test AV2 in the open

    An independent, repeatable comparison of AV2 with AV1 on films that anyone may copy. It reports the bitrate saved at equal quality, the encoding time per frame and whether the decoder plays the result, with every command and score published.

    The brief
    Credit · unclaimed

Open 01 · the brief

Measure streaming's electricity

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public dataset of the electricity streaming uses in real homes, measured at the wall second by second for the TV or streaming stick and for the router, while idle and while streaming, alongside the resolution and bitrate the player reports.

Why it is still open

The best-known figures are models, and they differ. The Carbon Trust estimated about 188 watt-hours (Wh) and 55 grams of CO2-equivalent per hour of streaming in Europe in 2020, 71 Wh of it from home routers, in a 2021 paper whose research was seed-funded by Netflix through the DIMPACT project. The International Energy Agency's George Kamiya estimated 0.077 kilowatt-hours (77 Wh) and 36 grams per hour in his revised analysis of December 2020. The Carbon Trust itself wrote that there are "only a limited number of consistently measured, publicly available data points." Greening of Streaming, an industry group, now measures devices with smart plugs, but its members get "access to pre-publication data," and this research found no public raw dataset.

carbontrust.comweb.archive.orgctoic.net

Built with what exists

Plug-in power meters that log readings once a second over the home network; the playback statistics some players show, such as YouTube's "Stats for nerds"; a Raspberry Pi to collect the logs; short Python scripts to work out watt-hours per hour of viewing.

  1. First step

    Put one logging plug on a TV and one on a router at home, stream for an hour at each resolution, and compare the readings with an hour with the TV on but idle.

  2. Done looks like

    At least 20 homes logged for four weeks, published under an open licence with the scripts. It gives watt-hours per viewing hour for each device and the extra router power while streaming, compared with the Carbon Trust's 188 Wh in total and 71 Wh for routers and the IEA's 77 Wh.

Open 02 · the brief

Test AV2 in the open

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An independent, repeatable comparison of AV2 with AV1 on films that anyone may copy. It reports the bitrate saved at equal quality, the encoding time per frame and whether the decoder plays the result, with every command and score published.

Why it is still open

The 30% figure comes from AOMedia's own working group. Jan Ozer of Streaming Media wrote in September 2025 that "there are no published real-world encoding or decoding benchmarks," and CNX Software's February 2026 trial of the draft software did not get a working decode. This research found no outside test of the final 1.0.0 software.

streamingmedia.comcnx-software.com

Built with what exists

AOMedia's reference software for AV2 (AVM) and for AV1 (libaom); Netflix's open test films, such as Meridian, Sol Levante and Sparks, released under a Creative Commons Attribution 4.0 licence; VMAF for scoring; rented cloud computers, because the AV2 reference encoder is slow.

opencontent.netflix.com

  1. First step

    Encode one 10-second clip from Meridian with both reference encoders at four quality settings, and plot bitrate against VMAF for each.

  2. Done looks like

    BD-rate figures, the bitrate saved at equal quality, for AV2 against AV1 on at least ten open clips by VMAF and PSNR, with encoding and decoding times. Scripts and encoded files are published so that another person can rerun them and get the same numbers.