The environmental footprint of the digital sector is still the subject of much debate in France and Europe. Due to the lack of knowledge in environmental sciences applied to the digital sector, many things are published and repeated without being confronted with recent and verified research. For this reason, I have grouped together the common questions that run through the public debate to answer them in a short way. For each question I also propose a long answer based on the most recent verified scientific knowledge. I will try to update this page twice a year.
The digital sector is the addition of data centres, transmission networks and user equipment (both professional and consumer). Entertainment and media equipment such as TVs and consoles are also normally added. Environmental studies currently focus on the impacts of manufacturing and operations of this broad scope without being able to include everything. For example, it is very difficult to integrate the environmental impacts of end-of-life, specific factors (material footprint, etc.), some equipment (satellites, etc.) or new uses (IoT, Blockchain, etc.).
In environmental studies, the digital sector is divided into three areas: data centres, telecommunications networks and user equipment. This division is now the consensus of the scientific community and also meets the modelling objectives of international standards for studying the life cycle of a system, a service or a product. These standards are ISO 14040 and 140441 and specifically the ITU L.1410 standard2 (ETSI 203 199) developed in part by the industry (Ericsson, Nokia, Huawei, etc.).
These three clusters also represent a range of equipment. While the scope for data centres and telecommunication networks is fairly stable, the inclusion of user equipment varies from study to study. Fixed and mobile phones, smartphones, consumer premise equipment (CPE), tablets, laptops, desktops and monitors can be modelled. There is then more variability in the way computer peripherals, projectors, public displays, surveillance cameras, payment terminals, wearables and smart meters are taken into account. Some studies separate these devices from those dedicated to entertainment (E&M): televisions, video boxes, video and audio players (DVD, etc.), sound systems, headphones, cameras, fixed or portable game consoles, etc. It is therefore always important to check the perimeters of each study (this separation of equipment between ICT and E&M is linked to the way the OECD divides up sectors and is defended by Jens Malmodin).
Taking into account connected objects is particularly difficult because it is complicated to obtain a consistent count and there is also a big difference in impacts between industrial sensors and consumer connected objects (connected speaker, doorbell, etc.). Similarly, no study takes into account the impacts of satellites. Even though only 1% of global traffic passes through satellites, their impact can be significant due to the manufacturing and launch phase. There is therefore no integrated public data on the old players or on the new ones (Starlink or Oneweb).
Some recent uses are not generally taken into account directly in the overall estimates, such as blockchain applications (cryptocurrencies, smart contracts, etc.), training of “intelligent” programmes (deep learning, machine learning) or new uses of video (VR/AR, cloud gaming, etc.). Taking all this into account means accounting for the manufacture, transport and end-of-life of dedicated equipment, in addition to their impact during use.
Finally, we need to be subtle about what we call ‘digital’. It seems inappropriate to lump together GAFAMs’ systems, digital search systems, current uses, autonomous corporate networks, Wikipedia, databrokers and online advertising players, etc. There is thus a methodological difference between ‘digital’ as a designation of extremely heterogeneous infrastructures with fundamentally different development models and ‘digital’ as a designation of a process of digitisation of human activities and the current uses of a part of the world’s population. The environmental issues of the ‘digital’ are in fact mainly directed at some very large systems (GAFAMs, advertising, manufacturers, etc.) and also question the net environmental balance of the process of increasing ‘digitalisation’ of human activities - to which many other considerations should be added.
Because the digital sector must find its place in the face of the rapid and massive reduction of greenhouse gas (GHG) emissions defined by the Paris Agreement and the COPs. Because the digital sector depends on many other sectors that will have to transform rapidly and affect global material structures (energy, mining, logistics, etc.). Because the environmental crisis increases the risks to many links in the sector’s extraction, supply and manufacturing chains.
The environmental crisis includes global warming, ocean acidification, the collapse of biodiversity and many other factors. It is in response to this crisis that we are creating, with varying degrees of success, new ecological and energy transition policies. The Paris Agreement is a legal framework that responds directly to the global warming factor and sets the objective of stabilising the global average temperature at +2°C by the end of the century. Achieving this target implies a drastic and rapid reduction in emissions for the most developed countries and stabilisation for developing countries, such as India. This means that the way we structure and develop all sectors of activity on Earth will have to change radically, whatever form they take today. This is the first reason why we are interested in the environmental footprint of the digital sector: whatever its footprint today, like all other sectors, it will have to find its place in a sustainable world.
The second point is that the sector will be affected by the many transformations of our planet and our societies. Firstly, geophysical, climatic and biological transformations are at work today and for the next decades, if not hundreds of years. The digital sector has taken 30 to 40 years to create a very complex - and highly interdependent - global network of extraction, manufacturing, logistics and so on. These globalisation processes have been structured around relative climate stability in order to extract resources, process them, transport them and manufacture products. How will these global chains be affected by the intensification of extreme weather events, changing weather patterns, increased environmental risks in extraction and production areas?3
In addition to the above-mentioned transformations, one must also consider the social, cultural and political transformations that will be associated with them. For example, one of the main factors that could slow down mining is the contestation by local communities. This resistance is increasing and in response the mining sector is becoming the sector with the highest number of killings of environmental defenders.4 In short, does the environmental crisis make the damage created by mining and its dependent sectors, including the digital sector, less acceptable? High-tech pollution is far from new. Silicon Valley has been deeply polluted by the electronics industry, creating superfund sites and “cancer villages”,5 and the same is true in Taiwan and many other places on Earth.6 How long will this damage be accepted before the physical foundation of the digital sector is affected? This is only one aspect of the question, however; one could also ask how transition policies will transform the sector, or how a war between power blocs affects global chains.
Even without mentioning its environmental footprint, if the digital sector does not integrate the increasing risks and the need for its transformation, then it will only amplify its own fragility. By integrating the issue of the sector’s environmental impact then we become able to consider its own trajectory and add it to the overall transition effort launched by other sectors, again with success commensurate with the will and means committed. Moreover, understanding the environmental impact of the digital sector calls into question the development models that underpin it. For example, the rapid consumption model for smartphones (which are increasingly difficult to repair) and the annual release cycles for new products (1.36 billion smartphones delivered in 20197) are, at first sight, difficult to reconcile with ecological transition policies. It is for all these reasons that we are interested in the environmental footprint of the digital sector.
According to the work of Charlotte Freitag et al. of Lancaster University, which synthesised and adjusted the various estimates, the carbon footprint of the digital sector would represent 2.1 and 3.9% of global emissions in 2020. The uncertainty is linked to the difficulty of obtaining data from the sector’s manufacturers and the opacity of the manufacturing chains. Thus, it is more rigorous to provide a range.
Environmental footprint is a term that generally implies more than one factor. At a minimum, energy consumption, resource consumption, water consumption and greenhouse gas emissions should be included in order to draw the contours of this footprint. Today the main studies we have focus on greenhouse gas emissions and electricity consumption (sometimes including part of the manufacturing), so at best we have carbon footprint estimates.
It is sometimes said that the carbon footprint of the digital sector is 4% or 1.4% in other literature. Why such variability? Global estimates come from 3 to 4 studies: Andrae & Edler (2015)8, Belkhir & Elmeligi (2018)9, Malmodin & Lundén (2018)10, Andrae (2020)11. The 2015 study by Huawei researchers Andrae and Edler was the first to be published with open data and an open model, and as a result it has been widely picked up by the rest of the community. However, this openness was paid for by the use of publicly available and therefore rather old data. Andrae & Elder’s results are significantly overestimated in the pessimistic scenario. Andrae will correct its estimates in 2020 and it is these that should be taken into account now, not those of 2015. Belkhir & Elmeligi also used open and old data, resulting in an overestimation of the carbon footprint, especially for data centres. Finally, Malmodin & Lundén, researchers at Ericsson and Telia respectively, published an estimate based partly on primary data (from direct measurements) through a partnership with other large digital companies. However, some of this data is under confidentiality agreements and is not accessible. In addition, Malmodin splits its results between ICT and E&M, so some of these footprints need to be added together to get an estimate that is consistent with the scope of the other studies. Theoretically, the Malmodin & Lundén study would be of better quality as it uses partly recent data and direct measurements. In practice, the results cannot be replicated because some of the data is not available.
When we say that the digital sector accounts for 4% of global GHG emissions, we use the estimates from Andrae & Edler 2015. When we say that the digital sector represents 1.4% of global GHG emissions, we use the estimates of Malmodin & Lundén 2018, which exclude the whole E&M part, with this share the emissions of the sector would rather be around 2.1%.
A team of researchers from Lancaster University, led by Charlotte Freitag, has attempted to standardise the results of all these studies. The biggest task is to harmonise the scope of each study. For example, Andrae counts televisions in ICT, Malmodin separates televisions in E&M. Malmodin includes fixed phones or IoT but not Andrae 2015 and Belkhir. Then there is also a need to harmonise the reference data, especially on the manufacturing part of the equipment. The Freitag team used an internal tool (which can be criticised) to realign the supply chain data and establish that manufacturing represents 30% of the digital sector’s footprint. In conclusion, they estimate, after harmonisation, that the carbon footprint of the digital sector in 2020 would be between 2.1 and 3.9% of global emissions.12 Given the level of uncertainty, it seems methodologically more coherent to use a range rather than a fixed value.
Why so much uncertainty? Because the most recent and most realistic data are not public. They are kept confidential by equipment manufacturers and industrialists for reasons of competition, industrial secrecy or sometimes because they are not favourable. Even the primary data in Malmodin & Lundén’s study is subject to this bias in reverse: those who provide their data under confidentiality are those who are already engaged in impact reduction policies. Similarly, much of the extraction and manufacturing data does not exist or is not consistent enough to be used. The increasing complexity of digital products increases the complexity and opacity of supply chains (55 metals for a smartphone!). Apple has more than 250 subcontractors who themselves have tens or hundreds of subcontractors and so on. Any researcher who has looked at the environmental footprint of the digital sector has always come to the same conclusion: there is not (enough) data. Ironically, the sector that derives its value from data extraction does not report on its own environmental footprint.
The choice today is between imperfect (and therefore perfectible) models with open data or supposedly better models with closed data and non-replicable results. The priority issue is the opening of data, so I tend to favour open models that can be improved - while maintaining results that respect the level of uncertainty (ranges), rather than believing the results of closed models on the good faith of the authors. Scientific knowledge is built on openness, replicability and confrontation of results.
Projections beyond 10 years are not used as there are too many uncertain factors in the evolution of the digital sector and associated technologies. Andrae estimates that the footprint will continue to grow slowly (1269 MtCO2e by 2030) while Malmodin projects the footprint to halve by 2030. There are many factors that can affect the evolution of the footprint so projection is a perilous exercise. In any case, despite the increasing digitalisation over the last 20 years, the major environmental trends (increasing GHG, energy consumption, material footprint) have continued to rise.
It is agreed that low-carbon transition operational scenarios should not be projected more than ten years ahead, as there are too many uncertainties regarding, among other things, the evolution of climate conditions or the evolution of the price of low-carbon technologies. The same applies to the digital sector, where it is difficult to predict the future equipment and efficiency of the equipment, or the emergence of new uses that are not yet known. Thus, Andrae and Malmodin project their estimates up to 2030. For some unknown reason, Belkhir projects his results to 2040, but the uncertainty, both climatically and technologically, in twenty years’ time disqualifies this estimate from the outset.
This uncertainty leads to a second methodological point: it is extremely risky to project the evolution of the sector based on historical data. As mentioned earlier, digital equipment and services are rapidly becoming more efficient at the moment, so it is not possible to estimate that their past footprint will evolve consistently. This goes in two directions: the relative power consumption (kWh or J / per operation) of a piece of equipment tends to decrease over time and does not remain constant over time. On the other hand, more complex or more efficient equipment can lead to an increase in the material footprint (manufacturing impacts). There are of course all the new uses that are not yet well modelled from an environmental point of view: AI, machine learning, IoT, blockchain, VR/AR, cloud gaming (Google Stadia, etc), etc.
Andrae estimates that the digital sector will account for 1500 to 3200 Twh of electricity consumption and 1269 MtCO2e of emissions by 2030.13 Malmodin has not yet presented any new publications, but at a doctoral seminar in Leuven he estimated that the carbon footprint of the digital sector could be halved by 2030. These differences are explained by different future assumptions: Andrae estimates that efficiency gains, even at a constant 20%, will not compensate for the emergence of new uses and equipment with a short life span. Malmodin believes that the pooling of equipment and efficiency gains will more than compensate for future developments. Perhaps Malmodin can be considered as the most optimistic scenario and Andrae as a rather pessimistic one, but it is still too early to say.
Charlotte Freitag’s Lancaster team has summarised the various hypotheses in the scientific literature that may change the environmental footprint of the sector in the years to come14 which I complete here :
- Efficiency gains continue steadily ;
- Renewable energy is increasingly replacing fossil fuels in the mix of large digital companies;
- Slowing efficiency gains reduce the growth of the sector and thus the growth of its emissions ;
- Activation effects of digitalisation (enablement effect) reduce the overall emissions of other sectors ;
- Increasing mutualisation of IT infrastructure (hyperscale, etc.) ;
- Market saturation in terms of facilities and services ;
- Reducing the weight of new equipment.
Among the assumptions that consider negative trends in the digital sector :
- Exponential growth of data is increasing the energy consumption of networks and data centres (growing infrastructure) despite efficiency gains ;
- Emergence of new equipment-intensive (IoT) and data-intensive (cloud gaming, etc.) services ;
- Efficiency gains are slowing down ;
- Rebound effects counterbalance efficiency gains at the aggregate level ;
- Efficiency gains through digitisation support the development of the sector and therefore its footprint ;
- Efficiency gains from digitalisation increase the economic activity of other sectors and thus their environmental footprint ;
- Increasing the material intensity of new equipment.
This list is embryonic as there are many other factors to be noted but this will be the subject of another article. However, this first draft has the merit of showing the complexity of the future projection.
The way emissions are accounted for in the two sectors are different, the scopes of calculation are not the same and the two sectors do not serve the same purposes so there is little point in comparing the two. The comparison was used to communicate that the digital sector has material and climate consequences like a highly visible sector, aviation. Once the materiality of the digital sector has become part of the collective unconscious it will be wise to abandon the metaphor.
Long answerIt is always difficult to compare two sectors that do not provide the same services, but the increasing use of the carbon indicator sometimes encourages comparison. However, are the carbon or greenhouse gas emissions calculated on the same perimeter in these two sectors? According to the reference publications, aviation (domestic and international, commercial and cargo) accounted for 1.9% of global greenhouse gas emissions in 2016, 2.5% of carbon emissions in 2018, and 3.5% of radiative forcing in the same year.15 In absolute terms, aviation carbon emissions were just over 1 GtCO2 in 2018. It does not appear that these estimates include the impacts of the aircraft manufacturing phase. It appears that the studies ‘just’ account for fuel consumption, related greenhouse gas emissions and contrails. In terms of usage, it would appear that 11% of the world’s population travelled by air in 2018, 4% of whom travelled internationally. It would appear that frequent flyers, 1% of the world’s population, account for 50% of the sector’s emissions.16 It is also easy to define the perimeter of use of air transport as the route is segmented (going to the airport, taking the plane, leaving the airport). The carbon footprint of the digital sector includes the manufacturing and use phase in its calculation. It can also be expressed in terms of greenhouse gas equivalents but, unlike research on the aviation sector, emissions of gases such as nitrogen oxides (NOx) are much less well modelled. Similarly, the digital sector does not have data on the radiative forcing associated with its activities. In terms of usage, the digital sector has more than 4 billion users with very different uses depending on their reference digital environment and the related infrastructure/usages. Some users have much more intense uses (time spent, data volume, computing power, etc.) than others (network impermanence, paid data, etc.). In contrast to air transport, it is much more complicated to set the perimeter of digital services and digitalisation. Both sectors, however, are unique in that they have made huge gains in efficiency over the last twenty years (kWh/GB or. L/Km). It is important to note that these efficiency gains have not reduced their footprint because of the increase in their users. It is generally understood that the aviation sector is a major emitter of GHGs so the comparison with this sector is used to show that the digital sector also has a very real impact. Beyond the communication exercise, there is little to compare between aviation and the digital sector. In any case, all sectors need to reduce their emissions as quickly as possible.
Greenhouse gas emissions from the digital giants can be reduced through massive investment in renewable energy. Not all players in the sector can do the same and they will use carbon offsetting to reduce their emissions from an accunting perspective. So carbon emissions can be kept under control. It is clear that a massive shift in environmental impacts is taking place from the use phase (electricity consumption and related emissions) to the manufacturing phase (which is generally outside the responsibility of service companies). That is, the reduction of GHG emissions related to electricity consumption comes at the cost of increased consumption of materials, energy and water during the manufacture of equipment. In the short term, the material and water footprints of the digital sector will be much more to watch at both global and territorial levels.
Carbon is at the centre of most discussions on the energy and ecological transition, and the digital sector is no exception. Eventually, the carbon footprint of some digital players is likely to reduce rapidly for a number of reasons: the giants (Google, Apple, Amazon, Facebook, Microsoft) are aggressively funding entire renewable energy fleets to power their offices and infrastructure. Other large companies in the sector whose activities are mainly based on electricity consumption will also be able to massively integrate low-carbon energy. Similarly, these same organisations will massively purchase carbon offsets to reduce their incompressible emissions. In short, large digital services companies will be able to reduce their carbon footprint during the usage phase of their services (electricity consumption and related emissions), sometimes physically, sometimes through offsetting.
While this reduction in carbon emissions (linked to electricity consumption during the use phase) is possible for these players with almost unlimited capital, this does not mean that the reasoning works for other players in the sector. For equipment manufacturing companies, the transition would be much harder and would rely much more on carbon offsetting (accounting reduction rather than physical reduction of emissions). Paradoxically, the more service companies reduce their carbon footprint during the use phase, the larger the relative share of manufacturing will be. Also, we are talking about only one of many environmental factors.
After four years of experience on the subject, I have learned not to think only from a global point of view. The territorial footprint is extremely important and its study reveals many points that are invisible at the global level. For example, the monopolisation of new renewable energy capacity by the digital sector (whose electricity consumption is growing) in one territory prevents other sectors from having access to it - as in Ireland, where data centres account for 11% of the country’s electricity consumption and put the national electricity network under strain. From this point of view, the territorial analysis shows a zero-sum game: the greening of digital actors is at the expense of other actors in the territory.
The second thing I learned is that the most worrying environmental factors in the long run are probably material consumption (mineral resources) and water consumption. Even though the digital sector does not represent a significant tonnage of mineral resources today compared to other sectors, the growth of its material needs is increasing rapidly. The diversity of metals demanded and their co-dependence on major metals will pose many ecological and logistical problems.
For specialist researchers it is now clear that we are in the process of organising a massive transfer of environmental impacts from the use phase (efficiency gains in electricity consumption) to the manufacturing phase (more resource and impact intensive) which is generally invisible. It is therefore necessary to better understand the impacts that are swelling due to this transfer.
Firstly, mining is extremely polluting and has a heavy ecological footprint on local ecosystems, communities and resources. Similarly, environmental disasters related to dam failures or deliberate spills are numerous. Secondly, it is becoming increasingly possible for cyclical shortages to occur in certain minerals, affecting downstream production chains. This is due to the co-dependence between metals, delays in opening mines, lower concentration of minerals in deposits, but also to local protests against mining projects. Finally, the digital sector provides a phenomenal amount of equipment with short life spans and incentives for the general public to renew. The sector is not committed to a resource economy and disperses small and rare metals that will be lost forever.
There are several points of significant water consumption in the digital sector including: the processing and refining of ores in mines, the etching of semiconductors (especially with new lithography methods), the cooling of data centres, and, indirectly, the consumption of water for electricity generation. Mining sites can be very water intensive depending on the minerals they extract and their location. Mining sites in South America are well known for the water problems they create for local communities, by monopolising water supplies and/or polluting them. As such, the effects of climate change on the extraction areas will be potentially violent. New lithography processes (EUV) for the sub-10nm semiconductor market could also impact on water supplies in manufacturing countries such as Taiwan. Finally, the cooling of the data centres of the American digital giants is increasingly problematic in the water-stressed areas where they are located (Utah, Texas, Nevada, California). Water problems are almost invisible at the global level but are particularly salient via the territorial approach.
To understand both positive and negative effects, it is necessary to have access to good quality, representative open data. Open data allows it to be independently audited by several actors to verify its reliability. Access to good quality data depends on the ability of researchers and experts to obtain real measurement data (primary data). The representativeness of the data will depend on the scale at which one wishes to extrapolate the results. Today, very few of these conditions are met and experts use global models to extrapolate their results in space and time.
We are slowly beginning to see more clearly the environmental footprint of the sector but the methods for estimating positive impacts (avoided emissions) are far too methodologically weak to be considered. Firstly, although it is easy to model two small scenarios (videoconferencing VS face-to-face conference) it is very complicated to integrate the effects of digitalisation at larger scales where there are many more factors to integrate. Secondly, the estimates aim to incorporate enablement effects, i.e. where the digital sector can reduce emissions from other sectors, but they do not incorporate enablement effects that increase emissions (e.g. increased oil barrel production due to the digitalisation of an oil platform). As such, the few estimates present a gross balance of gains and not a net balance (emissions avoided - emissions added).
The first negative effects of the sector are its environmental footprint, which we have already discussed in the previous questions. The second negative effects would be linked to the potential of the digital sector to increase the footprint of other sectors by increasing their volume of activity faster than their efficiency, or by making them more intensive in raw materials or energy or in environmental impacts (rebound effect). In this logic, the digital sector could create new uses that are also more resource-intensive and stack their environmental impacts on top of those of similar non-digital uses. Finally, there is the impact that digitisation could have on the scale of society when we add up new activities and new uses.
The exercise is reversed for positive impacts: does the digital sector reduce the footprint of other sectors, does it create new uses that are less resource-intensive and have less impact, can digitisation reduce the environmental impact of human activities at the societal level? The model of Horner et al17 summarises the classifications of positive and negative impacts of digitalisation for energy consumption, but it can be studied for other criteria.
We are able to estimate positive or negative effects at the microscopic level: replacing a conference abroad with an online conference. As soon as new dimensions are added, the exercise becomes proportionally more difficult: does teleworking reduce the environmental footprint of a household that works in the centre of Paris and lives in a metropolitan area? What is the variation for someone living and working in Brighton? There is no general calculation model because the concrete estimation of impacts requires the integration of behavioural data which quickly limits the scaling up of results.
Estimating the positive impacts of the digital sector and digitisation is an extremely complex exercise based on many assumptions. Two estimates circulate in industry circles: 1gCO2e emitted by the digital sector means 10gCO2e avoided in other sectors18 or that digitisation could reduce GHG emissions by 20% by 2030.19 After a thorough analysis available here I have come to the conclusion that these estimates are based on so many methodological flaws that they should not be used. This is also the opinion of the scientific community20 from all sides.
The methodological difficulties in estimating the overall positive or negative impacts of digitisation are :
- A case study provided by a company/industry group is not empirical research and has a low level of confidence ;
- If a case study shows the success of a technology deployment in a given context, can this success be extrapolated? Is the same procedure applied if it is a failure? Who informs about the failure of the deployment?
- At what level can we extrapolate without increasing the uncertainty too much? City, country, area, continent?
- Is digitisation the only factor that explains the efficiency of the activity or the change in behaviour, so how do we allocate the gains/losses between several factors (yet to be identified)?
- How can a multi-factorial environmental analysis be integrated to identify impact transfers? Is there any data available?
- If we count the case studies where digitisation reduces their GHG emissions, should we also count the case studies where it increases them (e.g. increased oil barrel production) in order to get a net balance?
- How to integrate the possible changes in behaviour due to digitalisation? How to extrapolate them?
- How to check the level of confidence in the results of a study on the positive impacts of digitalisation?
Today, the documentation and studies produced by companies in the sector only show the positive effects by extrapolating specific use cases to the global level; by highlighting the activation effects that reduce emissions, not those that increase them; and by proposing calculation methodologies that would be inconceivable from a scientific perspective. While progress is being made on calculating the environmental footprint of the sector, it is safe to say that we do not know how to estimate the overall effects (both positive and negative) of digitalisation on other human activities and behaviours. The only thing we can say is that since the democratisation of the digital sector (2005, arrival of the smartphone) no environmental trend has been changed (GHG emissions, material footprint, energy consumption, etc.).
Between 196 and 400 TWh worldwide. The variability is explained by the data sets used and the assumptions about the development of hyperscalers around the world. Whatever the global consumption, the location of data centres poses real questions of planning and energy transition at local level, as in Dublin, Singapore or Frankfurt for example.
Digital industry experts like to point out that data centres account for only 1% of global electricity consumption, or 196 TWh, quoting the International Energy Agency (IEA).21 How is this figure obtained and is it the only estimate available? The IEA figure comes from an article by Eric Masanet22 in the Science web magazine, itself based on a dataset compiled by Shehabi et al (including Masanet) in 2016 from the US data centres.23 This US dataset has the specificity of including an increasing amount of hyperscale data centres (up to 92% of the number of servers in 2020). Assumptions for the average volume of active servers between 2000 and 2010 are split between 45% hyperscale (50% for 2020), 20% service providers (25% for 2020) and 10% in-house centres (15% for 2020). Global estimates are then extrapolated from these assumptions by Masanet, and in turn by the IEA.
However, are the data reported representative of other areas to justify extrapolation? In Europe, the various estimates indicate a smaller number of hyperscalers or colocation data centres. Due to a different landscape, the estimates of the leading institute in Europe, Borderstep, are higher for the EU28 zone (76.8 TWh in 2018 compared to 40 TWh in the same year for Masanet).24 Worldwide, the Borderstep Institute estimates the consumption of data centres at 400 TWh.25 It is therefore more conservative to say that the overall electricity consumption of data centres is between 196 and 400 TWh.
Finally, it is worth remembering that the issue of electricity consumption by data centres is above all a territorial issue which will be expressed differently in the countries or cities where these infrastructures are installed.For example, 80% of data centres are in Western and Northern Europe and Irish data centres represent 11% of the country’s electricity consumption. The issues at stake here are concrete and require in-depth reflection on the development and capacity of electricity networks in the face of energy transition policies.
Why does the electrical measurement of the network transmission not match its environmental accounting?
There are two ways to report the power consumption of digital networks in the environmental footprint calculation. The conventional approach is to take global indicators such as data transfer and equipment power consumption to obtain a kWh/GB ratio. This approach allows for a posteriori accounting and all power consumption is allocated equally. The power model approach consists of allocating the base consumption of the equipment to each user/subscriber and then marginally allocating the surplus electricity related to the service used. The conventional approach is suitable for corporate environmental accounting and on a large scale, the power model approach is suitable for dynamic modelling of the power consumption of a data-carrying equipment/service. Whatever the approach, the power consumption of data centres and user equipment is calculated from the time of use.
It is easy to confuse the physical measurement of a piece of equipment - such as its power consumption - with a measurement of its environmental impact. For example, you can measure the power consumption of your box, router or computer according to the different services you use. Between streaming a video in 360p and in 1080p the power consumption of these devices will vary little or not at all in proportion to the flow of data transferred. At first sight, it is therefore legitimate to ask whether data transfer can be used to calculate the environmental footprint.
It is in fact necessary to understand that the environmental footprint is an impact allocation mechanism. For example, the carbon footprint is a mechanism for allocating the emissions of a nation to each of its citizens. Each citizen will be allocated the emissions of the nation’s industry, transport or public services even if they never used them. So a citizen’s carbon footprint is not their actual lifestyle or carbon measure. Its carbon footprint is just allocated to each citizen living in the country in question.
Most environmental footprint calculations are therefore impact allocations. In the case of the digital sector, each service mobilises a whole range of equipment and infrastructure. Thus a digital service will not always be allocated an impact for what it has consumed, but will also be allocated a proportion of the impacts associated with the manufacture and use of the infrastructure simply because it has used it. There are two main methodologies: the conventional approach and the power model approach.26 The conventional approach allocates a power consumption per volume of data (kWh/GB), so it is a unit of account, to be used a posteriori when the power consumption and data flow have been measured. The power model approach is a dynamic model that is used to monitor short-term consumption.
In the case of video streaming, these two methods are used. The conventional approach is to allocate energy consumption per data volume (kWh/GB). It is a top-down allocation mechanism like the carbon footprint of a country: one takes the total power consumption of the relevant parts of the sector and allocates it to the data transmission. The power model approach is a marginal allocation method that distributes the base power consumption of the network to each user and then marginally distributes a small portion of the network’s power consumption according to the required data transmission.27 In the streaming use case these two methods are used to model the power consumption of the network, the data centres and part of the user equipment. The manufacturing phase and other environmental factors (material, water, etc.) are not integrated.
Today, the conventional approach is widely used in industry as an accounting method to integrate the power consumption of all parts of a system. However, this method is not suitable for calculating the real time consumption of a service. The power model approach gives a much closer representation of the power consumption of a service in operation and its variation on the network (change in video quality for example). On the other hand, the model is still young and there are many open questions and data to be obtained in order to use this model in a routine way.
Whether they are in use or not, digital infrastructures (data centres, networks) are always more or less switched on, so their consumption does not vary much according to data traffic. Similarly, the renewal of user equipment or the construction of new data centres is not directly related to data traffic. Thus, reducing global traffic does not directly reduce the environmental impact of the sector. However, if global traffic were to stabilise, one might assume that this would indirectly encourage a slowdown in the construction of new infrastructure. But traffic is only one factor among many that structure the development of the sector.
If we go back to what was said in the previous point, we understand that data transfer is sometimes used as an allocation mechanism for the environmental footprint of digital networks. This allocation method, while not precise and detailed, is simple and allows for some shortcuts.
If we consider that data transmission is one of the main factors explaining the development of the sector then the hypothesis holds. For example, the increase in data transfer is often used as an argument to justify the evolution of certain layers of the digital infrastructure. For example, the deployment of 5G is partly explained by the saturation of the 4G network in densely populated urban areas. On the other hand, data transfer in no way explains the development trends in the data centre market, which are mainly linked to financial investments and greater geographical coverage rather than to measured demand.
From an accounting point of view, not correlating data transmission with electricity consumption is not problematic as long as it is seen as an allocation factor of the sector’s impact, and not a measure of physical phenomena (data transfer, equipment use and manufacture).
However, if we consider data transfer as an allocation factor of the footprint then it seems logical to want to reduce this transfer. However, one should not confuse the allocation factor with what is actually being allocated. We should not imagine that reducing the global transfer of data will have a direct impact on reducing the environmental footprint (energy, GHG, water, materials). In fact, reducing data transfer would increase the amount of carbon allocated to each GB, paradoxically increasing the relative carbon footprint of data transfer. On the other hand, as an allocation factor it could be argued that reducing global data transfer would have an indirect effect: reducing the speed of development of the sector (use, manufacture and construction of data centres/networks).
The question then is whether stabilising or reducing the flow of data would have an effect on the development of the sector. Telecommunication networks evolve every 10 to 15 years to accommodate higher throughput and to create new uses. 5G is partly creating new uses (IoT, industry, etc.) which will bring new data transmission vectors and new business opportunities for operators (they hope so anyway). As mentioned earlier, the rapid construction of new data centres around the world is not due to the increase in data traffic. Rather, the logic of financial investment and concentration of activities seems to govern this part of the sector. User equipment is subject to renewal cycles that are not correlated with data transmission. Also, data traffic does not also give a view of computing. For all these reasons, it seems a priori inaccurate to imagine that the reduction of global traffic would reduce the environmental footprint of the digital sector, even if increasing data traffic is part of the commercial arguments for developing part of the sector.
If nothing else, data traffic may be relevant as a mechanism for allocating the power consumption of networks. However, it should not be confused with a mechanism for reducing the environmental footprint of the digital sector. It is an overall factor, not the key.
From an environmental point of view, deleting a user’s emails has almost no impact. Deleting emails is more of an act of mental wellbeing.
n France, the idea of eliminating emails to reduce one’s footprint seems to come from an old communication from ADEME. At the time, the agency’s only available life cycle assessment in the digital sector was on email. Later, this ADEME communication campaign was taken up by companies such as Orange who popularised this practice.
In this case, deleting emails has almost no direct impact apart from freeing up a tiny bit of space on some servers and avoiding the resulting data transfer. However, email storage space is limited depending on the offer, ranging from 5 to 20 GB for the general public and more for professionals. Also, 85% of the emails sent each day would be spam28 and are therefore not controllable through a change in individual behaviour. Finally, emails represent a tiny part of the global data traffic.
You can delete your emails for your mental well-being, privacy or many other reasons, but deleting emails to reduce your environmental footprint has almost no effect.
The discussion was launched following the Shift Project’s report on the issue. Many major media outlets subsequently reported the estimates. However, the think tank made several calculation errors, pointed out by George Kamiya of the IEA, leading to an overestimation of the carbon impact of online video. Today, the white paper co-authored by the Carbon Trust and numerous researchers and experts on the subject is a reference. Streaming on Netflix is estimated at 100gCO2e per hour, video streaming in Europe is estimated at 56gCO2e per hour. These estimates only take into account the usage phase of the digital infrastructure and exclude the impacts of the manufacturing phase. Within this limited scope, the main source of electricity consumption is the television set (smart or not). The electricity mix of the country where the consumption is located varies greatly in terms of CO2e emissions, and there is little variability in the footprint depending on the resolution (this is linked in particular to the allocation method).
Video accounts for 80% of global data traffic according to Cisco.29 The issue of the carbon footprint of video streaming has been widely highlighted by the Shift Project’s report dedicated to it. In 2019, the think-tank estimated that online video represented 300 MtCO2e per year: 102 Mt for VoD (Netflix, etc.), 82 Mt for pornography, 65 Mt for ‘Tubes’ and 56 Mt for other types of video.30 This figure was picked up by the general press and opened a debate on the place of streaming in the transition effort.
This estimate was obtained by multiplying the traffic dedicated to video streaming (at 3Mbps) by a ratio of kWh/GB for the network and data centre part, and then by a ratio of kWh/min for the user equipment (computers, smartphones, televisions). The resulting electricity consumption was then multiplied by the global average carbon intensity (0.519gCO2e/kWh). However, the Shift Project made several errors in its estimate. The first is the choice of reference publication for the calculation of the footprint. The Shift Project relies on Andrae and Edler’s 2015 study, which is known to overestimate the energy consumption of networks and somewhat that of data centres. In addition, they made a conversion error between bit (b) and byte (B) (1 byte = 8 bits) which resulted in an 8-fold increase in impact in some parts of their calculation. Finally, the Shift Project would have underestimated the electricity consumption of user equipment by not including televisions.
These points were highlighted in a publication by George Kamiya,31 an IEA consultant, in Carbon Brief, a leading UK website on carbon transition issues. George Kamiya’s “fact check” uses Netflix’s figures to make a global point. Clearly, the Shift estimates are far too high, but is Netflix’s infrastructure sufficiently representative to make a general statement about video streaming? Netflix uses Amazon Web Services (AWS) for the bulk of its operations and a Content Delivery Network (CDN) system called Open Connect 2 to get the video files as close to the consumption areas as possible. Netflix has developed one of the most efficient infrastructures in the world for video streaming. The company has forged consumption habits specific to the platform that are now generalized to other platforms. In sum, Netflix represents a very optimistic use case and is not necessarily representative of the entire video streaming industry. Based on the case of Netflix, George Kamiya estimates that one hour of streaming would represent 36gCO2e / hour, compared to the Shift Project’s original figure of 3.6kgCO2e / hour in 2019, and 394gCO2e / hour in the think tank’s adjusted version in 2020.32
What does the main stakeholder say? In 2020, Netflix commissioned researchers from the University of Bristol, who specialise in the carbon footprint of video streaming and their DIMPACT model33, to estimate its footprint. Using primary data from Netflix, the team estimated that one hour of streaming on the company’s global infrastructure emitted 100gCO2e,34 3 times more than George Kamiya’s estimate and 4 times less than the Shift Project’s corrected estimate. All these estimates are for the use phase only and do not include the impacts of equipment manufacture.
Finally, the Carbon Trust, with the help of many of the digital sector’s power consumption experts, including the DIMPACT team, has published a white paper entitled “Carbon impact of video streaming”.35 This publication is now a reference on the issue and estimates that the carbon footprint of one hour of streaming in Europe is 56gCO2e / hour. This white paper recalls several key points: the main source of electricity consumption is the TV set (smart or not), the electricity mix of the country where the consumption is located is important, and there is little variability in the footprint depending on the resolution (this is notably linked to the allocation method).
All the necessary national and European policies must be put in place to extend the life of equipment, reduce the purchase of digital equipment over time and stabilise the development of digital infrastructure (data centres, networks) in order to benefit from efficiency gains.
When we know that a large part of the environmental footprint of digital technology is in the manufacture of equipment, particularly consumer equipment (smartphones, computers, screens, televisions, etc.), it seems logical to reduce the purchase of digital equipment.
At the collective level, how can we consume less equipment and increase its lifespan? Firstly, we need to extend the lifespan of equipment by extending manufacturers’ warranties and imposing reparability of equipment. We also need to regulate the duration of operating system (OS) updates to encourage people to keep their equipment longer without creating more security breaches. The repair professions must be encouraged through specific training programmes, the opening of repair guides and access to spare parts at no extra cost.
Once we have ensured that the lifespan of the equipment and the continuity of the software layer are really extended, we will have to reduce the flow of imports of this type of equipment into Europe, both for ecological reasons and because the market is saturated. This can be done through a targeted carbon tax on new consumer digital equipment.
During this century it is increasingly possible that Europe will experience shortages of certain minerals. So extending the life of our digital equipment, consuming less of it, strengthening our maintenance and circularity capabilities are one of the keys to reducing the environmental footprint of the digital sector and to be part of a long-term EU digital strategy.
At the individual level the most important actions are simply to keep your equipment for as long as possible, not to buy new if possible and to avoid over-equipment per person or per household. In an ideal world we should aim for consumer equipment (smartphones, etc.) with a functional life of 10 years.
Extending the functional life of equipment is a priority for users of the digital world in order to maximise the amortisation of the resources and pollution involved in its manufacture. It is also necessary to encourage the purchase of reconditioned or second-hand equipment to reduce the use of virgin materials or new products. Finally, the marketing of numerous connected objects for the general public is a siren call that must be resisted in order to avoid digital over-equipment (in addition to avoiding problems of security, connectivity, etc.).
Keeping a smartphone for as long as possible will depend on the type of phone already owned, the continuity of its update, the power and memory required by the applications requested and the obligation to use certain applications (such as for two-factor authentication via app). It will also depend on access to a network of repairers or the ability to repair one’s own equipment and therefore access to guides/tutorials if repairable and to spare parts. The repairability of digital equipment is something to fight for when it goes against the business models of manufacturers and the chain of actors they are linked to (operators, etc.). Similarly, you can choose not to own a phone and rent it through cooperatives like Commown.
The issue of over-equipment will be crucial in the years to come with the commercial development of connected objects. Thousands of new products responding to artificial needs or desires appear every day. Consumer connected objects generally have a short life span due to their short-term design, the short duration of software maintenance, or the possible superficiality of the need.
“ISO 14040:2006 – Environmental management — Life cycle assessment — Principles and framework,” ISO, 2006 ; “ISO 14044:2006 – Environmental management — Life cycle assessment — Requirements and guidelines,” ISO, 2006↩
Recommendation ITU-T L.1470: Greenhouse gas emissions trajectories for the information and communication technology sector compatible with the UNFCCC Paris Agreement,” ITU, 2020.↩
Shahana Althaf et Callie W.Babbitt, “Disruption risks to material supply chains in the electronics sector”, Resources, Conservation and Recycling 167, 2021.↩
Arnim Scheidel, “Environmental conflicts and defenders: A global overview”, Global Environmental Change 63, 2020.↩
Christophe Lécuyer, “From Clean Rooms to Dirty Water: Labor, Semiconductor Firms, and the Struggle over Pollution and Workplace Hazards in Silicon Valley”, Information & Culture: A Journal of History 52, n° 3, 2017, pp. 304-333.↩
Hua-Mei Chiu, “The Dark Side of Silicon Island: High-Tech Pollution and the Environmental Movement in Taiwan”, Capitalism Nature Socialism 22, n° 1, 2011, pp. 40-57.↩
Canalys, “Global smartphone market Q4 and full year 2019”, Canalys Newroom, 30 janvier 2020.↩
Anders S. G. Andrae and Tomas Edler, « On Global Electricity Usage of Communication Technology: Trends to 2030 », Challenges 6, 2020, pp. 117-157. ↩
Lofti Belkhir and Ahmed Elmeligi, « Assessing ICT global emissions footprint: Trends to 2040 & recommendations », Journal of Cleaner Production 177, 2018, pp. 448-463.↩
Jens Malmodin and Dag Lundén, « The Energy and Carbon Footprint of the Global ICT and E&M Sectors 2010-2015”, Sustainability 10, n° 9, 2018, p. 3027.↩
Anders S. G. Andrae, « New perspectives on internet electricity use in 2030 », Eng. Appl. Sci. Lett. 3, 2020, pp. 19-31.↩
Charlotte Freitag et al, “The real climate and transformative impact of ICT: A critique of estimates, trends, and regulations,” Patterns 2, 2021.↩
Anders S. G. Andrae, « New perspectives on internet electricity use in 2030 », Eng. Appl. Sci. Lett. 3, 2020, pp. 19-31.↩
Charlotte Freitag et al, “The real climate and transformative impact of ICT: A critique of estimates, trends, and regulations,” Patterns 2, 2021.↩
D.S. Lee, “The contribution of global aviation to anthropogenic climate forcing for 2000 to 2018”, Atmospheric Environment 244, n° 1, 2021.↩
Stefan Gossling, Andreas Humpe, “The global scale, distribution and growth of aviation: Implications for climate change”, Global Environmental Change 65, 2020.↩
Nathaniel C. Horner et al, “Known unknowns: indirect energy effects of information and communication technology,” Environ. Res. Lett. 11, 2016.↩
Carbon Trust, « The Enablement Effect – The impact of mobile communications technologies on carbon emission reductions », GSMA, 2019.↩
GeSI et Accenture Strategy, « SMARTer2030 – ICT Solutions for 21st Century Challenges », GeSI, 2015.↩
Jens Malmodin et Vlad Coroama, « Assessing ICT’s enabling effect through case study extrapolation – the example of smart metering », 2016 Electronics Goes Green 2016+ (EGG), 2016 ; Jens Malmodin et al., « Considerations for macro-level studies of ICT as enablement potential », ICT for Sustainability 2014 (ICT4S-14) 179, n° 188, 2014 ; Vlad C. Coroamă et Friedemann Mattern, “Digital Rebound – Why Digitalization Will Not Redeem Us Our Environmental Sins”, ICT4S (2019).↩
George Kamiya, “Data Centres and Data Transmission Networks”, IEA, novembre 2021.↩
Eric Masanet et al., “Recalibrating global data center energy-use estimates”, Science Mag 367, n° 6481, pp. 984-986.↩
Arman Shehabi et al.,“United States data center energy usage report”, Lawrence Berkeley National Laboratory, 2016.↩
Francesca Montevecchi et al., “Energy-efficient Cloud Computing Technologies and Policies for an Eco-friendly Cloud Market”, European Commission, 2020.↩
Ralph Hintemann, Data centers 2018. Efficiency gains are not enough: Data center energy consumption continues to rise significantly - Cloud computing boosts growth”, Borderstep Institute, 2018.↩
Carbon Trust, “Carbon Impact of Video Streaming”, Carbon Trust, 2021.↩
Jens Malmodin, “The power consumption of mobile and fixed network data services - The case of streaming video and downloading large files”, Electricity Goes Green, septembre 2020.↩
Nikolina Cvetićanin, “What’s On the Other Side of Your Inbox – 20 SPAM Statistics for 2021”, DataProt, 11 février 2021.↩
Cisco, “Cisco Visual Networking Index: Forecast and Trends, 2017–2022”, Cisco, 2018.↩
The Shift Project, “Climat : l’insoutenable usage de la vidéo en ligne”, The Shift Project, 2019.↩
George Kamiya, “Factcheck: What is the carbon footprint of streaming video on Netflix?”, Carbon Brief, 25 février 2020.↩
Netflix, “Environmental Social Governance 2020”, Netflix, 2021.↩
Carbon Trust, “Carbon impact of video streaming”, Carbon Trust, 2021.↩