What do we know about the environmental impacts of online video? The example of Netflix

This article is published in collaboration with Binaire, the blog to understand digital issues.

In view of the significant efforts that the Information and Communication Technologies (ICT) sector must make to move away from a trend of strong growth in its Greenhouse Gas emissions (+45% by 2030) to a significant reduction trajectory (-45% over the same period), all aspects of digital must necessarily question their contribution to this trend. In this sense it is interesting to map and estimate their environmental impacts in order to build a sustainable society.

As the chart below shows, online video accounts for the majority of internet traffic globally.

For France, video hosts are mainly responsible for increasing traffic to users at the interconnection level, with an increase of x2.4 from the start of 2020 to the end of 2022.

Video is therefore often singled out as one of the main causes of the environmental impacts of digital technology and is debated, for example, in The Shift Project’s 2019 report on the unsustainability of online video. But beyond TVs and other viewing terminals, what do we know about the materiality behind watching a film or series on video on demand (VoD)? What is this equipment actually used for?

Little public information exists on the very complex architectures operated by VoD service providers and on their sizing. However, we attempted a rough analysis of the Netflix service. The latter is in fact one of the most used and has the merit of making some information accessible regarding its operation. This platform is only an example and the aim is in no way to target this company specifically.

Through the information gathered via different sources (activity reports, videos of technical conferences, blog articles), we illustrate the difficulty of mapping the different parts of the architecture of a VoD service, work which is nonetheless essential before carrying out a quantification of the environmental impacts of such a service.

Although electricity consumption does not encompass all environmental impacts, it is one of the pieces of information made public by Netflix, as shown in the figure below. 2019 is the only year for which an estimate of the consumption of servers operated by third parties is provided.

Video streaming, how does it work?

The easiest part to explain about this consumption concerns the content delivery network (CDN). This is an infrastructure with several tens of thousands of servers for Netflix spread around the world and hosting the most popular titles in the catalog close to users. Some of these servers are hosted directly in the data centers of Internet service providers (ISPs) and thus escape the precise power consumption reported by Netflix.

ISPs in fact have their own data centers, in order to allow quality Internet traffic to be delivered while reducing pressure on network infrastructures.

So when a user accesses online video content, it is in reality most often on a CDN that this content is hosted. This data goes through all the network infrastructures (cables, antennas, routing equipment, etc.) necessary to connect this CDN to the user, without forgetting their Internet box and potentially other equipment (switch, wifi repeater, TV box , etc.) to the terminal where the video is viewed.

The rest of the consumption of third-party servers concerns the use of Amazon Web Services (AWS) for all operations upstream of the distribution of video content itself. This represents constant use of several hundred thousand virtual machines (more than one hundred thousand in 2016) in addition to storage requirements.

Indeed, once produced, a video is first stored in raw, uncompressed form. It is then encoded in around a hundred different versions in order to offer the best quality of user experience whatever the characteristics of the terminal and its screen (more than 1,500 types of equipment supported), the quality of the Internet network and of the operating system used. Video encoding thus requires several hundred thousand CPUs in parallel.

These multiple versions are duplicated on several servers within data centers located in different geographic areas (3 for Netflix in 2016) for security reasons and to guarantee access to content throughout the world.

The hidden face of VoD

The business models of VoD platforms are based on attention retention and the number of subscribers or views. Before being able to view video content, the user begins by navigating the platform, from the home page to choosing content. This navigation is personalized for each user, and is based on attention capture and retention mechanisms requiring the collection, storage and processing of numerous personal and usage data. A whole part of the infrastructure, and therefore of the environmental footprint, is thus linked not to the distribution of videos, but to the personalization of the user experience.

The set of data allowing the implementation of these mechanisms is commonly called “datahub”. This is made up of both data collected by the platform (user data and usage data), aggregated with data from other sources in the Netflix value chain: advertisers, payment providers, service providers , audience measurement providers, content reviews, social networks, etc. Netflix is, for example, a member of the Digital Advertising Alliance.

This datahub is of significant size, in 2016 for 89 million accounts it contained 60 PB (1 petabyte = 1 million GB) of data. It is not aberrant to imagine that it will be even more important today with 260 million subscribers in 2023.

On the other hand, the size of Netflix’s catalog is estimated to be between 50,000 and 60,000 hours of viewing. In this article, we are talking about 470 GB per hour of raw video, which gives an unencoded catalog of around 25 PB, or half of the 2016 datahub. Intuitively, however, we would expect that the catalog raw video accounts for a greater share of storage requirements than usage data.

In addition to this volume of data, there are video streams and data produced in connection with filming and editing by the Netflix studios themselves, which represent around 100 PB per year.

To support their economic model, the volume of original productions is growing strongly, as are the associated environmental impacts, which represent more than half of Netflix’s greenhouse gas emissions.

Beyond storage, each action on the platform (search, click on play, etc.) generates an event processed by Netflix, there were 500 billion per day in 2016. This data capture is used, for example, to generate ‘a personalized home page for each user account. The latter would require a total of more than 22,000 virtual servers hosted at AWS, and the storage of more than 14.3 PB of data for the management of a dynamic cache called EVCache.

Among the elements of personalization of the user experience, there is obviously the content offered, but also the way in which it is presented with personalization of the thumbnails used, or even the use of “Dynamic Sizzles”, generation of personalized videos aggregating content from several films or series.

These attention retention mechanisms rely on the use of increasingly advanced continuous machine learning algorithms, requiring both large amounts of data and computing power. For obvious reasons of updating, these algorithms are trained incrementally. The generalization of their use necessarily causes an increase in the data acquired, processed and stored, thus increasing the associated environmental impacts.

To all this data, we must add the backup policies necessary for business recovery in the event of an incident. They are applied at each level of this architecture, which can lead to more or less significant duplication of all of this content. Note that Netflix also implements sophisticated data purification methodologies, both in terms of the cache and the datahub and the data produced by the studios.

“Just” a video?

So, watching a video online involves a lot of steps and data generated well beyond the video content itself. Optimizing the user experience to the extreme relies on significant hardware resources compared to just watching video. The lack of available information on the complete functioning of the platforms and the associated infrastructures makes it hazardous at this stage to assess the environmental impacts of their activity with regard to planetary limits by an independent third party.

Some might retort that these impacts when compared to the number of subscribers would undoubtedly be negligible compared to many other consumption items. For example, for 2019, the power consumption of the servers used by Netflix only represents around 2.3 kWh/year per subscriber. This figure may seem insignificant, or even contradictory with the figures for the energy consumption of data centers in the world from the International Energy Agency, which would represent approximately the electricity consumption of a country like Italy or the United Kingdom. . This illustrates a difficulty with Information and Communication Technologies (ICT) which is a sector made up of a myriad of services, each seemingly insignificant, but whose sum of impacts is worrying. In other words, the overall reduction of the environmental impacts of ICT necessarily involves a myriad of “small wins”. In the context of video streaming, Netflix is ​​just one VoD service among a whole set of providers, and new modes of peer-to-peer video sharing or via social networks.

Can video on demand become sustainable?

In order to comply with a trajectory of reducing the environmental impacts of the online video sector, we can legitimately wonder what a VoD platform compatible with a sustainable environmental trajectory would look like. The preceding analysis raises at least four main areas of reduction:

• the compromise to be made between the weight (maximum resolution and number of variants) of the encoded videos (which impacts the calculation, storage and transmission requirements) and the quality actually perceived by users,

• questioning the personalization of the experience to the extreme,

• the need for high performance (in terms of quality of service, availability, etc.) for an entertainment service,

• and finally, the pace of production of new content.

Beyond environmental issues, the recommendation algorithms at the heart of these VoD platforms also raise many ethical and democratic questions.