An Attempt to Sort Out Digital Carbon Footprint Evaluations

  

1. Introduction

Ten years ago, together with Erol Gelenbe, we wrote a NATF report on the impact of ICT (Information & Communication Technology) on worldwide electricity consumption. You may also read the associated ACM article. This report was the result of a collective study on ICT electricity consumption, prompted by a growing concern that ICT CO2 impact was growing exponentially, especially from 2006 to 2010. The report showed that there was indeed growth, but no strong accelerations, and that many crazy forecasts that were just … forecasts. I decided six months ago to refresh this analysis and to address the larger issue of the impact of digital on CO2 emissions. My reasons for returning to this question were twofold.  First, as was the case 10 years ago, there is a rising cycle of concern, with lots of exaggerations and scary forecasts about what will happen in 2030. Second, at we are now much more knowledgeable about the lifecycle analysis of servers and digital devices, so we can address the “scope 3” questions more thoroughly than we could in 2010.

The goal of this blogpost is just to share a few numbers which I have painfully collected and sorted out. They are meant to be used as “orders of magnitude”, since the level of uncertainty is still quite high, but they may come handy to the reader when trying to assess the situation. To update this 10 years-old study, I will apply the following methodology. First, I will look at 2015 because we have plenty of data and many published studies, so I can make a synthesis with a reasonable level of confidence. I will look at the global numbers but also how they were produced (resource units x carbon unit costs). Then I will retrofit to 2010 to see how it fits with previous studies of 2010, including the NATF document. Using the resource units (number of servers, laptops, smartphones, TVs …) and unit costs model, I can extrapolate to 2020 and propose a fact-checking matrix to compare against what is found today on the Web. This is the main contribution of the post, reflected in the matrix that may be found in the conclusion. I will also offer my own prospective for 2030, which is not that difficult once you have the structure of the ICT footprint impact, but which is subjective by nature.

The main findings, compared to the previous study are a mix of confirmation and new information. On the server side, the progress made with individual consumption and PUE means that the total worldwide consumption of data centers has stayed globally constant over the past 20 years, despite a vigorous growth in the numbers of servers. However a new class of server-side devices, namely the blockchain mining infrastructure has made a stupendous entrance with a consumption that is growing fast and will soon be similar to the rest of the data centers. On the device side, the continuous growth of 2000-2015 has stalled, so the prospect for the future is rather good, while network usage will continue to grow, exponentially fast in term of traffic, but steadily and moderately in term of CO2 impact. When we add all CO2 impacts of our digital activity, we find today a share slightly below 3% of our total CO2 emissions, which represents approximately 2% of our total greenhouse gas emission (58Gt today).

This post is organized as follows. Section 2 focuses on “scope 2”, that is the electricity consumption due to digital activities. I will adopt the distinction between ICT (Information & Communication Technology) and E&M (Entertainment & Media) proposed by Jens Malmodin and Dag Lunden in their per “The Energy and Carbon Footprint of the Global ICT and E&M Sectors 2010–2015”. This paper serves as the backbone for this post, as the EINS report FP7-2888021 “Overview of ICT Consumption (D8.1)” was, 10 years ago, the backbone of the NAFT report. My goal is to reconstruct and share a few useful “key figures” to understand the number of devices and the units of electricity consumption. Once I have analyzed/reconstructed the 2015 numbers from their paper, I will reconcile them with other sources regarding the 2010 values (which is easy). I will then project them to extrapolate the 2020 values, which is harder and must include newcomers such as bitcoin mining. Section 3 deals with LCA (Lifecycle analysis) and the “scope 3” impact, with a specific focus on the CO2 footprint of manufacturing. I separate scope 2 and scope 3 because scope 3 analyses are both more difficult and more recent. The early studies of 10 years ago, including mine, were rather naïve. In this section, I start with the same source for 2015 and I propose some adjustment based on what we know today about manufacturing estimates. I will apply the same logic of reconciling with 2010 numbers and extrapolating to 2020 numbers. Section 4 applies this analysis, in a prospective manner, to 2030. On one hand, it is speculative and thus offered as food for thought. On the other hand, the literature is full of scary predictions about the exponential growth of ICT impact, so it is useful to understand what the drivers are and to make your own opinion about what realistic growth may be. As far as I am concerned, most of the forecasts I have seen in the past year are off by a factor of two.

Because the topic of ICT CO2 impact is very sensitive, and because I am not in a position of neutrality since I have been an ICT professional all my life, I need to point out that I am not an expert on this topic. I am leveraging several published research articles and applying the type of modelling and analysis that I have been doing for a long time (to be specific, what gives me credibility is not what I know, but the large number of errors I have made in the past 20 years performing similar analyses). I have been interested in the topic of CO2 impact for a really long time thanks to my father Paul CASEAU (since 1978 to be precise), former hear of EDF R&D, and because I had access to qualified experts when I was part of EDF Scientific council 10 years ago. Before that, I had been in charge of sustainable development for a few years at Bouygues Telecom, which gave me access to the great network built by Fabrice Bonnifet. I have a further motive in collecting data about ICT CO2 impact, which will be the topic of a future post, but which may be found in my research agenda. As always, the content of this blog is purely personal and does not reflect in any way the opinions of my past and present employers.

2. Digital Energy Consumption Update

One of the difficulties when discussing “the impact of digital life on the planet” is to be clear about the scope of digital and what kind of digital we are talking about. In this post I follow the approach proposed by Jens Malmodin and Dag Lunden:

-           ICT is mostly made of servers (data centers), networks (fixed & mobile), user devices (there is a long list, see Malmodin’s paper for detail, the most important ones are smartphones, laptops, desktops, telco interner/routers boxes) … to which I have added bitcoin mining (a new kind of data centers that was not there in 2010 or 2015)

-          E&M (Entertainment and Media) is made of TVs, STB (set-top boxes), home audio systems and many other small categories. Today I will simply focus on TV + others, since TV sets are a topic of interest of their own, are the bigger category of E&M, and it is critical to know whether TVs are included when you read a figure about “digital impact”.

I will start with what I have collected about ICT electricity consumption, which was the core of the NATF study that I quoted in the introduction.

 

Server consumption has remained relatively flat with worldwide value which is slightly above 200 TWh/year. The number of servers shipped worldwide has grown from 8.9M in 2010 to 11.09 in 2015 to 12.15M in 2020, according to Statista. Consequently, the number of servers has grown, from approximately 40M in 2010 to 55M in 2020. The servers have also grown more powerful to accommodate a strong growth in the workload, but their energy efficiency has significantly increased over the past 20 years. There are many papers to read on this topic, but I strongly suggest “Beyond the Energy Techlash: The Real Climate Impacts of Information Technology” by Colin Cunliff. The paper spends some time debunking some of the myths and false claims associated to ICT consumption. It is very similar to the NATF paper, but being written in 2020, it is more relevant. The key message in both documents is that the progresses made on (1) PUE (2) Typical server energy intensity (3) average number of servers per workload (thanks to the cloud) and (4) the average storage drive energy use (kilowatt-hour/terabyte) have made significant progress from 2010 to 2018, which compensate the growth in servers and workloads. You can find another analysis in the article “Recalibrating global data center energy-use estimates” by Eric Masanet, Armand Shehabi, Nuoa Lei, Sarah Smith and Jonathan Koomey (sciencemag.org), which quote a total consumption of 205 TWh in 2018. The values that I found for 2010 and 2015 vary in the 200-240 TWh range (for 2010, Malodin reports 240 and ITU reports 205). It is important, when comparing with other published articles, such as “Assessing ICT global emission footprint : Trends to 2040 & recommendations” by Lofti Belkhir and Ahmed Elmeligi, to check if the datacenter consumption figures are collected or forecasted (easy to track when the source is anterior to the date of the reported value !). There has been a large stream of papers forecasting huge growth of datacenter electricity consumption, and this dramatic growth had not occurred.

Network consumption is growing at a constant rate. While the consumption of network was less than datacenters in 2010 (185TWh for Malmodin), it grew to 242TWh in 2015. In addition to the Malmodin paper, a great source on this topic is the ITU document (L.1470) : “Greenhouse gas emissions trajectories for the information and communication technology sector compatible with the UNFCCC Paris Agreement”. This report gives similar numbers (230TWh in 2015), to be compared with 220 TWh in the IEA report. It is harder to extrapolate the 2020 numbers, but it seems reasonable to expect a 20% growth, so I selected 280 TWh as my estimate in the table below.

On the other hand, the total consumption of user devices has probably peaked in 2015. The history of ICT device electricity consumption shows two stages: significant growth the previous decade, up to 342TWh in 2015, followed by stabilization because we see a clear stabilization in the number of devices. The number of smartphones sold each year went from 300M in 2010 to 1420M in 2015 and then 1433 in 2021 (2020 was a special year, see the curve on Statista). For laptops, the shipment figures are 201M in 2010, 163M in 2015 and 222M in 2020 (COVID rebound). For desktops, we have 157, 113 and 80M.  Malmodin & Lunden’s paper works with the estimate of 3700 M smartphones, 970M laptops and 370 desktop PC in operations in 2015. If you make the division you will see that he assumes 34 kWh as the unit (yearly) electricity consumption, which is less what most “back of the envelope study” (365d x 8h x 60W -> more than 150kWh) assume, but more than what field studies have found (20kWh).  The 2015 number matches that of the ITU report (345 TWh). Because the number of device has stabilized and it looks like the consumption of user devices such as a laptop is fairly stable, I have used 2015 figure as my estimate for 2020. Notice that the 342 TWh number includes 45TWh for phones, 34TWh for laptops, 109 for PC and 30 TWh for displays.

The arrival of Blockchain and bitcoin mining makes for a totally different picture, that of true exponential growth with a worldwide consumption of approximately 100TWh in 2022, that is the half of all other servers worldwide.  On this topic the best source that I found is the IEA report “Data Centres and Data Transmission Networks” (quite recent : September 2022). The figures are pretty consistent what what is reported by Statista on Bitcoin Energy Consumption (from which I drew the 70 TWh figure reported in the table below).

The most significant contributor of the E&M segment is the consumption of TV sets. There are conflicting trends at work : on the one hand, to move from LCD to LED has reduced the unit consumption of a pixel; on the other hand the growth of the TV sets in size and number of pixels (from HD to fullHD to 4K and, maybe, to 8K). However, according to Malmodin’s paper, our previous CRT and LCD TVs were such power hogs that the newer larger LED sets result in a continuous improvement from 200kWh in 2010 to 140 kWh in 2020. The estimate total of 1900M of TV sets worldwide in 2015 generated 160TWh of electricity consumption.  Together with STB, home theaters and other devices, the electricity consumption of the E&M sector in 2015 was 467kWh.

 

3. LCA (Life Cycle Analysis) and Carbon Footprint for Manufacturing

Let us start with data centers. The carbon footprint reported by Malmodin and Lunden in 2015 is 160 Mt of CO2, with a breakdown of 135 Mt for operations, 25Mt for scope 3, mostly manufacturing. Operations being mostly electricity consumption, it means that Malmodin uses a rather large CO2 intensity for the worldwide average, namely 560 g/kWh. We will return to that question later on. If we divide the manufacturing number by the number of servers, we get approximately 500kg/year per server for scope 3, which is consistent with both a number of spec sheets from Dell or HP, actually on the high side. We notice that scope3 is roughly 16% of the total footprint, a figure that we find in many documents from server manufacturer, but which often suppose a high CO2 intensity for operations. With more realistic CO2 intensity, the previously mentioned Dell server represents 320kg/year for manufacturing and 1760 kWh/year of electricity consumption, which can be evaluated as 680 kg/year for a better mix of 380g/kWh, leading a 30% scope3 contribution to the footprint. This discussion may be found in “The carbon footprint of servers”.  To better understand the carbon footprint a of server, I recommend to read the longer Boavizta article. To retrofit to 2010 and to extrapolate to 2020, we need to understand the variation of the manufacturing footprint, which is hard because the topic is newer than electricity consumption for server. In the following table, I have assumed constant carbon unit costs (cost per server). The result is a slow progression between 2010 and 2020 that reflects the growth with the number of servers.

As far as devices are concerned, Scope 3 (manufacturing) is the major driver of the carbon footprint. For 2015, Malmodin & Lunden report a total scope 3 footprint of 196 Mt, including 64 Mt for smartphones, 32 Mt for laptops and 28 Mt for desktops. The associated CUC (carbon unit cost) is 200 kg (163 M new units shipped producing 32 Mt), which is lower than the typical 300 kg that we find in other more recent studies. The typical CUC numbers for a modern laptop are 300kg for manufacturing and 100kg for usage (for instance, from the circular computing article  “what is the carbon footprint of a laptop”). As before (electricity consumption), it is hard to get a common number from different sources (for instance, look at the Oxford IT services paper). When looking at the PCF (Product Carbon Footprint) of the Lenovo T490, the first figure one sees is the 615 kg, lifetime footprint until you read later that, because of the large variation, Lenovo reports the 95 percentile confidence number (safe by overestimation) rather than the average of 421 kg which is written below (+/- 108). When trying to adjust to 2010 and 2020 values, it looks like the unit costs have grown from 2010 to 2015 as laptops became more sophisticated, but that the manufacturing unit costs has stabilized since then. These are the hypotheses that I took more generally to extrapolate the 2015 total device footprint to 2010 and 2020.

To compute the carbon footprint of ICT, we need the evaluation of Networks. For 2015, Malmodin reports a total footprint of 169 Mt, that has grown from 144Mt in 2010. It is hard to find much information about the scope3 / scope 2 structure of the network footprint (neither in the document or in the references such as “The electricity consumption and operational carbon emissions of ICT network operators” but the report states that total scope 3 for network and data center is approximately 50Mt, which I have used here (25 Mt for both). Altogether, we get a carbon footprint of 730 Mt for ICT, and a retrofit that gives 700Mt in 2010. The 2010 number is very consistent with what was found in the NATF report, and also what is reported in the study “The climate impact of ICT: A review of estimates, trends and regulations” by Charlotte Freitag, Mike Bernes-Lee, Kelly Widdicks, Bran Knowles et al. For 2015, we have a similar value of 730 Gt proposed by Colin Cunliff in the previously quoted ITIF paper.

To get a reasonable estimate for 2020, I have used the previously mentioned CUC together with the expected resource unit. The big change is the necessity to include bitcoin mining as a new category. I have applied the same scope 2 / scope 3 structure than data centers to extrapolate the CO2 footprint from the electricity consumption. There is a lot of uncertainty about the CO2 intensity of electricity consumption for bitcoin mining, as told in the Cambridge Bitcoin Web site. Thus, I evaluated the 2020 footprint of Bitcoin at 75Mt.

To get a full estimate of the “digital footprint”, we need to add the E&M sector. Malmodin & Lunden estimate the total footprint at 280Mt in 2015, out of which 160Mt are due to TV sets (note that STB are worth 53.3 Mt). For TV sets, Scope 3 (manufacturing mostly but also shipping) accounts for 70Mt, which yields a CUC (carbon unit footprint) of approximately 300kg (obtained from the shipment volume). The number of TV sets has grown from 1.47 billion in 2010 to 1.6 in 2015 according to Statista (while Malmodin quotes a higher figure of 1.9 billion) but the growth is now very slow. When we add all the numbers, we get a digital footprint of 1154 Mt in 2020, which is consistent with the values presented in the Freitag paper.

4. Prospective Analysis for 2030

This third section is very different from the previous two. Before, I tried to collect and sort out published numbers with my own attempt to select sources that I believe to be credible. Here I will build my own analysis and share with you my prevision for 2030 consequently. This is just food for thoughts, you should do your own … I believe in sharing my thoughts and exposing myself to constructive criticism, but I have no crystal ball and what follows only reflects what I think today and does not pretend to be right or accurate.

For the server parts, I consider that the progress on server unit consumption will continue, as the world moves to AMD servers of newer generations that are indeed more energy-frugal. For the manufacturing part, there is also some hope coming from the lab since the most energy-consuming part, that is the lithography, is also making progress. We can expect a significant reduction of server’s CUC in the future, but here I only assumed a small improvement in 2030.  There is a more complex question of using “green energy” for data centers. Here I apply the CO2 intensity of electricity that is produced globally (which is what the studies that I have used are doing). You might think that, as large cloud providers are switching to green energy, their scope 2 footprint should become zero. On the other hand, green energy that is obtained through certificates does not really impact the planet if the regional mix  does not change. The benefits of zero-carbon energy source only materialize for the planet when their share becomes significant in the total mix, which is not the case at the worldwide scale. This is a complex topic, which I do not have the time to address here, so my forecast does not include the benefits of “greener sources of energy” for data centers.

The 2030 forecast reported in the conclusion totals at 1060 Mt of CO2. The growth is mostly the growth of bitcoin mining, which is very hard to forecast. I propose 300TWh in this table, but I really hope to be shown wrong. Any guess here is as good as mine, since 300TWh is a huge slice of the electricity pie. For the other ICT categories, there is a fair amount of continuity since, at the first level of analysis, I have used constant CUC (unit carbon footprint) and regular (linear) growth of resource units. As we learn more about the expected improvements for manufacturing chipset, this table will be revised. As it stands, it is pretty conservative : it does not reflect much improvement, but no “crazy growth” either (with the exception of bitcoin mining).

The goal of the table was also to challenge some of the ratios (share of ICT or digital in the total CO2 / GHG footprint) so I have added the worldwide figures in the table to show the matching ratios. This requires a few comments:

1.       The values for 2010, 2015 and 2020 are easy to find, the value that I decide to put in 2030 is highly controversial, since they are related to political claims of various governments. For instance, I would not feel comfortable to propose a number for France. However, at the worldwide scale, the CO2 and GHG (greenhouse gas) have evolved regularly enough so that a conservative forecast is reasonably safe (+/- 10%)

2.       Beware of the difference between CO2 and GHG, that has been growing constantly over the past decades. Today, the non-CO2 gases account for half the warming effects of the emitted CO2. The 52Gt figure popularized by Bill Gates in his last book is no longer current. As a result, it seems fair to evaluate Digital with its share of CO2 emissions (the bold line in my table), but if you divide by GHG, you get a smaller number, significantly smaller. In 2020, the share of digital in the GHG emissions is 2%.

 

5. Conclusion

The following table reports all the numbers presented in this blog post. I have colored the cell to reflect my level of confidence. Green means that I was able to cross-check against a couple of reference with an error level that seems below 10%. Orange cells contain figures where I still lack enough sources or where I have a pending interrogation but consider the figure to be “a practical hypothesis” with an expected incertitude level at 30%. Pink cells are “educated guesses”, offered to give a complete prospective, but with no confidence. Obviously, this is “Work in Progress” (WIP) and is subject to regular updates.

 

 I decided to share this table, albeit its WIP status, because it is hard to collect these numbers, which are badly needed to make one’s opinion. Not only ICT impact is often exaggerated, but mostly the trends are inflated. It take a long time to follow the “data thread” to find out which data source was used in the first place. To illustrate this with an example, I received some “digital footprint ratios” while participating in the “fresque du numérique”, which quoted ARCEP numbers, which are based on ADEME, which quotes Carbon Shift Project and and greenit.fr, which eventually quote published research papers that are mostly about forecasts. A similar story could be told about the report of the “Convention Entreprise Climat” (cf. page 118).

 

To conclude, I will propose three pieces of advice:

-          Beware of percentages (like, the “impact of digital is 4% of greenhouse gases”), make sure that you are told which ratio is used, that is if you are given A/B as a percentage, make sure that you know which values are taken for A and B. When you see a document that has only percentages and no CO2 Mt values, keep a critical eye!

-          Once you are given A & B, make sure that the scope is clear : what is taken in and out, what are the resource units and the “carbon unit costs” (CUC)

-          Always check the references to find which sources are actual collected data analysis versus prospective studies. I have read more than 100 articles to prepare this post, while the actual number of trustable sources (published scientific articles) is less than 10, half of which are more of the “prospective” kind, trying to predict the future, as I did in Section 4. To illustrate this point, there are far more opinions about what the consumption of a laptop should be than scientific studies based on collected consumption of real users.