PreprintPDF Available

From Twin Transition to Twice the Burden: Digitalisation, Energy Demand, and Economic Growth

December 2024

DOI:10.2139/ssrn.5069660

Authors:

Jérôme Hambye-Verbrugghen

University of Strasbourg

Paul Edward Brockway

University of Leeds

Emmanuel Aramendia

University of Leeds

Show all 6 authorsHide

Preprints and early-stage research may not have been peer reviewed yet.

Content uploaded by Jérôme Hambye-Verbrugghen

Content may be subject to copyright.

From Twin Transition to Twice the Burden:

Digitalisation, Energy Demand, and Economic Growth

J´erˆome Hambye-Verbrugghen∗1, Stefano Bianchini1,

Paul Edward Brockway2, Emmanuel Aramendia2,

Matthew Kuperus Heun2,3,4, Zeke Marshall2

1BETA, CNRS, Strasbourg University 2School of Earth and Environment, University of Leeds

3Engineering Department, Calvin University 4School for Public Leadership, Stellenbosch University

Abstract

In this paper, we evaluate the potential of digitalisation to drive structural transformations

toward a sustainable economy. We apply an index decomposition analysis (IDA) to understand

the factors inﬂuencing energy demand in a panel of 31 high-income countries from 1971 to

2019. The IDA framework includes four factors related to the scale and sectoral composition

of the economy and technical improvements, accounting for the quality of energy ﬂows and

actual work potential through useful exergy measures. We ﬁrst apply the model at the sector

level across 16 productive industries to explore cross-sector heterogeneity in the structure of

energy demand. Industries are then classiﬁed by digital intensity categories, allowing us to

compare results across diﬀerent levels of digitalisation. We ﬁnd that economic growth is the

primary driver of energy use. While digitalisation alone does not fully explain trends in energy

demand, it is associated with substantial value added growth in high digital intensity sectors

and intensiﬁes the use of energy. This suggests that digitalisation may, in fact, exacerbate

economic-ecological tensions rather than alleviate them. We discuss the implications of these

ﬁndings in the context of recent policy actions aimed at accelerating the green and digital—

“twin”—transition.

Keywords: Structural Change, Energy, Energy Eﬃciency, Digitalisation, Technological Change

∗Corresponding author: jhambye@unistra.fr

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Highlights:

•Economic dynamics are most conductive to changes in energy demand.

•Cross-sector heterogeneity shapes trends in energy demand.

•Digitalisation polarises rates of economic growth, and thus energy use patterns.

•Eﬃciency gains must target energy demand reductions, addressing rebound.

•Greener technological trajectories require addressing economic-ecological tensions.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

1 Introduction

Climate change is one of the greatest challenges humanity has ever faced, with high risks of disrup-

tion to all life on Earth (Rockstr¨om et al. 2009,Steﬀen et al. 2015,Richardson et al. 2023,Lee et al.

2023). Public actions taken in the coming years will be decisive in determining whether we can

achieve a future aligned with the 1.5◦target set by the Paris Agreement and meet the Sustainable

Development Goals (SDGs), particularly SDG13: Take urgent action to combat climate change and

its impacts (UN General Assembly 2015). Can technological change—here, the widespread adop-

tion of digital technologies—drive structural shifts in energy use and support the transition to more

sustainable economic models? This is the overarching question we address in this paper.

Since the 1970s, economists have increasingly questioned the sustainability of economic growth

in light of increasing environmental degradation and resource depletion. This debate is well exem-

pliﬁed by the “Georgescu-Roegen/Daly vs. Solow/Stiglitz” controversy, which contrasts the belief

in unbounded productivity gains from natural resources and technology (Solow/Stiglitz) with the

argument that physical limits imposed by thermodynamics ultimately constrain production and

growth (Georgescu-Roegen/Daly) (Georgescu-Roegen 1975,Daly 1997,Solow 1997,Stiglitz 1997,

Mayumi et al. 1998,Couix 2019). The divide has persisted for long and the debate remains open

(see, e.g., Germain 2019,Couix 2019,Polewsky et al. 2024). Yet, optimistic assumptions regarding

productivity gains and the compatibility of Gross Domestic Product (GDP) growth with environ-

mental sustainability have shaped the (smart )green growth narrative of recent decades. Innovation

has been been placed at the cornerstone of the dominant climate change mitigation strategy in

high-income countries, reﬂecting an enduring belief in the role of technological change in decou-

pling economic growth from environmental impacts.1As a matter of fact, while green growth

requires absolute decoupling, there is a growing consensus that the scarce observations of such de-

coupling remain insuﬃcient to achieve mitigation targets (Savona & Ciarli 2019,Parrique et al.

2019,Le Qu´er´e et al. 2019,Haberl et al. 2020,Wiedenhofer et al. 2020). Regardless, the focus on

innovation and eﬃciency has also been rooted in the current discourse on the twin transition in Eu-

rope, which emphasises the integration of advanced digital technologies (ADTs) into environmental

strategies (Perez 2019,Lesher et al. 2019,Bianchini et al. 2022,2024).

The assumption that digital and green transitions can progress in tandem has attracted growing

criticism. While empirical evidence remains limited, there are in fact indications that digitalisa-

tion may not automatically align with sustainability goals (Fouquet & Hippe 2022). Information

technologies (IT) and some ADTs—e.g., Artiﬁcial Intelligence—are considered by many as General

Purpose Technologies (GPTs), and as such, they can unlock new opportunities and expand the

range of possibilities (Bresnahan & Trajtenberg 1995,Brynjolfsson & Yang 1996,David & Wright

1Decoupling refers to the “uncoupling” of resource use or environmental impacts from economic growth (Browne

et al. 2011,Regueiro-Ferreira & Alonso-Fern´andez 2022). Decoupling can be relative, meaning that resource use or

environmental impacts grow at a slower rate than GDP; or absolute, in which GDP growth is accompanied by a

reduction in resource use or environmental impacts (Parrique et al. 2019).

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

1999,Lee & Lee 2021), including the development of new products, production processes and ser-

vices that oﬀer environmental beneﬁts (Montresor & Vezzani 2023,Verendel 2023,Damioli et al.

2024). However, digitalisation also incurs substantial direct demand for energy and material related

to the production, use, and disposal of digital technologies (see Williams 2011,Jones 2018,Strubell

et al. 2019,Freitag et al. 2021,OECD 2022,Williams et al. 2022 for examples), all of which are

expected to keep growing in the future. Besides, there are potential indirect, or structural, eﬀects

which are complex and diﬃcult to quantify (Schulte et al. 2016,Yang & Shi 2018,Zhang & Wei

2022,Niebel et al. 2022,Ahmadova et al. 2022,Bartekov´a & B¨orkey 2022,Kunkel et al. 2023).

In a seminal article, Lange et al. (2020) propose an analytical framework to capture interactions

between digitalisation and energy consumption. They identify four channels through which digital-

isation may aﬀect the environment: (1) direct eﬀects related to the production, use and disposal of

ADTs; (2) digitally-induced gains in energy eﬃciency; (3) economic growth following productivity

gains; and (4) shifts in sectoral composition. Channels (1) and (3) are expected to intensify energy

use, while (2) and (4) should moderate demand through technical improvements and the expansion

of sectors with lower energy requirements. This framework informs our empirical analysis, which

uses data from a panel of 31 high-income countries over 1971–2019 to assess whether technological

change can drive the structural changes needed for a green transition. We propose some extensions

by integrating concepts from ecological and exergy economics with evolutionary economics (see

Section 2), allowing us to account for both structural and technological changes as well as the phys-

ical processes underlying economic production. Speciﬁcally, we conduct an energy decomposition

analysis across 16 productive sectors, comparing the results between digital-intensive sectors—as

deﬁned by the OECD (Calvino et al. 2018)—to understand the structural eﬀects of digitalisation.

We ﬁnd that digitalisation polarises the dynamics of energy demand through its boosting eﬀect

on economic growth, which remains the strongest driver of energy use. The expected eﬃciency

improvements following the adoption of ADTs are not observed, and the risk of digitally-induced

energy rebounds remains important. Our work highlights the existing economic-ecological tensions

that must be addressed for greener technological trajectories.

Our work contributes to the literature on the environmental impacts of digitalisation in several

ways. First, most studies focus on a single dimension of digitalisation—such as the number of ma-

chines or internet usage (Hig´on et al. 2017,Haseeb et al. 2019,Chimbo 2020,Oteng-Abayie et al.

2023), ICT capital (Bernstein & Madlener 2010,Khayyat et al. 2016,Schulte et al. 2016,Niebel

et al. 2022), ICT sectors (Zhou et al. 2018,2019,Wang et al. 2022), or ICT patents (Yan et al.

2018); and when multiple dimensions are considered (e.g., robots, skills, and digital capital), they

are often examined as independent variables (Matthess et al. 2023). Here, we rely on a taxonomy

that captures multiple facets of digital transformation, so that sectors vary in their development and

adoption of ADTs, the human capital needed to integrate them into production, and the extent to

which digital tools are used in interactions with clients and suppliers. Second, many existing studies

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

overlook the (strong) sectoral heterogeneity of technological change, often focusing on country-level

digitalisation (Ahmadova et al. 2022,Zhang & Wei 2022,Benedetti et al. 2023) or highly aggregated

sectors (Oteng-Abayie et al. 2023). In contrast, we propose a more granular, sector-level analysis

that considers both the economic and energy impacts of digitalisation. Third, and perhaps most

critically, many studies underestimate the role of energy and energy eﬃciency in economic produc-

tion, due to inconsistent deﬁnitions, misconceptions, mismeasurements, and limited accounting of

advances in ecological and exergy economics. In this study, we do our best to address these gaps.

The remainder of the paper is organised as follows: Section 2lays out the analytical framework

of our research; Section 3presents the decomposition model and data; Section 4reports the main

results and discusses their implications; and Section 5concludes with some remarks and implications

for policy.

2 Background

In this study, we investigate the indirect eﬀects of digitalisation on energy, focusing on the struc-

tural drivers of energy demand. Structural change is not limited to mere changes in economic

composition—as usually understood in energy analysis (see details in Section 2.1 and 2.3)—but is

a multi-layered process with interconnection among its components, that accompanies economic

growth through perpetual changes in technologies and products (Savona & Ciarli 2019,Ciarli &

Savona 2019). Economic composition is but one part of structural change, alongside growth dynam-

ics, technical improvements, the evolution of institutions or changes in the international division of

labour. Our analytical framework and empirical model draw on recent advances in ecological and

exergy economics—which emphasize the role of physical processes in economic production—as well

as on some concepts from evolutionary economics—which highlight the complex and heterogeneous

nature of technological change. In the following section, we ﬁrst elaborate on these theoretical foun-

dations (Section 2.1 and 2.2), then introduce the structural drivers of energy demand considered

in our analysis (Section 2.3.1), and ﬁnally present the taxonomy of sectors based on their level of

digitalisation used in this work (Section 2.3.2).

2.1 Metrics and modelling tools for energy analysis

Energy quality is central to economic production. Yet, surprisingly, most studies on the relation-

ship between digitalisation and energy have overlooked the quality of energy ﬂows and their actual

work potential, known as exergy (see Brockway et al. 2018 for a detailed outline).2Exergy eco-

2As a reminder: energy represents the total (heat) quantity of energy in a system, which is conserved (ﬁrst law of

thermodynamics); while exergy measures the work potential or available energy in a system, reﬂecting its quality

(second law of thermodynamics). Exergy accounts for irreversibilities and is essential when assessing the eﬃciency

of the system.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

nomics brings thermodynamic principles into economic analysis, considering energy across the three

stages of the energy conversion chain (ECC): primary, ﬁnal, and useful—along which energy/exergy

quantities can be measured (see Aramendia et al. 2021, Fig. 1 and Section 1.2 for details). The

primary stage represents raw energy resources extracted from nature; the ﬁnal stage captures the

energy/exergy purchased by end-users; and the useful stage measures energy/exergy at the point

of use in the production of energy services—such as heating, cooling, mechanical drive, lighting,

electronics, and muscle work (see Guevara 2014, Section 2.1.4; or Brockway et al. 2018, Table 8.1

for details).3The provision of energy services accounts for end-use device eﬃciencies, so assessing

energy/exergy at the useful stage captures improvements in second-law eﬃciency from advances in

energy technologies.

Despite the progress in exergy economics, however, many studies frequently conﬂate energy

intensity with thermodynamics-based (second-law) eﬃciency, neglecting the work potential, or ex-

ergy, of energy ﬂows (Guevara 2014,Proskuryakova & Kovalev 2015). In other words, they rely on

energy intensity as a (poor) proxy for energy eﬃciency. Energy intensity (I) is typically calculated

by dividing energy quantities (E) by the monetary value of GDP, gross output, or value added (Y),

or by physical quantities (Q) for speciﬁc goods or services:

I=E/Y or I=E/Q (1)

Yet this approach has clear limitations. For instance, energy intensity primarily captures changes in

ﬁrst-law (energy) eﬃciency, which measures only the quantity of energy input versus output, often

with signiﬁcant delays (Stern 2004,Proskuryakova & Kovalev 2015,Saunders et al. 2021). Moreover,

as shown in Equation (1), energy intensity depends on economic metrics that are frequently reported

with insuﬃcient detail or lack precise deﬁnitions, all of which aﬀect its accuracy (Semieniuk 2024).

Here, we explicitly accounts for the quality and work potential of energy ﬂows. Two main

empirical ﬁndings further underscore the importance of an exergy-based analysis: ﬁrst, qualitative

improvements in energy conversion technologies (i.e., second-law eﬃciency) is fundamental in ex-

plaining total factor productivity and, hence, economic growth (Santos et al. 2018,Sakai et al. 2018,

Santos et al. 2021). Second, these eﬃciency improvements appear much less substantial when the

quality of energy ﬂows is factored in (Aramendia et al. 2021,Brockway et al. 2024).

With this in mind, our empirical analysis relies on Index Decomposition Analysis (IDA), a

method particularly suited for investigating the evolution of energy use (or emissions).4Technical

details are provided in Section 3.1; for now, it is suﬃcient to note that IDA models decompose an

aggregate variable—i.e., energy use—into multiple components, oﬀering insights into the underlying

3In the remainder of this paper, exergy and work potential or work are used interchangeably, as well as useful exergy

and useful work.

4Since their development in the 1970s for energy balance analysis, decomposition methods (IDA, SDA, etc.) have

been continually reﬁned, with over 10,000 publications recorded as of 2023 (Wang & Yang 2023).

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

factors driving its variation. Few studies only have recently also integrated useful energy or work

potential within energy decomposition analyses (see, e.g., Guevara 2014,Brockway et al. 2015,

Silverio 2015,Guevara et al. 2016,Hardt et al. 2018,Aramendia et al. 2021,Ecclesia & Domingos

2022).

Diﬀerent IDA models vary in their methodological basis, but the core idea is to decompose an

aggregate indicator into three factors: production (or scale), structure (economic composition), and

technology (Hoekstra & Van den Bergh 2003). For example, in analysing national energy consump-

tion, changes in consumption can be decomposed into: the production eﬀect, which captures the

scale of overall activities using energy; the structure eﬀect, which reﬂects shifts in the composition of

these activities and thus the sectoral structure of energy demand; and the technology eﬀect, which

indicates the impact of energy converting technologies.5Due to its simplicity, IDA can be ﬂexi-

bly adapted to various dimensions (e.g., temporal and spatial) and scales (e.g., economies, sectors,

ﬁrms), and is therefore well-suited to our aims.

2.2 Multidimensionality, heterogeneity, and temporality of changes

In studying energy dynamics, it is important to account for the multidimensionality, heterogeneity

and temporality of technological and structural changes. And to this end, we believe it is useful to

consider (at least) three main caveats from evolutionary economics.

First, most studies on the environmental impact of digitalisation focus on a single dimension:

the number of machines or internet usage (Hig´on et al. 2017,Haseeb et al. 2019,Chimbo 2020,

Oteng-Abayie et al. 2023), ICT capital (Bernstein & Madlener 2010,Khayyat et al. 2016,Schulte

et al. 2016,Niebel et al. 2022), speciﬁc ICT sectors (Zhou et al. 2018,2019,Wang et al. 2022), or

ICT patents (Yan et al. 2018). We believe that digitalisation should instead be understood and

measured as a multidimensional transformation, encompassing ICT alongside other key ADTs.6At

a minimum, these should include artiﬁcial intelligence (AI), big data, IT infrastructure, and robotics

(see Bianchini et al. 2022 for a comprehensive classiﬁcation). In addition to technological change,

shifts in skills, markets, and business strategies also evolve and should be included in the analysis

(Calvino et al. 2018,Benedetti et al. 2023).

Second, there is a tendency to overlook the sectoral heterogeneity of technological change by

focusing on country-level digitalisation (Ahmadova et al. 2022,Benedetti et al. 2023) or using highly

aggregated sectoral data (Oteng-Abayie et al. 2023). Yet evolutionary economics teaches us that

heterogeneity matters (see, e.g., Dosi 2023, Ch.3 and 9). In a recent review, Zhang & Wei (2022)

conﬁrms that sector-level studies that address both the economic and environmental impacts of ICT

remain scarce. Moreover, a substantial body of literature shows that production technologies vary

5The structure eﬀect in decomposition methods is estimated considering only economic composition as part of

structure, and thus leads to a narrow deﬁnition of structural change in interpreting this eﬀect.

6In this analysis, we do not draw a strict distinction between ADTs and ICT. Instead, we consider ICT the historical

foundation for the emergence of ADTs and, therefore, a subset of them (Lee & Lee 2021).

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

signiﬁcantly by sector, with distinct patterns of diﬀusion and use across industries (Fierro et al.

2022,McElheran et al. 2024); and recent research aiming to classify sectors by levels of digitalisation

conﬁrms this strong heterogeneity (Calvino et al. 2018,Matthess et al. 2023). To account for all this,

as discussed further below, we conduct a sector-level analysis that captures multiple dimensions of

digital transformation.

Finally, most studies do not capture the path-dependent, long-term patterns of adoption and

use of (advanced) digital technologies, nor do they consider potential structural breaks in their

environmental impacts. The limited temporal scope of many analyses—understandable given the

challenges of accessing reliable long-term data—reﬂects a general tendency to treat technical change

as exogenous. This limitation has been suggested as a reason why the productivity eﬀects of

information technology were initially diﬃcult to observe; timing, therefore, matters too (David &

Wright 1999,Brynjolfsson & Hitt 2000). Cumulative eﬀects over time are particularly important;

for instance, it has been shown that “computer-enabled” organizational changes show much larger

impacts over longer periods (Brynjolfsson & Hitt 2000, p. 33). Thus, as further discussed in Section

3.2.1, we apply decomposition analysis on a long time series spanning almost 50 years (1971–2019)

to capture the cumulative eﬀects of technological and structural changes.

2.3 The analytical framework

We combine the elements discussed above into a single analytical framework. Conceptually, we ﬁrst

consider that digitalisation aﬀects various energy-related structural components across diﬀerent

sectors; and second, these components shape and deﬁne the dynamics of energy demand. Our

empirical approach is reversely structured: we begin with a decomposition model (brieﬂy introduced

in Section 2.1) that disaggregates energy demand into structural drivers—technical details in Section

3.1. Second, we examine the heterogeneous eﬀects of digitalisation across sectors based on the

methodology from the OECD (Calvino et al. 2018) (outlined in Section 1), comparing results across

levels of digital intensity—technical details in Section 2.3.2.

Our preference for sector-level decomposition over country-level analysis is motivated by three

main reasons. First, as discussed in previous sections, sectors exhibit distinct patterns of digital

penetration that cannot be captured at the country level. While ﬁrm-level studies may be best

suited to identify these changes, data limitations and the inability to aggregate results for country-

wide eﬀects make sector-level models better suited to our research question. Second, estimating

decomposition models directly at the sector level allows us to avoid the aggregation issues common

in country-level analysis (Weber 2009,Mulder & De Groot 2012,Guevara 2014). Indeed, as we

directly estimate separate models for each sector, we sidestep issues that may arise from diﬀerent

aggregation strategies. Third, sector-level decomposition often shows that the scale of production

(or activity eﬀect) plays a signiﬁcant role in energy use (Hajko 2012,Brockway et al. 2015,Heun &

Brockway 2019). Many studies, however, focus only on relative changes in sectoral composition—a

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

narrow view of structural change (see Henriques & Kander 2010,Mulder & De Groot 2012 for

example). When these changes are aggregated at the country level, some important sector-speciﬁc

patterns may be hidden, potentially underestimating structural shifts and overestimating the role

play by the scale of production.7

2.3.1 Structural drivers of energy demand

We distinguish two components of structural change which materialise through four driving fac-

tors. First, the composition component captures sectoral growth dynamics as drivers of economic

composition, and is measured using the scale eﬀect. Second, the technical components are split

between physical measures regarding the exergy-to-energy conversion ratio (conversion eﬀect) and

the second-law eﬃciency (eﬃciency eﬀect), and a hybrid physical-monetary measure for energy

productivity (productivity eﬀect).8Table 1lists the driving factors used to understand structural

changes in energy demand and their formula.

Table 1: List of driving factors for the decomposition model

Structural

Component Decomposition Factor Formula

Composition Scale eﬀect: S V A

Technical

Exergy-to-energy conversion eﬀect: IC Xf/Ef

Thermodynamic eﬃciency eﬀect: IE Xu/Xf

Productivity eﬀect: IP V A/Xu

Note: V A is value added, Xfis ﬁnal exergy, Efis ﬁnal energy, Xuis useful exergy.

Some concrete examples of causes for changes in the decomposition factors from Table 1include

the following. The scale eﬀect may be inﬂuenced by changes in total sales volume, market share,

or markups. The conversion eﬀect can result from shifts between ﬁnal energy carriers (e.g., from

heat to electricity) that have diﬀerent exergy-to-energy coeﬃcients (for more details, see Table 1

in Brockway et al. 2024). The eﬃciency eﬀect reﬂects changes in production processes, such as

the adoption of more or less eﬃcient machines for converting ﬁnal to useful energy, or shifts across

categories and subcategories of useful work (e.g., from medium temperature heat—MTH—200◦C

to MTH 300◦C). Finally, the productivity eﬀect may result from changes in the quality of products

without altering the requirements for useful work, or from shifts in the product structure of an

industry towards less (or more) exergy-intensive products.

7Forin et al. (2018) is one example of decomposition analysis adopting a sectoral perspective, but sectors are aggre-

gated across countries to capture potential eﬀects of industry oﬀshoring. Although this is an interesting and original

perspective, it diﬀers from the aim of our study. Mulder & De Groot (2012) also consider cross-sector heterogeneity

and conﬁrm diverging trends across sectors, particularly between manufacturing and services.

8While energy intensity was computed as I=E/Y , energy productivity (its inverse) is computed as P=I−1=Y/E.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

2.3.2 Measuring sectoral digitalisation

To account for the heterogeneous diﬀusion and use of digital technologies, skills, and business

models, we adopt the multidimensional framework proposed by the OECD (Calvino et al. 2018).

The framework identiﬁes three components that aﬀect the degree of digitalisation of a sector: a

technological component, a human capital component, and a market component.9

The technological component consists of ﬁve sub-indicators: intensity of investment in ICT

tangibles (1) and intangibles (2), intensity of intermediate expenditure in ICT goods (3) and services

(4), and robots density (5). Investment intensities are computed with total capital investment as

the denominator, and investment in computer hardware and telecommunications equipment (for

tangibles) or in software and databases (for intangibles) as numerators. Intermediate expenditure

intensities use input-output data to identify purchases made to ICT goods and ICT services sectors,

as a share of total output. Expenditure of ICT goods are characterised as purchases to the sector

Manufacture of computer, electronic and optical products (ISIC division 26), which mostly concerns

microchips and intermediate electronic components. Expenditure of ICT services are characterised

as purchases to the sector IT and other information services (ISIC divisions 62-63), which includes

hardware & software consultancy, computing equipment maintenance, and data processing. Finally,

robots density is computed by dividing the stock of industrial robots in a sector by hundreds of

employees.

The human capital component focuses on skills and is measured as the intensity of ICT spe-

cialists, computed as the percentage of ICT specialists over total employment. Finally, the market

component is measured as the share of turnover from online sales. The seven sub-components are

eventually aggregated to build an indicator for digital intensive sectors. For each sub-component,

all sectors are ranked based on their quartile position across four categories: low, medium-low,

medium-high, and high digital intensity. The global indicator is an average of the sector’s posi-

tion across the diﬀerent sub-components. This implies a sector may be classiﬁed in the low digital

intensity category while being ranked at the top of one sub-component. For instance, low digital

intensive sector Food products, beverages, and tobacco (ISIC divisions 10-12) is poorly ranked for

ICT investments, expenditures, and specialists, despite important online sales. Opposing exam-

ples for high digital intensity sectors are Transport equipment (ISIC divisions 29-30) with its high

stock of robots per employee, its online sales, and its ICT specialists; or sector Scientiﬁc research

and development (ISIC division 72) with its hardware and communications infrastructures, and its

expenditures in ICT goods and services, but no online sales.

9Data sources and speciﬁc metrics used to compute each indicator can be found in the Methodological Appendix of

Calvino et al. (2018); here we limit to a broad overview of the indicators.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

3 Methods and data

3.1 The sector-level decomposition model

Using the factors introduced in Section 2.3.1 (see Table 1), our decomposition model is based on

Equation (2). Efis ﬁnal energy, V A is value added, Xfis ﬁnal exergy, and Xuis useful exergy; S

is the scale eﬀect, IC is the conversion eﬀect, IE is the eﬃciency eﬀect and I P is the productivity

eﬀect. Subscripts iand jaccount for the i-th country and j-th sector.

ij =V Aij ×Ef

×Xf

×Xu

V Aij

=Sij ×ICij ×IEij ×IPij (2)

The choice between multiplicative and additive models does not aﬀect decomposition results, and

results can be converted straightforwardly from one form to another (Ang 2015, Section 3.2, p.236-

237). We choose the multiplicative version as the results are normalised around 1, which allows to

smoothly visualises the dynamics of change, and to make direct cross-sector comparisons regardless

of diﬀerences in aggregation levels.

Following Ecclesia & Domingos (2022), inverse values of the formulas presented in Table 1

are taken for the conversion,eﬃciency, and productivity eﬀects to ﬁt the accounting equality of

Equation (2). Inverse values for the three technical components reﬂect that improvements in these

metrics will translate into decreasing factors, thus contributing to reduce energy demand. Indeed,

improvements in the exergy-to-energy conversion ratio, the ﬁnal-to-useful exergy eﬃciency, and

the useful work productivity will lead to reductions in the conversion,eﬃciency, and productivity

eﬀects. Taking the rates of change in Equation (2), we get the following multiplicative relationship.

Denergy =ET

E0=DS×DIC ×DIE ×DIP (3)

Our model is based on the multiplicative LMDI model with type-I weights (details about the

mathematical properties for the LMDI-I model, and its diﬀerence with respect to LMDI-II, can

be found in Ang 2015). The general country-level estimation procedure for any driving factor DV

in country iis reproduced in Equation (4). L(x, y) = (x−y)/log(x/y) is the logarithmic mean

function, Trefers to the subsequent period, and 0 to the previous period. At the country-level,

the results are ﬁrst weighted by ωij , the ratio of the logarithmic mean function applied to the j-th

sector and to the entire country, then summed across all jsectors in country i.

DVi=X

expL(ET

ij , E0

ij )

L(ETi, E0i)·ln VT

ij = expωij ·lnVT

ij (4)

In our situation, the decomposition model is estimated for each country-sector (i, j) pair separately,

and there is no need for aggregation across sectors. Equation (4) is thus transformed to focus

directly on the changes at the “subgroup” (sector) level. Our estimation procedure for each factor

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

V={scale,conversion,eﬃciency, and productivity}is reported in Equation (5). The weight

parameter ωij cancels out from Equation (4) to Equation (5) due to our focus on country-sector

pairs, but our model remains structured to reﬂect the inﬂuence of the usual LMDI procedure.10

DVij = expL(ET

ij , E0

ij )

L(ET

ij , E0

ij )·lnVT

ij = exp1·lnVT

ij =VT

(5)

We obtain chained times series of DVij for each (i, j) pair from estimating Equation (5).11

Chained series refer to series computed dynamically, where each year tis compared directly to

its preceding year t−1, instead of a base year. Cumulative results are aggregated either over

the entire period—by taking the cumulative product of the whole time series—or by decade, by

taking the cumulative product of sub-series grouped by decade. This allows us to account for the

long-term dynamics of energy demand, by observing the contributions of each factor aggregated

over time. The size of the sample under consideration (433 (i, j ) pairs in total) and the focus on

sectoral dynamics make the analysis of individual series tedious. For this reason, we analyse the

distribution of the results and compare them across sector, digital intensity categories, and time.

Decomposition results in their multiplicative version are strictly positive: results in the interval

[0; 1] imply downward eﬀects for DVij , while results in the interval [1; +∞[ imply upward eﬀects

(see Ang 2015, Model 2 and 4, Table 3, p. 237 for an numerical example; and Heun & Brockway

2019, Fig. 7, p.9 for a graphical visualisation of index decomposition results).

3.2 Data

3.2.1 Energy and economic data

Energy and exergy data at the ﬁnal and useful stage are derived from the International Energy

Agency IEA 2023 Extended World Energy Balances (International Energy Agency 2023) and ac-

cessed through the country-level primary-ﬁnal-useful (CL-PFU) energy and exergy database (Heun

et al. 2024,Brockway et al. 2018,Marshall et al. 2024). These are available across 158 IEA coun-

tries, between periods from 1960-2020 (OECD countries) or 1971-2000 (non-OECD), and available

for sectors based on IEA classes (see United Nations Statistical Division 2018, for information about

the IEA classiﬁcation of sectors). Sector-level value added data comes from the STructural ANalysis

(STAN) OECD database, spanning 38 countries over the 1971-2019 period. To join the energy and

economic data, we must match sectors across the data sources: we aggregate IEA sectors to match

10Although our adaptation of the model omits the Log Mean component to which it owes its name, the simpliﬁcation

to mere rates of change does not compromise the eﬀectiveness of our approach or the validity of our conclusions.

11The results presented in Section 4, exclude outliers where the annual rate of change exceeds a tenfold increase.

This methodological choice only concerns sectors that are newly included in the database and are observed for

the ﬁrst time in a given year. Based on these observations, we make the conjecture that these substantial yearly

changes are due to statistical errors in early years of accounting for new sectors. By ﬁltering any instance where

DVij exceeds a tenfold increase, we only remove 17 observations, which constitutes less than 0.01% of our total

sample.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

ISIC (Rev.4) 2-digit divisions of sectors. After matching the data sources, we are able to build a

large panel of 31 high-income countries (see Appendix B), mostly OECD or EU countries, across 16

sectors that represent the entire productive economy from 1971 to 2019 (see Appendix Cfor details

about the data collection and aggregation mapping).

Final energy is our variable of interest and exists as such in the CL-PFU database (as do ﬁnal

exergy, useful energy and useful exergy) and is measured in terajoules (TJ). The data used in the

analysis accounts for gross energy, which includes energy producing sectors’ own energy use (i.e.,

energy industry own energy use) but excludes non-energy uses and muscle work (see Appendix Cfor

details). The scale eﬀect is measured with value added data from the STAN database, concerning

gross value added and expressed in millions of national currency, with chained prices (previous year

base). The conversion,eﬃciency, and productivity eﬀects are computed using ﬁnal energy and value

added, to which ﬁnal exergy and useful exergy data from the CL-PFU database—also measured

in TJ—are added. Having multiple currencies for the monetary and physical-monetary measures

used here does not pose a problem for our analysis; although this prevents direct comparison of

value added and energy productivity levels across countries with diﬀerent currencies, decomposition

analysis relies on rates of change in factors, and thus allows us to compare results across all countries

regardless of their currency.12

3.2.2 Sectoral digitalisation

The full time series for the digital-intensive sector classiﬁcation developed by Calvino et al. (2018)

is not publicly available; however, we have obtained the most recent classiﬁcation from Horv´at

& Webb (2020). Consistent with the rest of the STAN database, the classiﬁcation of sectors is

based on the ISIC Rev.4 industry classiﬁcation and thus requires to be matched with IEA sector

(United Nations 2008). Three of the productive sectors selected for this work cannot be exactly

matched, namely: Other industries;Machinery, electrical & electronic equipment ; and Commercial

and public services. Taking Other industries as an example, it is composed of Manufacture of

rubber and plastic products (ISIC division 22), Manufacture of furniture (ISIC division 31), and

Other manufacturing (ISIC division 32); respectively classiﬁed as medium-low digital intensity,

medium-high digital intensity and medium-high digital intensity. In this situation there is no

perfect matching strategy that could allow to assign one single digital intensity (DI) category to

Other industries. Hence, each ISIC division within our selection of 16 industries is given an equal

weight in determining the DI classiﬁcation. Consequently, the assignment of a sector to a speciﬁc

DI category is guided by the predominant classiﬁcation among its constituent ISIC divisions. In the

example above, Other industries has 66% of its composing ISIC divisions classiﬁed as medium-high

12Note that it would be appropriate to use the expression exergy productivity,useful exergy productivity, or useful

work productivity instead of energy productivity. For simplicity of writing and to refer to the common concept of

energy productivity, we use the term energy as a generic term encompassing all stages of the energy conversion chain

(ECC), and the qualitative measure of exergy. Thus in the remainder of the paper, the term energy productivity

may be used, but will always refer to our metric of productivity based on useful work.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Table 2: List of sectors classiﬁed by ISIC division and level of digital intensity.

Sector Full Name ISIC Div. Rev.4 DI-4

AGRI Agriculture, forestry, ﬁshing 01-03 L-DI

MINING Mining, quarrying 05-09 L-DI

FOOD Food products, beverages, tobacco 10-12 L-DI

TEXTIL Textiles, wearing apparel, leather 13-15 ML-DI

WOOD Wood, wood products 16 MH-DI

PAP Paper, pulp, printing 17-18 MH-DI

CHEMPHA Chemicals, chemical products, pharmaceutical

products

20-21 ML-DI

MINERAL Non-metallic minerals 23 ML-DI

METAL Metals, metal products 24 ML-DI

MACHIN∗Machinery, electrical and electronic products 25-28 MH-DI(1)

TRANSPEQ Transport equipment 29-30 H-DI

OTIND∗Other industries 22, 31-32 MH-DI(2)

COKE Coke & reﬁned petroleum products 19 ML-DI

ELECGAS Electricity, gas, steam, air conditioning 35 L-DI

CONSTR Construction 41-43 L-DI

COMSER∗Commercial & public services 33, 36-39, 45-96 H-DI(3)

Note: L-DI is low digital intensity, ML-DI is medium-low digital intensity, MH-DI is medium-high digital intensity,

H-DI is high digital intensity.

*Sectors among the 16 selected for which a perfect matching with the DI classiﬁcation was not possible. See Table

3.2 (p.18) in Horv´at & Webb (2020) for the original classiﬁcation.

(1) MACHIN: 25% ML-DI (ISIC division 25) and 75% MH-DI (ISIC divisions 26-28).

(2) OTIND: 33% ML-DI (ISIC division 22) and 66% MH-DI (ISIC divisions 31 and 32).

(3) COMSER: 19% L-DI (ISIC divisions 36-39, 49-53, 55-56, 68), 8.5% ML-DI (ISIC divisions 85-88), 25.5% MH-DI

(ISIC divisions 33, 45-47, 58-60, 84, 90-93) and 46.8% H-DI (ISIC divisions 61-66, 69-82, 94-96).

digital intensity; therefore, we assign this category to the sector.

Table 2presents the 16 sectors included in our analysis along with their digital intensity (DI)

classiﬁcations. Our main analysis uses four categories of digital intensity (DI-4); however, for

robustness, we also conduct a comparative analysis that consolidates digital intensity into just two

categories: low and high (see Supplementary Information Section S.1).

4 Results

This section is organised as follows. We ﬁrst present general results for the entire sample and across

sectors; next, we compare the results across categories of digital intensity. Tables 3.1–3.4 provide

the main set of results, with summary statistics calculated as unweighted mean and median values

for each dimension of interest (sectors, digital intensity categories, and time).13 Sectors in Tables

3.1–3.4 are ordered by their digital intensity (DI-4) category: L-DI is shown in Table 3.1; ML-DI

in Table 3.2; MH-DI in Table 3.3; and H-DI in Table 3.4. Figures 1and 2display the cumulative

13The unweighted statistics should not be interpreted as aggregate eﬀects. Instead, they represent the eﬀects for the

average country-sector pair, or the average country when results are grouped by sector, digital intensity category,

or time.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

results for a selection of sectors and the distribution of cumulative results aggregated by decade

across digital intensity categories, respectively. Figure 3summarises the relative contributions of

technical and composition components across all sectors and DI-4 categories. Additional results,

including those for the Conversion eﬀect, have been moved to Supplementary Information Section

S.2 and Appendix Dfor readability, as they show no signiﬁcant eﬀects to discuss.

4.1 Aggregate and sectoral energy demand

4.1.1 Economic dynamics matter more than physical processes

Across the entire sample, we observe a signiﬁcant mean increase in energy demand, with a 3.82-fold

rise from 1971 to 2019, while the median increase is more moderate at 1.07-fold (Tables 3.1–3.4).

The strongest driver of energy demand over this period is growth in value added, reﬂected in the

scale eﬀect (30.73; 4.24).14 We ﬁnd evidence of relative decoupling between energy use and economic

growth, as well as absolute decoupling on a few occasions: while the growth rate of value added

generally exceeds that of energy demand, we also observe periods with reductions in energy demand.

Nevertheless, economic growth mostly oﬀsets technical gains, despite progress towards more eﬃcient

(0.935; 0.880) and more productive (0.560; 0.280) processes. The magnitude of value added growth

is 16.1 (mean) or 1.04 (median) times stronger than the combined downward eﬀects of eﬃciency

and productivity.

The pivotal role of economic dynamics in driving energy demand remains observed across sectors

and across time: the growth rate in value added is the most striking diﬀerence between sectors

with strong growth in energy demand and those with low growth or reductions in energy demand

(Figure 1). Indeed, sectors with high growth in energy use are those with the strongest scale

eﬀects (e.g., TRANSPEQ,MINING,CONSTR, or COMSER). On the contrary, sectors with low

growth or reductions in energy demand are those with scale eﬀects that are below the full sample

mean or median, or that decrease over time (e.g., TEXTIL,OTIND,CHEMPHAR, or PAP).

This suggests that absolute or strong relative decoupling is facilitated when economic growth is

low, which is consistent with empirical evidence (Le Qu´er´e et al. 2019). Furthermore, we ﬁnd

that strong productivity gains help moderate growth in energy demand (e.g., TRANSPEQ), while

productivity declines amplify the impact of economic growth on energy use (e.g., CONSTR). Overall,

productivity plays a stronger role than eﬃciency in oﬀsetting energy demand growth associated

with low economic growth. Taken together, these results conﬁrm that economic factors (scale and

productivity eﬀects) are stronger drivers of energy demand compared to factors associated with

physical processes (conversion and eﬃciency eﬀects).

However, this should not be understood to mean that eﬃciency gains play no role in moderating

energy demand. While eﬃciency gains display lower variation, the absence of improvements for this

14Values in parentheses correspond to (mean; median). This notation applies throughout the remainder of this

section.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Table 3.1: Decomposition results by sector and decade for L-DI sectors, 1971-2019

Sector Period Energy Scale Eﬃciency Productivity

Avg. Med. Avg. Med. Avg. Med. Avg. Med.

Tot. Tot. 3.82 1.07 30.73 4.24 0.935 0.880 0.560 0.280

L-DI Tot. 3.62 1.30 24.20 5.24 0.900 0.872 0.741 0.280

AGRI Tot. 2.79 1.36 44.96 3.86 0.859 0.876 0.786 0.327

AGRI 1971-80 2.17 1.38 3.12 2.14 0.929 0.941 1.31 0.692

AGRI 1980-90 1.72 1.38 3.50 2.06 1.13 0.988 0.707 0.575

AGRI 1990-00 1.05 1.02 1.22 1.11 0.982 0.926 1.02 1.07

AGRI 2000-10 1.10 0.906 1.28 0.962 0.936 0.968 1.50 0.977

AGRI 2010-19 1.07 0.997 1.59 1.46 0.946 0.947 0.749 0.748

MINING Tot. 5.55 1.40 39.08 4.18 0.892 0.871 0.791 0.221

MINING 1971-80 2.29 1.10 6.89 2.94 0.978 0.971 0.469 0.408

MINING 1980-90 2.14 1.16 2.83 2.19 0.958 0.955 1.95 0.814

MINING 1990-00 1.20 1.07 1.31 1.19 0.956 0.926 1.07 1.05

MINING 2000-10 1.12 1.05 2.46 1.58 1.05 0.977 0.765 0.468

MINING 2010-19 1.23 0.955 1.25 1.10 0.957 0.943 1.17 1.07

FOOD Tot. 2.83 1.04 12.98 3.56 1.01 1.02 0.437 0.251

FOOD 1971-80 1.61 1.24 2.83 2.26 1.01 0.997 0.720 0.579

FOOD 1980-90 1.78 0.950 2.41 2.29 1.01 1.04 0.719 0.428

FOOD 1990-00 1.24 1.05 1.49 1.34 0.972 0.964 0.941 0.760

FOOD 2000-10 0.985 0.893 1.45 1.33 1.04 1.03 0.690 0.665

FOOD 2010-19 1.05 1.02 1.20 1.16 0.989 0.986 0.894 0.876

ELECGAS Tot. 1.48 1.22 8.09 4.22 0.770 0.793 0.447 0.306

ELECGAS 1971-80 1.39 1.41 2.48 2.50 0.919 0.911 0.691 0.649

ELECGAS 1980-90 1.40 1.35 2.18 2.07 0.890 0.933 0.794 0.729

ELECGAS 1990-00 1.07 1.07 1.55 1.30 0.976 0.968 0.829 0.834

ELECGAS 2000-10 1.14 1.07 2.23 1.83 0.977 0.975 0.692 0.599

ELECGAS 2010-19 0.947 0.897 1.20 1.08 0.858 0.857 0.966 0.967

CONSTR Tot. 5.36 1.85 14.97 9.04 0.965 0.962 1.22 0.165

CONSTR 1971-80 3.26 1.19 2.31 2.37 1.02 0.983 1.72 0.589

CONSTR 1980-90 1.68 0.956 2.93 2.31 0.948 0.960 0.935 0.350

CONSTR 1990-00 1.74 1.17 1.81 1.69 0.945 0.915 1.20 0.827

CONSTR 2000-10 1.27 1.05 1.98 1.68 1.04 1.02 0.740 0.631

CONSTR 2010-19 1.36 1.06 1.53 1.36 1.01 0.990 1.10 0.749

Note: Summary statistics for the full sample (Tot.) correspond to the unweighted cross-country and cross-sector

mean (Avg.) and median (Med.) values of the cumulative decomposition results, where cumulative series are

aggregated over the total period. Results by digital intensity category (L-, ML-, MH-, H-DI) correspond to the

unweighted cross-country average and median values of the (total) cumulative decomposition results. For each

DI-4 category, the summary statistics are derived for the (total) cumulative decomposition results across all the

sectors composing the DI-4 category. Results by sector correspond to the unweighted cross-country average and

median values of the (total) cumulative decomposition results. Results by sector and period correspond to the

unweighted cross-country average and median values across sectors for each decade, where cumulative results by

decade are obtained by multiplying the results for all periods in the same decade. For cross-sector comparison, the

two minimum and maximum values for each factor (across all Tables 3.1–3.4) have been highlighted.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Table 3.2: Decomposition results by sector and decade for ML-DI sectors, 1971-2019

Sector Period Energy Scale Eﬃciency Productivity

Avg. Med. Avg. Med. Avg. Med. Avg. Med.

Tot. Tot. 3.82 1.07 30.73 4.24 0.935 0.880 0.560 0.280

ML-DI Tot. 1.27 0.798 7.61 2.67 1.02 0.920 0.418 0.301

TEXTIL Tot. 1.20 0.285 5.37 1.39 0.959 0.936 0.305 0.249

TEXTIL 1971-80 1.58 1.11 2.70 1.62 0.991 0.993 0.873 0.794

TEXTIL 1980-90 1.36 0.852 1.78 1.53 1.11 1.06 0.618 0.502

TEXTIL 1990-00 0.965 0.931 1.27 1.10 0.964 0.953 0.884 0.897

TEXTIL 2000-10 0.486 0.494 0.853 0.801 1.01 0.960 0.637 0.599

TEXTIL 2010-19 0.804 0.782 1.10 1.12 0.962 0.934 0.779 0.769

CHEMPHAR Tot. 1.08 1.06 12.78 4.71 0.854 0.841 0.295 0.248

CHEMPHAR 1971-80 0.971 0.991 2.65 2.55 0.918 0.922 0.493 0.429

CHEMPHAR 1980-90 1.09 1.03 2.37 2.42 0.956 0.907 0.524 0.493

CHEMPHAR 1990-00 1.10 1.00 1.85 1.41 0.991 0.939 0.748 0.673

CHEMPHAR 2000-10 1.00 0.923 1.75 1.48 0.941 0.948 0.697 0.636

CHEMPHAR 2010-19 1.05 0.961 1.34 1.29 1.00 0.986 0.856 0.760

MINERAL Tot. 1.12 0.885 4.31 2.71 0.960 0.906 0.504 0.432

MINERAL 1971-80 2.26 1.54 1.79 1.74 1.03 1.03 1.20 0.867

MINERAL 1980-90 0.755 0.624 2.07 1.95 1.01 1.01 0.349 0.335

MINERAL 1990-00 1.03 0.996 1.64 1.42 0.969 0.974 0.733 0.786

MINERAL 2000-10 0.957 0.877 1.41 1.07 1.02 1.02 0.798 0.789

MINERAL 2010-19 1.08 0.991 1.28 1.20 0.960 0.955 0.901 0.908

METAL Tot. 1.20 0.822 4.93 1.94 1.31 0.906 0.605 0.466

METAL 1971-80 1.40 1.38 2.43 2.49 0.940 0.920 0.643 0.625

METAL 1980-90 0.890 1.01 2.46 2.50 0.963 0.940 0.370 0.419

METAL 1990-00 1.18 1.06 1.48 1.26 0.945 0.940 0.983 0.974

METAL 2000-10 1.09 0.848 1.59 1.18 1.07 0.967 0.728 0.774

METAL 2010-19 0.951 0.935 1.29 1.17 1.20 0.993 0.774 0.811

COKE Tot. 1.86 1.16 11.53 6.84 1.00 0.996 0.376 0.202

COKE 1971-80 1.62 1.13 5.18 4.39 0.816 0.876 0.624 0.557

COKE 1980-90 1.18 1.06 1.28 1.08 1.18 1.06 0.842 0.791

COKE 1990-00 1.02 0.973 1.96 1.38 0.948 0.958 0.929 0.805

COKE 2000-10 1.28 1.08 3.04 1.58 1.01 0.966 0.805 0.848

COKE 2010-19 0.955 0.953 1.59 1.37 1.05 1.03 0.809 0.736

Note: See bottom of Table 3.1 above.

factors may results in strong growth in energy demand. Indeed, we ﬁnd that strong scale eﬀects

with no gains in eﬃciency result in strong growth in energy demand, even when the productivity

eﬀect is strong. Without signiﬁcant eﬃciency improvements, even the strongest improvements in

productivity are not enough to eﬀectively achieve reductions in energy demand (e.g., COMSER).

In some cases, eﬃciency gains also compensate for weak productivity improvements, which on their

own are not enough to signiﬁcantly reduce the growth in energy demand, even with low value

added growth (e.g., WOOD or PAP ). While we mentioned above that economic dynamics, and

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Table 3.3: Decomposition results by sector and decade for MH-DI sectors, 1971-2019

Sector Period Energy Scale Eﬃciency Productivity

Avg. Med. Avg. Med. Avg. Med. Avg. Med.

Tot. Tot. 3.82 1.07 30.73 4.24 0.935 0.880 0.560 0.280

MH-DI Tot. 6.62 1.04 57.06 3.46 0.888 0.865 0.626 0.376

WOOD Tot. 2.02 1.70 3.61 2.34 0.874 0.888 1.05 0.541

WOOD 1971-80 3.04 1.41 2.40 2.46 1.14 1.14 1.10 0.492

WOOD 1980-90 1.20 0.886 1.66 1.61 1.08 1.04 0.685 0.604

WOOD 1990-00 1.31 1.20 1.67 1.40 0.897 0.907 0.987 0.884

WOOD 2000-10 1.30 1.06 1.31 1.20 1.01 1.01 1.07 0.854

WOOD 2010-19 1.09 1.01 1.26 1.27 0.931 0.963 1.07 0.881

PAP Tot. 1.20 1.02 2.91 1.61 0.831 0.825 0.829 0.842

PAP 1971-80 1.50 1.28 1.96 2.05 1.05 1.02 0.799 0.874

PAP 1980-90 1.29 1.19 2.49 2.40 1.03 1.03 0.516 0.518

PAP 1990-00 1.22 1.17 1.55 1.37 0.938 0.924 0.924 0.944

PAP 2000-10 1.02 0.924 1.22 1.01 0.916 0.939 1.04 0.884

PAP 2010-19 0.913 0.896 1.04 0.995 0.961 0.947 0.942 0.940

MACHIN∗Tot. 21.83 1.25 209.1 7.00 0.915 0.890 0.344 0.151

MACHIN 1971-80 2.61 1.17 5.02 2.54 1.08 1.06 0.568 0.513

MACHIN 1980-90 1.84 0.962 3.19 2.69 0.958 0.985 0.546 0.421

MACHIN 1990-00 0.976 0.890 2.29 1.70 0.996 0.998 0.571 0.541

MACHIN 2000-10 1.15 1.06 1.61 1.36 0.983 0.962 0.874 0.910

MACHIN 2010-19 1.17 0.946 1.30 1.25 0.931 0.929 1.10 0.881

OTIND Tot. 0.844 0.513 7.11 4.42 0.939 0.865 0.253 0.156

OTIND 1971-80 0.122 0.060 2.10 1.94 1.00 0.991 0.112 0.019

OTIND 1980-90 2.18 2.30 2.36 2.47 1.26 1.33 0.992 0.824

OTIND 1990-00 1.12 0.860 2.33 1.71 1.09 0.968 0.612 0.451

OTIND 2000-10 0.984 0.887 1.56 1.19 0.983 0.938 0.805 0.602

OTIND 2010-19 0.924 0.797 1.31 1.24 0.973 0.944 0.741 0.646

∗The value of MACHIN for energy and scale is surprisingly high, driven by the substantial value added growth of

this sector in South Korea, with a 5,347-fold increase between 1971 and 2018. When South Korea is removed from

the sample, the mean total cumulative change in energy demand drops to 1.2, compared to 21.83. While all countries

are kept in our results to avoid arbitrary outlier exclusions, it is worth noting that MACHIN is no longer among the

sectors with the strongest growth in energy demand once South Korea is excluded.

Note: See bottom of Table 3.1 above.

more particularly economic growth, may be more conductive to energy demand, it remains that

improving the eﬃciency of physical processes is fundamental to achieving the targeted reductions.

4.1.2 Heterogeneity across sectors is strong

The above results are conﬁrmed in most sectors, but do however conceal notable diﬀerences, con-

ﬁrming the importance of cross-sector heterogeneity for the dynamics of energy demand (Tables

3.1–3.4, see also Appendix Figures D.1 and D.2). Only 4 sectors display either mean or median val-

ues below 1, indicating an average or median reduction in energy demand in 2019 relative to 1971:

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Table 3.4: Decomposition results by sector and decade for H-DI sectors, 1971-2019

Sector Period Energy Scale Eﬃciency Productivity

Avg. Med. Avg. Med. Avg. Med. Avg. Med.

Tot. Tot. 3.82 1.07 30.73 4.24 0.935 0.880 0.560 0.280

H-DI Tot. 4.95 1.40 51.50 8.23 0.923 0.881 0.284 0.180

TRANSPEQ Tot. 6.86 1.15 51.29 7.78 0.821 0.838 0.298 0.187

TRANSPEQ 1971-80 1.47 1.38 3.14 2.60 0.973 0.976 0.712 0.484

TRANSPEQ 1980-90 0.953 0.985 4.52 3.06 1.06 1.02 0.342 0.297

TRANSPEQ 1990-00 4.53 1.12 2.99 1.85 0.971 0.929 1.54 0.619

TRANSPEQ 2000-10 1.08 0.993 1.73 1.19 0.942 0.968 0.733 0.711

TRANSPEQ 2010-19 0.987 0.970 1.37 1.40 0.900 0.876 0.868 0.823

COMSER Tot. 3.27 1.59 51.68 9.26 1.01 1.00 0.272 0.173

COMSER 1971-80 1.69 1.27 3.35 3.09 0.894 0.920 0.914 0.591

COMSER 1980-90 1.34 1.13 3.22 2.86 1.06 1.12 0.415 0.381

COMSER 1990-00 1.31 1.21 2.54 1.77 0.994 0.999 0.709 0.625

COMSER 2000-10 1.32 1.25 2.15 1.88 1.06 1.07 0.639 0.609

COMSER 2010-19 1.00 0.954 1.37 1.31 0.961 0.932 0.814 0.804

Note: See bottom of Table 3.1 above.

OTIND,TEXTIL,MINERAL, and METAL. In contrast, all other 12 sectors display moderate to

high increases in energy use, with up to a 22-fold increase for MACHIN, and 7 other sectors with

2-fold to 7-fold increases.15 The median values for increases in energy demand are much lower and

indicate, as might be expected, that outliers are driving the mean values upward. Despite observ-

able improvements in eﬃciency and productivity across most sectors, these gains are insuﬃcient to

counter the upward eﬀects of value added, with the few exceptions cited above.

Heterogeneity across sectors is stronger for economic dynamics (scale and productivity eﬀects)

than it is for thermodynamics-based measures of eﬃciency (conversion and eﬃciency eﬀects). Mean

scale and productivity values respectively range from 2.91 (PAP) to 51.68 (COMSER, if we disregard

the extreme mean value of MACHIN ) and from 0.253 (OTIND ) to 1.22 (CONSTR). Their median

values range from 1.37 (TEXTIL) to 9.26 (COMSER) and from 0.150 (MACHIN ) to 0.842 (PAP). In

contrast, the mean eﬃciency eﬀects range from 0.770 (ELECGAS) to 1.31 (METAL), and its median

values from 0.793 (ELECGAS) to 1.02 (FOOD). CONSTR,MACHIN,TRANSPEQ, and COMSER

are the sectors with the strongest value added growth, while TEXTIL,METAL,WOOD, and PAP

have the lowest. When considering technical components, there is no clear sector outperforming

for both factors: ELECGAS,PAP, and TRANSPEQ have the strongest eﬃciency improvements;

OTIND,MACHIN, and COMSER perform best in terms of productivity. In contrast, FOOD,

METAL,COKE, and COMSER perform poorly in terms of eﬃciency with no improvements or

deteriorations; CONSTR and WOOD perform poorly for productivity.

15MACHIN has a surprisingly strong mean scale eﬀect, 209.1, which leads to the strongest growth in energy demand.

This is due to the strong economic growth observed in this sector in South Korea over the entire period (5,347-fold).

If South Korea is excluded from the analysis, the mean growth in energy for MACHIN falls to 1.20, but its mean

scale eﬀect remains among the strongest (11.48).

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Figure 1: Bar charts of decomposition results for selected sectors

ENERGY

SCALE

EFFICIENCY

PRODUCTIVITY

Top 3

Worst 3

Top 3

Worst 3

Top 3

Worst 3

Top 3

Worst 3

0.2

0.6

1.4

1.8

−50

−25

−7

−3

SECTORS TEXTIL OTIND CHEMPHAR CONSTR MINING TRANSPEQ

Note: The bar chart displays the cumulative decomposition results aggregated over the full period for the Top 3 and

Worst 3 sectors. Top 3 sectors—TEXTIL,OTIND, and CHEMPHAR—display either reductions or low growth in

energy demand, while Worst 3 sectors—CONSTR,MINING, and TRANSPEQ—display the strongest increases. The

bars in the chart represent the mean value, while the blue square represents the median value. The horizontal red

line sets the threshold between upward and downward eﬀects. From left to right, the factors appear in the following

order: Energy, Scale, Eﬃciency, and Productivity. The scale of the y-axis is the same for Eﬃciency and Productivity.

4.1.3 The magnitude of eﬀects reduces over time

Over time, we ﬁrst notice that the periods of economic expansion in the 1970s–1980s and early 2000s

are characterised by the strongest increases in energy demand (Tables 3.1–3.4). The scale eﬀect

was strongest in these periods, which conﬁrms the strong connection between the size of economic

activities and energy use.

Even during economic slowdowns and periods of recession, this connection is conﬁrmed by the

associated slowdowns or reductions in energy demand. Although energy demand has increased in

most decades, its rate of increase has been declining over time, occasionally resulting in absolute

decreases. The reduced magnitude of the scale eﬀect is observed in recent decades for all sectors

but four: AGRI,METAL,COKE and CONSTR. We ﬁnd the same reduction in magnitude for the

productivity eﬀects, with a stable convergence towards of eﬀects 1, with some exceptions again. For

the few sectors for which economic growth has not slowed down over the decades, it may be the

reason for which productivity has remained stable or kept improving (e.g., AGRI,METAL). The

eﬃciency eﬀect displays more variation across time and no clear trend.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Overall, this generalised reduction in the magnitude of most eﬀects (both upward and downward)

over the decades, as seen with a convergence towards 1, may explain why reductions in energy

demand in the latest decades have seemed to spread to more sectors. This again underlines the

facilitating role of lower economic growth to achieve energy demand reductions through productivity

gains. Even with productivity gains converging towards 1, we still ﬁnd more reductions in energy

demand in the ﬁnal decades of the sample. If, however, productivity deteriorates in the ﬁnal

periods (i.e., DIP >1), energy demand may strongly increase, even with lower economic growth

(e.g., CONSTR or MACHIN ). Once again, variations in economic factors—economic growth and

energy productivity—are more strongly associated to variations in energy demand that those related

to physical processes.

4.2 Energy demand by digital intensity categories

4.2.1 Structural drivers of energy demand vary with levels of digitalisation

The dynamics of energy demand diﬀer across digital intensity categories (Figure 2; see also Table

D.1 and Figure D.3 in the Appendix). Mean and median values are consistently above 1 for L-DI

and H-DI, while ML-DI and MH-DI show overall lower eﬀects. For instance, the mean increase in

energy demand for ML-DI from 1971 to 2019 is only 27%, whereas for the other categories, increases

range from 3.62-fold (262%) to 6.62-fold (562%). Over time, the growth rates of energy demand

generally decline across all categories, though with distinct patterns. Energy demand slows and

begins declining in the ML-DI category starting in the 1990s, while similar declines appear in the

other categories only from the 2000s or 2010s.

Interestingly, the structure of energy demand diﬀers across categories, particularly for the scale

and productivity eﬀects. The distribution of value added growth across digital categories mirrors

the patterns observed for energy demand: L-DI and, especially, H-DI show higher values than ML-

DI and MH-DI (this is even more pronounced in the total cumulative results displayed in Appendix

Figure D.3). Diﬀerences in productivity eﬀects across categories are less clear-cut, though H-DI

consistently shows lower values than other groups (Appendix Table D.1). Finally, variation in the

eﬃciency eﬀect is minor, with weak cross-category diﬀerences.

4.2.2 Digitalisation reveals polarised dynamics of energy demand

The diﬀerences across categories reveal that digitalisation creates some polarisation of energy de-

mand dynamics, where disparities are mainly driven by economic growth rather than by eﬃciency

or productivity gains (Figure 2, see also Appendix Figure D.3). Table 4reinforces this conclusion:

sectors with high or low digital intensity (L-DI and H-DI) tend to have both high economic growth

and high energy demand, while sectors with medium digital intensity (ML-DI and MH-DI) have

lower growth and lower energy demand. One can also note that the sectors identiﬁed as best and

worse in terms of energy demand growth rate (Figure 1) respectively fall in ML-DI/MH-DI and

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Figure 2: Cumulative decomposition results by digital intensity categories

PRODUCTIVITY

EFFICIENCY

SCALE

ENERGY

0 1 2 3 4 5

DIGITAL INTENSITY LOW MEDIUM−LOW MEDIUM−HIGH HIGH

Note: The decomposition results in this ﬁgure are cumulative and have been aggregated by decade (the same box

plots with cumulative results aggregated over the entire period are available in Appendix Figure D.3). The yellow,

green, blue and purple correspond respectively to H-DI, MH-DI, ML-DI and L-DI sectors. Central box plot lines

correspond to the median values, and the blue diamonds to the mean values. The vertical red dashed line sets the

threshold between upward and downward eﬀects. From top to bottom, the factors appear in the following order:

Energy, Scale, Eﬃciency, and Productivity.

L-DI/H-DI categories. Although polarisation is less clear for the technical components (Appendix

Table D.1), improved eﬃciency and productivity, combined with lower scale eﬀects, result in ML-DI

and MH-DI sectors experiencing lower growth—or even reductions—in energy demand.

Overall, our results suggest that digital intensity does not impact energy demand in a straight-

forward, “linear” way. One might expect that moving from low to high digital intensity would

linearly result in both higher growth rates for value added and greater technical improvements

(Niebel et al. 2022,Zhang & Wei 2022).16 In contrast, our results indicate that energy demand

dynamics across levels of digital intensity are polarised, and that the primary driver is the disparity

in value added growth.

A few details on the dynamics speciﬁc to high digital intensity sectors are noteworthy. These

sectors form a distinct cluster, with value added growth substantially higher than in any other sector

or category, and coupled with generally stronger technical (mostly productivity) improvements

16By linear increase we do not mean an increase from a factor αfrom one DI category to another, but that the

direction of variation from one category to the next one remains the same such that DI-4 categories were always

ordered as such: L-DI, ML-DI, MH-DI, H-DI.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Table 4: Average rankings by digital intensity and factor

DI-4 Energy Scale Eﬃciency Productivity

Avg. Med. Avg. Med. Avg. Med. Avg. Med.

L11 11.4 10.6 9 7.8 8.6 11.6 8.6

ML 4.8 4.8 5.8 6.4 10.8 10.2 7.2 10

MH 7.5 8.25 6.25 7.25 6.5 6 9 8.5

H13.5 11 14.5 15 8.5 9 3 4.5

Note: The table provides the average rankings of sectors according to their digital intensity (DI-4) category. The

rankings were calculated using both the cross-country (unweighted) mean and median values of cumulative decom-

position results, where cumulative results were aggregated over the entire period. For each factor (Energy,Scale,

Eﬃciency and Productivity ), sectors were ranked from 1 to 16, where 1 corresponds to the lowest and 16 to the high-

est. Once the sectors were ranked across the mean and median values for each factor, the rankings were averaged

across the sectors within each DI-4 category, resulting in an overall average rank for each category. Higher averages

indicate stronger eﬀects for the underlying factors, while lower average indicate weaker eﬀects.

(Figure 3; see Figure E.1 in the Appendix for the same ﬁgure corrected from two outliers). In

contrast, sectors in the other categories display stronger within-category dispersion. ML-DI and

MH-DI sectors vary both across technical and composition components, and their mean values

(represented by the triangles in Figure 3) shift in parallel to the bisection line. This suggests that for

the sectors within these categories, even when economic growth is stronger, technical improvements

help to somewhat moderate the growth in energy demand. L-DI sectors are generally characterised

by lower variation for technical improvements, but vary widely across rates of economic growth.

This suggests that these disparities cannot be related directly to variations in technical components.

4.2.3 Economic growth intensiﬁes energy demand in digital intensive sectors

While scale eﬀects remain the strongest driver of energy use for both L-DI and H-DI, H-DI sectors

experience substantially stronger economic growth than L-DI, along with greater productivity im-

provements. In fact, the H-DI category shows a stronger correlation between combined technical

improvements and growth in value added, as illustrated by the dashed yellow line in Appendix

Figure E.1.17

This may indicate the occurrence of a digitally-induced energy rebound—a largely under-researched

empirical question (Coroam˘a & Mattern 2019,Kunkel & Tyﬁeld 2021,Kunkel et al. 2023)—or al-

ternatively, that strong technical improvements are facilitated by strong economic growth.18 The

second hypotheses is less plausible, however, as other sectors achieve strong technical improvements

even without a boost in economic growth (e.g., PAP). Our analysis thus conﬁrms that high digital

intensity is associated with stronger economic growth, which is consistent with previous evidence

17Figure 3also displays the correlation between technical improvements and changes in composition, but the corre-

lation for the L-DI category is strongly inﬂuenced by an outlier: the AGRI sector in Iceland displays a substantial

(1,143-fold) value added growth rate, and we remove this outlier in Appendix Figure E.1.

18It should be noted that rigorously assessing the potential digitally-induced energy rebound would require further

investigation and a formal analysis to ensure that energy eﬃciency precedes boosts in economic growth. This is

out of the scope of our work.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Figure 3: Total cumulative changes in composition components vs. technical components, by sector and

digital intensity category

205

215

0 10 20 30

(DconversionDefficiencyDproductivity)−1

Dscale

Country−sector value Cross−country mean value by sector Cross−country cross−sector mean value by DI−category

Digital intensity Low Medium−low Medium−high High

Note: The y-axis corresponds to changes in the composition component and is equivalent to the scale eﬀect. The

x-axis corresponds to (inverse) changes in the technical components and is equivalent to the (inverse) product of the

conversion,eﬃciency, and productivity eﬀects. Moving from the bottom to the top implies growth in value added

cumulative over the entire period, while moving from the left to the right implies stronger combined gains in technical

components. The red vertical and horizontal lines separate between upward and downward changes. Beneath the

horizontal line value added has decreased, while it has increased above. On the left of the vertical line deterioration

of technical components is observed, while technical gains are found on the right side of the vertical line. Dashed

coloured lines correspond to the linear regression line showing the correlation between the composition component

and the technical components. Observations resulting in increased energy demand over the entire period fall above

the black bisection line, while observations resulting in reduced demand fall below. Figure E.1 in the Appendix

displays the same plot with two outliers removed.

(Zhang & Wei 2022). While it is also associated with stronger productivity gains, these do not

translate into reductions in energy demand. Strong scale eﬀects systematically translate into higher

energy demand, whether technical improvements are strong (e.g., TRANSPEQ ) or low (e.g., CON-

STR). The degree to which lower scale eﬀects translate into reductions in energy demand varies

signiﬁcantly, and depends on the relative magnitude of technical improvements.

Finally, it remains true that L-DI sectors, with the exception of ELECGAS, struggle with techni-

cal gains, mostly eﬃciency. In these sectors, eﬃciency remains a critical challenge and future gains

might be fostered by digitalisation. However, strong value added growth should also be addressed

to ensure technical improvements translate to reductions in energy demand, as they do in some oc-

casion in ML-DI and MH-DI sectors. With respect to H-DI sectors, our work ﬁnds that irrespective

of their productivity improvements, the overall scale of economic activity must be questioned in

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

order to achieve targeted reductions. Digitalisation still falls short on the promises made in the twin

transition or smart green growth narratives, instead carrying twice the burden. It not only fails to

deliver the expected eﬃciency gains, but also shows little potential to drive the economic transfor-

mations needed—here, the decline of energy-hungry sectors—to achieve energy demand reductions.

While strategies to manage energy use should be tailored to the speciﬁc context of each sector, it is

fundamental that the risk of digitally-induced rebound is addressed, particularly for sectors already

strongly beneﬁting from ADTs.

Before concluding, it is important to acknowledge some limitations of our analysis as a caution-

ary note. First, our focus on advanced economies omits the eﬀects of outsourcing and globalisation,

which likely explain some of the reductions in energy use that we observe (Hardt et al. 2018,Niebel

et al. 2022). Second, the classiﬁcation of digital intensive industries suﬀers its own limitations

(Calvino et al. 2018), and may beneﬁt from improvements to capture the dynamics of digitalisation

with other emerging technologies such as AI (or generative AI, GenAI ), or the cross-country dif-

ferences in such developments. Third, the high aggregation of the Commercial and public services

sector in energy accounting only allows a limited understanding of the changes in sectoral com-

position occurring in this aggregate sector. Improving the disaggregation in data collection would

allow to better understand which of its constituent sectors are responsible for the growth in energy

demand. Finally, while large sample studies like ours allow to observe general trends in the data,

further research could beneﬁt from focusing on speciﬁc sectors to better understand the mechanisms

and impacts of ADTs on economic growth and energy demand at a more detailed level.

5 Conclusion

Combining evolutionary economics with ecological and exergy economics provides a fruitful theo-

retical background for our empirical analysis, aimed at reconciling a broader deﬁnition of structural

change with the physical groundings of production processes. Our analysis highlights clear cross-

sector structural diﬀerences in both the dynamics of energy demand and its driving factors. We

also observe structural eﬀects from digitalisation, though these eﬀects are more complex than an-

ticipated. Indeed, we ﬁnd that the dynamics of energy demand are polarised across high-growth,

high (energy) demand sectors and low-growth, low-demand sectors; with both low and high digital

intensity (L-DI and H-DI) sectors displaying high-growth and high-demand. Rather than driving

energy demand itself, digitalisation appears more as a booster to sector-speciﬁc dynamics of energy

demand, as it is associated with much higher economic growth.

Strong growth in value added thus remains the primary driver of ﬁnal energy demand, and signif-

icant reductions in demand are only achieved when both eﬃciency and productivity improvements

are combined to lower economic growth. Over time, the magnitude of scale and productivity eﬀects

reduces and converge to lower levels, which is consistent with the economic slowdown observed in

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

advanced economies in recent decades. At the same time, changes in physical processes measured

by the conversion and eﬃciency eﬀects display less variation and distinct patterns across sectors.

Eﬃciency gains alone are insuﬃcient to trigger energy savings during periods of strong economic

growth. In contrast, productivity improvements play a stronger role in mitigating scale eﬀect in

periods of modest value-added growth. Notably, energy demand reductions are observed primarily

during economic slowdowns, conﬁrming earlier ﬁndings (Le Qu´er´e et al. 2019). However, caution is

needed in interpreting these reductions as absolute decoupling, as they often occur during periods of

economic decline, reﬂecting recessions rather than true decoupling. These may also result from the

relocation of speciﬁc sectors to other countries outside of our sample (Hardt et al. 2018,Bogmans

et al. 2020).

The fact that energy demand reductions mainly occur during periods of low-growth or recession

highlights the challenge of reducing ecological impacts in a growing economy. This is particularly

true in the context of the twin transition, where hopes are high about potential eﬃciency gains.

Instead, our analysis seems to point to the double burden related to digitalisation: in addition to

its direct energy requirements, it is associated to an increase in output, but does not lead to the

expected technical improvements. With the pursuit of innovation and eﬃciency gains dominating

the public discourse and policy proposals for sustainability, technological change should be consid-

ered with respect to its broader economic, social, and ecological implications. This emphasises the

need to critically question the relevance of sustained economic growth in relation to societal and

environmental needs. Technological change is neither neutral nor rational; it is strongly connected

to rent-seeking and accumulation, and thus serves as a primary engine of economic growth in the

ﬁrst place (Schmelzer et al. 2022).

Post-growth and degrowth may oﬀer alternative paths for future research and policy strategies

that explicitly address these risks (Creutzig et al. 2018,Hardt et al. 2021). Yet these agendas should

not underestimate the transformational potential of technological change, which, as we argued ear-

lier, is a dynamic, complex, and heterogeneous process.

We therefore conclude by suggesting to exercise caution with respect to public policies only

targeting technical improvements through digitalisation, as the empirical evidence for this connec-

tion remains weak. Our analysis ﬁnds that digitalisation has not yet be able to produce absolute

and suﬃcient rates of decoupling, and while this may change in the future, refusing to address the

role played by economic growth seems unwise. The risk lies in anticipating the development of

digital technologies regardless of the knowledge on its environmental (e.g., digitally-induced energy

rebound eﬀects) or social implications (e.g., labour displacements or risks of monopoly), and to face

the long-term consequences of technological lock-in (Matthess et al. 2023).

Digitalisation itself is no longer an option. But guiding its trajectory is a matter of choice, and

must rely on sound evidence. Whether it may trigger sustainable structural transformations in the

future—be they technical, institutional, behavioural, organisational, compositional, or related to

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

the scale of overall economic activities—will likely remain a debated question. It is our hope that

the considerations presented here will inform economic research.

Acknowledgements

JHV acknowledges this work of the Interdisciplinary Thematic Institute MAKErS, as part of the

ITI 2021-2028 program of the University of Strasbourg, CNRS and INSERM, was supported by

IdEx Unistra (ANR-10-IDEX-0002), and by SFRI-STRAT’US project (ANR-20-SFRI-0012). We

acknowledge support for SB from the SEED project – Grant agreement n°ANR-22-CE26-0013. We

acknowledge support for PEB under EPSRC fellowship Award EP/R024251/1.

CRediT authorship contribution statement

Author contributions for this paper are shown in Table 5.

Table 5: Author contributions following CRediT (contributor roles taxonomy) (NISO 2023)

CRediT Role JHV SB PEB EA MKH ZM

Conceptualisation • •

Data curation • ••••

Formal analysis •

Funding acquisition • • •

Investigation ••••••

Methodology • • •

Project administration • •

Software • • • •

Supervision • •

Validation ••••••

Visualisation •

Writing – original draft • •

Writing – review & editing • • • •

Declaration of competing interest

The authors declare that they have no known competing ﬁnancial interests or personal relationships

that could have appeared to inﬂuence the work reported in this paper.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

References

Ahmadova, G., Delgado-M´arquez, B. L., Pedauga, L. E. & Leyva-de la Hiz, D. I. (2022), ‘Too

good to be true: The inverted u-shaped relationship between home-country digitalization and

environmental performance’, Ecological Economics 196, 107393. 4,5,7

Ang, B. (2015), ‘LMDI decomposition approach: A guide for implementation’, Energy Policy

86, 233–238.

URL: https://linkinghub.elsevier.com/retrieve/pii/S0301421515300173 11,12

Aramendia, E., Brockway, P. E., Pizzol, M. & Heun, M. K. (2021), ‘Moving from ﬁnal to useful

stage in energy-economy analysis: A critical assessment’, Applied Energy 283, 116194. 6,7

Bartekov´a, E. & B¨orkey, P. (2022), ‘Digitalisation for the transition to a resource eﬃcient and

circular economy’. 4

Benedetti, I., Guarini, G. & Laureti, T. (2023), ‘Digitalization in europe: A potential driver

of energy eﬃciency for the twin transition policy strategy’, Socio-Economic Planning Sciences

89, 101701. 5,7

Bernstein, R. & Madlener, R. (2010), ‘Impact of disaggregated ict capital on electricity intensity in

european manufacturing’, Applied Economics Letters 17(17), 1691–1695. 4,7

Bianchini, S., Bottero, P., Colagrossi, M., Damioli, G., Ghisetti, C., Michoud, K. et al. (2024),

‘Sustainable development goals and digital technologies: Mapping scientiﬁc research’. 3

Bianchini, S., Damioli, G. & Ghisetti, C. (2022), ‘The environmental eﬀects of the “twin” green

and digital transition in European regions’, Environmental and Resource Economics .

URL: https://link.springer.com/10.1007/s10640-022-00741-7 3,7

Bogmans, C., Kiyasseh, L., Matsumoto, A. & Pescatori, A. (2020), ‘Energy, eﬃciency gains and

economic development: when will global energy demand saturate?’. 26

Bresnahan, T. F. & Trajtenberg, M. (1995), ‘General purpose technologies ‘engines of growth’?’,

Journal of econometrics 65(1), 83–108. 3

Brockway, P. E., Heun, M. K., Marshall, Z., Aramendia, E., Steenwyk, P., Relph, T., Widjanarko,

M., Kim, J., Sainju, A. & Franzius, J. I. (2024), ‘A country-level primary-ﬁnal-useful (cl-pfu)

energy and exergy database v1. 2, 1960-2020’. 6,9,39,40

Brockway, P. E., Steinberger, J. K., Barrett, J. R. & Foxon, T. J. (2015), ‘Understanding china’s

past and future energy demand: An exergy eﬃciency and decomposition analysis’, Applied Energy

155, 892–903. 7,8

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Brockway, P., Sorrell, S., Foxon, T. & Miller, J. (2018), Exergy economics: new insights into

energy consumption and economic growth, in ‘Transitions in Energy Eﬃciency and Demand:

The Emergence, Diﬀusion and Impact of Low-Carbon Innovation’, p. 23. 5,6,12

Browne, D., O’Regan, B. & Moles, R. (2011), ‘Material ﬂow accounting in an irish city-region

1992–2002’, Journal of Cleaner Production 19(9-10), 967–976. 3

Brynjolfsson, E. & Hitt, L. M. (2000), ‘Beyond computation: Information technology, organizational

transformation and business performance’, Journal of Economic perspectives 14(4), 23–48. 8

Brynjolfsson, E. & Yang, S. (1996), ‘Information technology and productivity: a review of the

literature’, Advances in computers 43, 179–214. 3

Calvino, F., Criscuolo, C., Marcolin, L. & Squicciarini, M. (2018), A taxonomy of digital intensive

sectors, OECD Science, Technology and Industry Working Papers 2018/14. Series: OECD

Science, Technology and Industry Working Papers Volume: 2018/14.

URL: https://www.oecd-ilibrary.org/industry-and-services/a-taxonomy-of-digital-intensive-

sectorsf404736a−en4,7,8,10,13,25

Chimbo, B. (2020), ‘Information and communication technology and electricity consumption in

transitional economies’, International Journal of Energy Economics and Policy 10(3), 296–302.

4,7

Ciarli, T. & Savona, M. (2019), ‘Modelling the Evolution of Economic Structure and Climate

Change: A Review’, Ecological Economics 158, 51–64.

URL: https://linkinghub.elsevier.com/retrieve/pii/S0921800917313277 5

Coroam˘a, V. C. & Mattern, F. (2019), Digital rebound–why digitalization will not redeem us

our environmental sins, in ‘Proceedings 6th international conference on ICT for sustainability.

Lappeenranta. http://ceur-ws. org’, Vol. 2382. 23

Couix, Q. (2019), ‘Natural resources in the theory of production: the georgescu-roegen/daly

versus solow/stiglitz controversy’, The European Journal of the History of Economic Thought

26(6), 1341–1378. 3

Creutzig, F., Roy, J., Lamb, W. F., Azevedo, I. M., Bruine de Bruin, W., Dalkmann, H., Ede-

lenbosch, O. Y., Geels, F. W., Grubler, A., Hepburn, C. et al. (2018), ‘Towards demand-side

solutions for mitigating climate change’, Nature Climate Change 8(4), 260–263. 26

Daly, H. E. (1997), ‘Georgescu-roegen versus solow/stiglitz’, Ecological Economics 22(3), 261–266.

Damioli, G., Bianchini, S. & Ghisetti, C. (2024), ‘The emergence of a ‘twin transition’scientiﬁc

knowledge base in the european regions’, Regional Studies pp. 1–17. 4

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

David, P. & Wright, G. (1999), ‘General purpose technologies and surges in productivity: Historical

reﬂections on the future of the ict revolution.’. 3,8

Dosi, G. (2023), The foundations of complex evolving economies: Part one: Innovation, organiza-

tion, and industrial dynamics, Oxford University Press. 7

Ecclesia, M. V. & Domingos, T. (2022), ‘An Index Decomposition Analysis of Final and Useful

Exergy Intensities of the Portuguese Economy’, SSRN Electronic Journal .

URL: https://www.ssrn.com/abstract=4126061 7,11,40

Fierro, L. E., Caiani, A. & Russo, A. (2022), ‘Automation, job polarisation, and structural change’,

Journal of Economic Behavior & Organization 200, 499–535. 8

Forin, S., Radebach, A., Steckel, J. C. & Ward, H. (2018), ‘The eﬀect of industry delocalization on

global energy use: A global sectoral perspective’, Energy Economics 70, 233–243. 9

Fouquet, R. & Hippe, R. (2022), ‘Twin transitions of decarbonisation and digitalisation: A historical

perspective on energy and information in European economies’, Social Science p. 18. 3

Freitag, C., Berners-Lee, M., Widdicks, K., Knowles, B., Blair, G. S. & Friday, A. (2021), ‘The

real climate and transformative impact of ICT: A critique of estimates, trends, and regulations’,

Patterns 2(9), 100340.

URL: https://linkinghub.elsevier.com/retrieve/pii/S2666389921001884 4

Georgescu-Roegen, N. (1975), ‘Energy and economic myths’, Southern economic journal pp. 347–

381. 3

Germain, M. (2019), ‘Georgescu-roegen versus solow/stiglitz: back to a controversy’, Ecological

economics 160, 168–182. 3

Guevara, Z. H. Z. (2014), Three-level energy decoupling: Energy decoupling at the primary, ﬁnal

and useful levels of energy use, PhD thesis, Universidade de Lisboa (Portugal). 6,7,8

Guevara, Z., Sousa, T. & Domingos, T. (2016), ‘Insights on energy transitions in mexico from the

analysis of useful exergy 1971–2009’, Energies 9(7), 488. 7

Haberl, H., Wiedenhofer, D., Vir´ag, D., Kalt, G., Plank, B., Brockway, P., Fishman, T., Hausknost,

D., Krausmann, F., Leon-Gruchalski, B., Mayer, A., Pichler, M., Schaﬀartzik, A., Sousa, T.,

Streeck, J. & Creutzig, F. (2020), ‘A systematic review of the evidence on decoupling of GDP,

resource use and GHG emissions, part II: synthesizing the insights’, Environmental Research

Letters 15(6), 065003.

URL: https://iopscience.iop.org/article/10.1088/1748-9326/ab842a 3

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Hajko, V. (2012), ‘Changes in the energy consumption in eu-27 countries’, Review of Economic

Perspectives 12(1), 3–21. 8

Hardt, L., Barrett, J., Taylor, P. G. & Foxon, T. J. (2021), ‘What structural change is needed for

a post-growth economy: A framework of analysis and empirical evidence’, Ecological Economics

179, 106845. 26

Hardt, L., Owen, A., Brockway, P., Heun, M. K., Barrett, J., Taylor, P. G. & Foxon, T. J. (2018),

‘Untangling the drivers of energy reduction in the uk productive sectors: Eﬃciency or oﬀshoring?’,

Applied energy 223, 124–133. 7,25,26

Haseeb, A., Xia, E., Saud, S. et al. (2019), ‘Does information and communication technologies

improve environmental quality in the era of globalization? an empirical analysis’, Environmental

Science and Pollution Research 26, 8594–8608. 4,7

Henriques, S. T. & Kander, A. (2010), ‘The modest environmental relief resulting from the transition

to a service economy’, Ecological Economics 70(2), 271–282. 9

Heun, M. K. & Brockway, P. E. (2019), ‘Meeting 2030 primary energy and economic growth goals:

Mission impossible?’, Applied Energy 251, 112697. 8,12

Heun, M. K., Marshall, Z. & Aramendia, E. (2024), ‘Clpfudatabase: A suite of r packages for energy

conversion chain analysis’, Journal of Open Source Software 9(93), 6057. 12

Hig´on, D. A., Gholami, R. & Shirazi, F. (2017), ‘Ict and environmental sustainability: A global

perspective’, Telematics and Informatics 34(4), 85–95. 4,7

Hoekstra, R. & Van den Bergh, J. C. (2003), ‘Comparing structural decomposition analysis and

index’, Energy economics 25(1), 39–64. 7

Horv´at, P. & Webb, C. (2020), The OECD STAN Database for industrial analysis: Sources and

methods, OECD Science, Technology and Industry Working Papers 2020/10. Series: OECD

Science, Technology and Industry Working Papers Volume: 2020/10.

URL: https://www.oecd-ilibrary.org/science-and-technology/the-oecd-stan-database-for-

industrial-analysisece98fd3−en 13,14,39,47

International Energy Agency (2023), World energy balances, Report, International Energy Agency.

Jones, N. (2018), ‘Data centres are chewing up vast amounts of energy’. 4

Khayyat, N. T., Lee, J. & Heo, E. (2016), ‘How ict investment inﬂuences energy demand in south

korea and japan’, Energy Eﬃciency 9, 563–589. 4,7

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Kunkel, S., Neuh¨ausler, P., Matthess, M. & Dachrodt, M. F. (2023), ‘Industry 4.0 and energy in

manufacturing sectors in china’, Renewable and Sustainable Energy Reviews 188, 113712. 4,23

Kunkel, S. & Tyﬁeld, D. (2021), ‘Digitalisation, sustainable industrialisation and digital rebound–

asking the right questions for a strategic research agenda’, Energy Research & Social Science

82, 102295. 23

Lange, S., Pohl, J. & Santarius, T. (2020), ‘Digitalization and energy consumption. does ict reduce

energy demand?’, Ecological economics 176, 106760. 4

Le Qu´er´e, C., Korsbakken, J. I., Wilson, C., Tosun, J., Andrew, R., Andres, R. J., Canadell, J. G.,

Jordan, A., Peters, G. P. & van Vuuren, D. P. (2019), ‘Drivers of declining CO2 emissions in 18

developed economies’, Nature Climate Change 9(3), 213–217.

URL: http://www.nature.com/articles/s41558-019-0419-7 3,15,26

Lee, H., Calvin, K., Dasgupta, D., Krinner, G., Mukherji, A., Thorne, P., Trisos, C., Romero, J.,

Aldunce, P., Barrett, K. et al. (2023), Climate change 2023: synthesis report. Contribution of

working groups I, II and III to the sixth assessment report of the intergovernmental panel on

climate change, The Australian National University. 3

Lee, J. & Lee, K. (2021), ‘Is the fourth industrial revolution a continuation of the third industrial

revolution or something new under the sun? analyzing technological regimes using us patent

data’, Industrial and Corporate Change 30(1), 137–159. 4,7

Lesher, M., Gierten, D., Attrey, A., Carblanc, A. & Ferguson, S. (2019), ‘Going digital: Shaping

policies, ımproving lives’. 3

Marshall, Z., Brockway, P. E., Heun, M. K., Aramendia, E., Steenwyk, P., Relph, T., Widjanarko,

M., Kim, J., Sainju, A. & Franzius, J. I. (2024), ‘A country-level primary-ﬁnal-useful (cl-pfu)

energy and exergy database v1.2, 1960-2020’. [Dataset]. 12

Matthess, M., Kunkel, S., Dachrodt, M. F. & Beier, G. (2023), ‘The impact of digitalization on

energy intensity in manufacturing sectors–a panel data analysis for europe’, Journal of cleaner

production 397, 136598. 4,8,26

Mayumi, K., Giampietro, M. & Gowdy, J. M. (1998), ‘Georgescu-roegen/daly versus solow/stiglitz

revisited’, Ecological Economics 27(2), 115–117. 3

McElheran, K., Li, J. F., Brynjolfsson, E., Kroﬀ, Z., Dinlersoz, E., Foster, L. & Zolas, N. (2024),

‘Ai adoption in america: Who, what, and where’, Journal of Economics & Management Strategy

33(2), 375–415. 8

Montresor, S. & Vezzani, A. (2023), ‘Digital technologies and eco-innovation. evidence of the twin

transition from italian ﬁrms’, Industry and Innovation 30(7), 766–800. 4

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Mulder, P. & De Groot, H. L. (2012), ‘Structural change and convergence of energy intensity across

oecd countries, 1970–2005’, Energy Economics 34(6), 1910–1921. 8,9

Niebel, T., Axenbeck, J. & B¨oning, J. (2022), ‘Climate protection potential of digital

transformation-the role of production relocation’. 4,7,22,25

NISO (2023), ‘Contributor roles taxonomy (CRediT)’. Accessed: 23 August 2024.

URL: https://credit.niso.org 27

OECD (2022), Measuring the environmental impacts of artiﬁcial intelligence compute and applica-

tions: The AI footprint, OECD Digital Economy Papers 341. Series: OECD Digital Economy

Papers Volume: 341.

URL: https://www.oecd-ilibrary.org/science-and-technology/measuring-the-environmental-

impacts-of-artiﬁcial-intelligence-compute-and-applications7babf571 −en4

Oteng-Abayie, E. F., Dramani, J. B., Adusah-Poku, F., Amanor, K. & Quartey, J. D. (2023),

‘Decomposition and drivers of energy intensity in ghana’, Energy Strategy Reviews 47, 101090.

4,5,7

Parrique, T., Barth, J., Briens, F., Kerschner, C., Kraus-Polk, A., Kuokkanen, A. & Spangenberg,

J. H. (2019), ‘Decoupling debunked’, Evidence and arguments against green growth as a sole

strategy for sustainability. A study edited by the European Environment Bureau EEB .3

Perez, C. (2019), ‘19. transitioning to smart green growth: lessons from history’, Handbook on green

growth 1929, 447. 3

Polewsky, M., Hankammer, S., Kleer, R. & Antons, D. (2024), ‘Degrowth vs. green growth. a

computational review and interdisciplinary research agenda’, Ecological Economics 217, 108067.

Proskuryakova, L. & Kovalev, A. (2015), ‘Measuring energy eﬃciency: is energy intensity a good

evidence base?’, Applied energy 138, 450–459. 6

Regueiro-Ferreira, R. M. & Alonso-Fern´andez, P. (2022), ‘Ecological elasticity, decoupling, and

dematerialization: Insights from the eu-15 study (1970–2018)’, Ecological Indicators 140, 109010.

Richardson, K., Steﬀen, W., Lucht, W., Bendtsen, J., Cornell, S. E., Donges, J. F., Dr¨uke, M.,

Fetzer, I., Bala, G., Von Bloh, W. et al. (2023), ‘Earth beyond six of nine planetary boundaries’,

Science advances 9(37), eadh2458. 3

Rockstr¨om, J., Steﬀen, W., Noone, K., Persson, ˚

A., Chapin, F. S., Lambin, E. F., Lenton, T. M.,

Scheﬀer, M., Folke, C., Schellnhuber, H. J. et al. (2009), ‘A safe operating space for humanity’,

nature 461(7263), 472–475. 3

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Sakai, M., Brockway, P. E., Barrett, J. R. & Taylor, P. G. (2018), ‘Thermodynamic eﬃciency gains

and their role as a key ‘engine of economic growth”, Energies 12(1), 110. 6

Santos, J., Borges, A. S. & Domingos, T. (2021), ‘Exploring the links between total factor produc-

tivity and energy eﬃciency: Portugal, 1960–2014’, Energy Economics 101, 105407. 6

Santos, J., Domingos, T., Sousa, T. & Aubyn, M. S. (2018), ‘Useful exergy is key in obtaining

plausible aggregate production functions and recognizing the role of energy in economic growth:

Portugal 1960–2009’, Ecological Economics 148, 103–120. 6

Saunders, H. D., Roy, J., Azevedo, I. M., Chakravarty, D., Dasgupta, S., de la Rue du Can, S.,

Druckman, A., Fouquet, R., Grubb, M., Lin, B. et al. (2021), ‘Energy eﬃciency: what has research

delivered in the last 40 years?’, Annual review of environment and resources 46(1), 135–165. 6

Savona, M. & Ciarli, T. (2019), ‘Structural changes and sustainability. a selected review of the

empirical evidence’, Ecological economics 159, 244–260. 3,5

Schmelzer, M., Vetter, A. & Vansintjan, A. (2022), The future is degrowth: A guide to a world

beyond capitalism, Verso Books. 26

Schulte, P., Welsch, H. & Rexh¨auser, S. (2016), ‘Ict and the demand for energy: Evidence from

oecd countries’, Environmental and resource economics 63, 119–146. 4,7

Semieniuk, G. (2024), ‘Inconsistent deﬁnitions of gdp: Implications for estimates of decoupling’,

Ecological Economics 215, 108000. 6

Silverio, A. G. B. (2015), ‘Energy eﬃciency in denmark: perspectives from useful work accounting

and useful work decomposition analysis’. 7

Solow, R. M. (1997), ‘Georgescu-roegen versus solow-stiglitz’, Ecological Economics 22(3), 267–268.

Steﬀen, W., Richardson, K., Rockstr¨om, J., Cornell, S. E., Fetzer, I., Bennett, E. M., Biggs, R.,

Carpenter, S. R., De Vries, W., De Wit, C. A. et al. (2015), ‘Planetary boundaries: Guiding

human development on a changing planet’, science 347(6223), 1259855. 3

Stern, D. I. (2004), ‘The rise and fall of the environmental kuznets curve’, World development

32(8), 1419–1439. 6

Stiglitz, J. E. (1997), ‘Georgescu-roegen versus solow/stiglitz’, Ecological Economics 22(3), 267–268.

Strubell, E., Ganesh, A. & McCallum, A. (2019), ‘Energy and Policy Considerations for Deep

Learning in NLP’. arXiv:1906.02243 [cs].

URL: http://arxiv.org/abs/1906.02243 4

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

UN General Assembly (2015), ‘Transforming our world: The 2030 agenda for sustainable devel-

opment’. A/RES/70/1, 21 October 2015. Available at: https://www.refworld.org/legal/

resolution/unga/2015/en/111816 [Accessed 09 August 2024]. 3

United Nations (2008), International Standard industrial classiﬁcation of all economic activities

(ISIC), number no. 4, rev. 4 in ‘Statistical papers. Series M’, rev. 4 edn, United Nations, New

York. OCLC: ocn277738900. 13

United Nations Statistical Division (2018), International recommendations for energy statistics

(IRES). OCLC: 1081333526. 12,39

Verendel, V. (2023), ‘Tracking artiﬁcial intelligence in climate inventions with patent data’, Nature

Climate Change 13(1), 40–47. 4

Wang, H. & Yang, Y. (2023), ‘Decomposition analysis applied to energy and emissions: A literature

review’, Frontiers of Engineering Management 10(4), 625–639. 6

Wang, J., Dong, X. & Dong, K. (2022), ‘How digital industries aﬀect china’s carbon emissions?

analysis of the direct and indirect structural eﬀects’, Technology in Society 68, 101911. 4,7

Weber, C. L. (2009), ‘Measuring structural change and energy use: Decomposition of the us economy

from 1997 to 2002’, Energy policy 37(4), 1561–1570. 8

Wiedenhofer, D., Vir´ag, D., Kalt, G., Plank, B., Streeck, J., Pichler, M., Mayer, A., Krausmann, F.,

Brockway, P., Schaﬀartzik, A. et al. (2020), ‘A systematic review of the evidence on decoupling of

gdp, resource use and ghg emissions, part i: bibliometric and conceptual mapping’, Environmental

research letters 15(6), 063002. 3

Williams, E. (2011), ‘Environmental eﬀects of information and communications technologies’, nature

479(7373), 354–358. 4

Williams, L., Sovacool, B. K. & Foxon, T. J. (2022), ‘The energy use implications of 5G: Reviewing

whole network operational energy, embodied energy, and indirect eﬀects’, Renewable and Sus-

tainable Energy Reviews 157, 112033.

URL: https://linkinghub.elsevier.com/retrieve/pii/S1364032121012958 4

Yan, Z., Shi, R. & Yang, Z. (2018), ‘Ict development and sustainable energy consumption: a

perspective of energy productivity’, Sustainability 10(7), 2568. 4,7

Yang, S. & Shi, X. (2018), ‘Intangible capital and sectoral energy intensity: Evidence from 40

economies between 1995 and 2007’, Energy Policy 122, 118–128. 4

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Zhang, X. & Wei, C. (2022), ‘The economic and environmental impacts of information and com-

munication technology: A state-of-the-art review and prospects’, Resources, Conservation and

Recycling 185, 106477. 4,5,7,22,24

Zhou, X., Zhou, D. & Wang, Q. (2018), ‘How does information and communication technology aﬀect

china’s energy intensity? a three-tier structural decomposition analysis’, Energy 151, 748–759.

4,7

Zhou, X., Zhou, D., Wang, Q. et al. (2019), ‘How information and communication technology drives

carbon emissions: A sector-level analysis for china’, Energy Economics 81, 380–392. 4,7

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Appendix A List of abbreviations

Acronym/Term Deﬁnition/Description

General Terms:

ADTs Advanced digital technologies

DI Digital intensive/intensity

L-DI = low

ML-DI = medium-low

MH-DI = medium-high

H-DI = high

ECC Energy/exergy conversion chain

GDP Gross domestic product

GPTs General purpose technologies

ICT Information and communication technologies

MTH Medium-temperature heat

SDGs Sustainable development goals

Methods:

IDA Index decomposition analysis

LMDI Logarithmic Mean Divisia Index

SDA Structural decomposition analysis

Data:

CL-PFU Country-level primary, ﬁnal, useful (database )

IEA International Energy Agency

ISIC International standard industrial classiﬁcation

STAN STructural ANalysis (database)

TJ Terajoules

Equations:

IEnergy intensity

EEnergy

EfFinal energy

QProduction in physical quantities

V A Value added

XfFinal exergy

XuUseful exergy

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Acronym/Term Deﬁnition/Description

YProduction in monetary quantities

Φ Exergy-to-energy coeﬃcient

Driving factors:

DVThe rate of change of Vwith V={ﬁnal energy,S,IC,IE,IP}

S = Scale

IC = Inverse conversion

IE = Inverse eﬃciency

IP = Inverse productivity

Appendix B List of countries and time span with available

data

Table B.1: List of countries and time span with available data.

Country Time Span Country Time Span

AUS – Australia 1990–2018 ITA – Italy 1971–2019

AUT – Austria 1977–2018 JPN – Japan 1971–2019

BEL – Belgium 1971–2019 KOR – South Korea 1971–2018

CHE – Switzerland 1991–2018 LTU – Lithuania 1996–2018

CZE – Czech Republic 1971–2019 LUX – Luxembourg 1986–2018

DEU – Germany 1992–2019 LVA – Latvia 1996–2018

DNK – Denmark 1971–2018 NLD – Netherlands 1971–2018

ESP – Spain 1981–2018 NOR – Norway 1971–2018

EST – Estonia 1996–2018 NZL – New Zealand 1978–2018

FIN – Finland 1971–2018 POL – Poland 1996–2018

FRA – France 1971–2019 PRT – Portugal 1978–2018

GBR – Great Britain 1971–2019 SVK – Slovakia 1994–2019

GRC – Greece 1971–2019 SVN – Slovenia 1996–2018

HUN – Hungary 1992–2018 SWE – Sweden 1981–2019

IRL – Ireland 1996–2018 TUR – Turkey 1999–2019

ISL – Iceland 1974–2019

Note: The time span accounts for the ﬁrst year for which some industry data is available, but these time spans do

not account for perfectly balanced data. This means for the early periods, only some sectors may appear while data

for other sectors only start in the 1990s or early 2000s. Additional countries were available in both the CL-PFU and

the STAN databases but were excluded due to substantial missing values.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Appendix C Description of data collection and selection

from the CL-PFU and STAN OECD databases

More information on the CL-PFU database and access to the data can be found on the following

GitHub repository and link:

https://github.com/EnergyEconomyDecoupling/CLPFUDatabase

https://doi.org/10.5518/1199.

More information about the STAN OECD database can be found in Horv´at & Webb (2020) or on

the following link:

https://www.oecd.org/en/data/datasets/structural-analysis-database.html

C.1 Merging IEA products with ISIC Rev.4 2-digits divisions of sectors

The CL-PFU database includes data across 7 aggregate sectors, 46 detailed sub-sectors, and 68 ﬁnal

energy products. It also accounts for non-energy uses of energy across 16 sub-sectors, which track

energy resources used for purposes other than generating heat, electricity, or power, such as chem-

ical or plastic production. The CL-PFU aggregation mapping is available in the Data Availability

Statement in Brockway et al. (2024). This paper uses the sectoral level data from the CL-PFU

database, covering 34 IEA products. The data include energy industry own use (EIOU), which is

of interest for capturing potential structural transformations within energy industries. There is no

double accounting: energy industries are treated as ﬁnal energy consumers, similar to other sectors.

Non-energy uses of energy are not included. This approach helps explore the eﬀects of digitalisation

on the use of energy resources, focusing only on energy purposes. Our analysis excludes muscle work

(including feedstock inputs and human or animal labour) as it focuses on the structural impact of

technological change on resource use, not on human or animal labour.

The 34 IEA products correspond to sectors, sub-sectors, or ﬁnal energy products and are mapped

to their respective ISIC Rev.4 classes or divisions, based on Table 5.1 (p. 59) and Table 5.3 (p. 66)

from United Nations Statistical Division (2018). This mapping links the 34 IEA products in the

CL-PFU database to 18 ISIC Rev.4 2-digit divisions (or groups of divisions, e.g., the Commercial

and public services sector is composed of multiple ISIC divisions). The 16 productive sectors used

in this paper’s decomposition model are listed in Table 2, along with 2 non-productive sectors:

Residential and Transport. The ﬁnal mapping ﬁle and Rcode are available upon request.

C.1.1 The nuclear industry

The CL-PFU database relies on the International Energy Agency (IEA) Extended World Energy

Balances (EWEB) data, which presents aggregation challenges in some cases, notably the nuclear

industry. It is the only IEA energy industry that cannot be perfectly mapped with speciﬁc ISIC

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

divisions. The IEA nuclear energy industry covers both the extraction and processing of nuclear fuels

in combination, making it impossible to separate between these processes. Extraction corresponds

to ISIC class 0721 (Mining of uranium and thorium ores), while processing aligns with ISIC class

2011 (Manufacture of basic chemicals), placing the nuclear industry between ISIC divisions 05-09

(Mining & quarrying) and 20-21 (Manufacture of chemicals and chemical products). To the best

of our knowledge, there is no empirical basis for preferring one ISIC division over the other for

aggregating the nuclear industry’s own use of energy. In this analysis, we arbitrarily include its

energy use in the Mining & quarrying sector. Upon review, we ﬁnd this choice has a negligible eﬀect

on the aggregate results. However, in speciﬁc countries where the nuclear industry is important,

such as France, Slovakia, or Belgium, decomposition results may vary considerably between the two

sectors involved in nuclear energy production.

C.2 Exclusion of non-productive sectors

The two non-productive sectors from the CL-PFU data, Residential and Transport, account for a

large share of total energy use (Brockway et al. 2024, Figure 6, p.13). While decomposition analyses

have been adapted to account for non-productive sectors (see Ecclesia & Domingos 2022), conduct-

ing such analyses on a large panel is challenging for two main reasons. First, alternative measures of

energy intensity or productivity are required due to the absence of monetary metrics (value added,

gross output) for these sectors. One option is to approximate energy intensity by the ratio of energy

use to total value added or gross output. Another approach is to use physical measures, such as

energy intensity per ﬂoor area or per kilometers travelled, which requires additional data.

The second issue concerns how the IEA accounts for transport energy use. The Transport sector

can be divided into six sub-categories (road, rail, domestic aviation, domestic navigation, pipeline

transport, and not elsewhere speciﬁed), but commercial and private transport data are combined

and cannot be diﬀerentiated. As a result, it is impossible to separate productive (commercial) from

non-productive (private) uses of energy for transport. The method in Ecclesia & Domingos (2022)

to split productive and non-productive transport energy use is not applicable to the large sample

in our analysis. Therefore, both non-productive sectors are excluded.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Appendix D Complementary Tables and Figure for decom-

position results

Figure D.1: Boxplots of cumulative (decadal) decomposition results by sector

PRODUCTIVITY

EFFICIENCY

SCALE

ENERGY

0 1 2 3 4 5

SECTORS AGRI

MINING FOOD

ELECGAS CONSTR

TEXTIL CHEMPHAR

MINERAL METAL

COKE WOOD

PAP MACHIN

OTIND TRANSPEQ

COMSER

Note: The decomposition results in this Figure are cumulative and have been aggregated by decade. All sectors

are distincted by their colour and are ordered by digital intensity (DI-4) groups, going from L-DI (purple) to H-DI

(yellow). Central boxplot lines corresponds to the median values, and the blue diamonds to the mean values. The

vertical red dashed line sets the threshold between upward and downward eﬀects. From top to bottom, the factors

appear in the following order: Energy, Scale, Conversion, Eﬃciency, and Productivity.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Figure D.2: Boxplots of cumulative (total) decomposition results by sector

PRODUCTIVITY

EFFICIENCY

SCALE

ENERGY

0 10 20 30 40 50

SECTORS AGRI

MINING FOOD

ELECGAS CONSTR

TEXTIL CHEMPHAR

MINERAL METAL

COKE WOOD

PAP MACHIN

OTIND TRANSPEQ

COMSER

Note: The decomposition results in this Figure are cumulative and have been aggregated over the entire period.

All sectors are distincted by their colour and are ordered by digital intensity (DI-4) groups, going from L-DI (purple)

to H-DI (yellow). Central boxplot lines corresponds to the median values, and the blue diamonds to the mean values.

The vertical red dashed line sets the threshold between upward and downward eﬀects. From top to bottom, the

factors appear in the following order: Energy, Scale, Conversion, Eﬃciency, and Productivity.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Table D.1: Decomposition results by DI-4 category and decade, 1971-2019

Sector Period Energy Scale Eﬃciency Productivity

Avg. Med. Avg. Med. Avg. Med. Avg. Med.

Tot. Tot. 3.82 1.07 30.73 4.24 0.935 0.880 0.560 0.280

L Tot. 3.62 1.30 24.20 5.24 0.900 0.872 0.741 0.280

L 1971-80 2.23 1.26 3.66 2.34 0.973 0.957 1.02 0.582

L 1980-90 1.77 1.09 2.88 2.17 0.999 0.960 1.05 0.534

L 1990-00 1.26 1.07 1.48 1.31 0.966 0.955 1.01 0.891

L 2000-10 1.12 0.993 1.88 1.48 1.01 0.992 0.881 0.662

L 2010-19 1.13 1.00 1.36 1.23 0.953 0.955 0.976 0.854

ML Tot. 1.27 0.798 7.61 2.67 1.02 0.920 0.418 0.301

ML 1971-80 1.50 1.10 3.18 2.28 0.930 0.950 0.733 0.587

ML 1980-90 1.15 0.967 1.88 1.74 1.07 0.977 0.605 0.535

ML 1990-00 1.06 0.999 1.63 1.31 0.964 0.954 0.854 0.817

ML 2000-10 0.943 0.846 1.66 1.15 1.01 0.966 0.729 0.715

ML 2010-19 0.965 0.935 1.31 1.18 1.03 0.973 0.823 0.795

MH Tot. 6.62 1.04 57.06 3.46 0.888 0.865 0.626 0.376

MH 1971-80 2.08 1.12 3.58 2.33 1.07 1.06 0.626 0.490

MH 1980-90 1.69 1.07 2.72 2.40 1.03 1.00 0.624 0.506

MH 1990-00 1.16 1.13 1.95 1.51 0.977 0.947 0.776 0.729

MH 2000-10 1.11 0.986 1.42 1.17 0.973 0.962 0.950 0.831

MH 2010-19 1.03 0.938 1.22 1.17 0.949 0.944 0.967 0.891

H Tot. 4.95 1.40 51.50 8.23 0.923 0.881 0.284 0.180

H 1971-80 1.59 1.38 3.25 2.92 0.932 0.968 0.818 0.580

H 1980-90 1.17 0.989 3.79 2.87 1.06 1.04 0.383 0.375

H 1990-00 2.78 1.14 2.75 1.77 0.984 0.974 1.09 0.623

H 2000-10 1.21 1.10 1.96 1.63 1.01 0.995 0.682 0.680

H 2010-19 0.994 0.963 1.37 1.35 0.933 0.922 0.839 0.823

Note: Summary statistics for the full sample (Tot.) correspond to the unweighted cross-country and cross-sector

mean (Avg.) and median (Med.) values of the cumulative decomposition results derived, where cumulative series

are aggregated over the total period. Results by digital intensity category (L-, H-DI) correspond to the unweighted

cross-country average and median values of the (total) cumulative decomposition results. For each DI-4 category, the

summary statistics are derived for the (total) cumulative decomposition results across all the sectors composing the

DI-4 category. Results by DI-4 category and period correspond to the unweighted cross-country average and median

values across sectors for each decade, where cumulative results by decade are obtained by multiplying the results for

all periods in the same decade. For cross-sector comparison, the minimum and maximum values for each factor

have been highlighted.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Figure D.3: Boxplots of cumulative (total) decomposition results by sector

PRODUCTIVITY

EFFICIENCY

SCALE

ENERGY

0 20 40

DIGITAL INTENSITY LOW MEDIUM−LOW MEDIUM−HIGH HIGH

Note: The decomposition results in this Figure are cumulative and have been aggregated over the total period. The

yellow, green, blue and purple correspond respectively to H-DI, MH-DI, ML-DI and L-DI sectors. Central boxplot

lines corresponds to the median values, and the blue diamonds to the mean values. The vertical red dashed line sets

the threshold between upward and downward eﬀects. From top to bottom, the factors appear in the following order:

Energy, Scale, Eﬃciency, and Productivity.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Appendix E Scatterplot of cumulative changes in compo-

sition versus technical improvements, without

outliers

Figure E.1: Figure 3without the MACHIN sector from South Korea and the AGRI sector from Iceland

205

215

0 10 20 30

(DconversionDefficiencyDproductivity)−1

Dscale

Country−sector value Cross−country mean value by sector Cross−country cross−sector mean value by DI−category

Digital intensity Low Medium−low Medium−high High

Note: The y-axis corresponds to changes in the composition component and is equivalent to the scale eﬀect. The

x-axis corresponds to (inverse) changes in the technical components and is equivalent to the (inverse) product of the

conversion,eﬃciency, and productivity eﬀects.

Moving from the bottom to the top implies growth in value added cumulative over the entire period, while moving

from the left to the right implies stronger combined gains in technical components. The red vertical and horizontal

lines separate between upward and downward changes. Beneath the horizontal line value added has decreased, while

it has increased above. On the left of the vertical line deterioration of technical components is observed, while

technical gains are found on the right side of the vertical line.

Dashed coloured lines correspond to the linear regression line showing the correlation between the composition

component and the technical components.

Observations resulting in increased energy demand over the entire period fall above the black bisection line, while

observations resulting in reduced demand fall below.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Supplementary Information for:

From twin transition to twice the burden: digi-

talisation, energy demand, and economic growth

J´erˆome Hambye-Verbrugghen (jhambye@unistra.fr)1,

Stefano Bianchini1,

Paul Edward Brockway2,

Emmanuel Aramendia2,

Matthew Kuperus Heun2,3,4,

Zeke Marshall2

1BETA, CNRS, Strasbourg University.

2School of Earth and Environment, University of Leeds.

3Engineering Department, Calvin University.

4School for Public Leadership, Stellenbosch University

November 26, 2024

Introduction

This document provides supplementary information and results to support the ﬁndings reported in

the main article titled ”From twin transition to twice the burden: digitalisation, energy demand,

and economic growth.”

Contents

S.1 Robustness analysis: two categories of digital intensity (DI-2)

S.2 Results for the Conversion eﬀect

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

S.1 Robustness analysis: two categories of digital intensity

(DI-2)

In this section, we perform a robustness analysis to verify whether the previous results hold when

going from four categories of digital intensive industries (DI-4) to only two (DI-2). In particular, this

allows to address the issue raised in Section 3.2.2 of the main article, namely that perfect matching

is not possible for three of the macro-sectors analysed: Other industries;Machinery, electrical &

electronic equipment; and Commercial and public services. As a robustness, we simply include both

ML-DI and L-DI in a single category (L-DI), and both MH-DI and H-DI in another (H-DI). Table

S.1 presents the ﬁnal 16 sectors under scrutiny with their DI-4 and DI-2 classiﬁcation.

Table S.1: List of industries classiﬁed by ISIC division and level of digital intensity.

Sector Full Name ISIC Div. Rev.4 DI-2 DI-4

AGRI Agriculture, forestry, ﬁshing 01-03 L-DI L-DI

MINING Mining, quarrying 05-09 L-DI L-DI

FOOD Food products, beverages, tobacco 10-12 L-DI L-DI

TEXTIL Textiles, wearing apparel, leather 13-15 L-DI ML-DI

WOOD Wood, wood products 16 H-DI MH-DI

PAP Paper, pulp, printing 17-18 H-DI MH-DI

CHEMPHA Chemicals, chemical products,

pharmaceutical products

20-21 L-DI ML-DI

MINERAL Non-metallic minerals 23 L-DI ML-DI

METAL Metals, metal products 24 L-DI ML-DI

MACHIN∗Machinery, electrical and electronic

products

25-28 H-DI(1) MH-DI(1)

TRANSPEQ Transport equipment 29-30 H-DI H-DI

OTIND∗Other industries 22, 31-32 H-DI(2) MH-DI(2)

COKE Coke & reﬁned petroleum products 19 L-DI ML-DI

ELECGAS Electricity, gas, steam, air conditioning 35 L-DI L-DI

CONSTR Construction 41-43 L-DI L-DI

COMSER∗Commercial & public services 33, 36-39, 45-96 H-DI(3) H-DI(3)

Note: L-DI is low digital intensity, ML-DI is medium-low digital intensity, MH-DI is medium-high digital intensity,

H-DI is high digital intensity.

*Sectors among the 16 selected for which a perfect matching with the DI classiﬁcation was not possible. See Table

3.2 (p. 18) in Horv´at & Webb (2020) for the original classiﬁcation.

(1) MACHIN: 25% ML-DI (ISIC division 25) and 75% MH-DI (ISIC divisions 26-28).

(2) OTIND: 33% ML-DI (ISIC division 22) and 66% MH-DI (ISIC divisions 31 and 32).

(3) COMSER: 19% L-DI (ISIC divisions 36-39, 49-53, 55-56, 68), 8.5% ML-DI (ISIC divisions 85-88), 25.5% MH-DI

(ISIC divisions 33, 45-47, 58-60, 84, 90-93) and 46.8% H-DI (ISIC divisions 61-66, 69-82, 94-96).

The results for this robustness analysis are presented and commented below. Note that the

results for the Conversion eﬀect, as in the main analysis, have been removed and placed in SI

Section S.2 below.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Figure S.1: Boxplots of cumulative (decadal) decomposition results by digital intensity (DI-2)

PRODUCTIVITY

EFFICIENCY

SCALE

ENERGY

0 1 2 3 4 5

DIGITAL INTENSITY LOW HIGH

Note: The decomposition results in this Figure are cumulative and have been aggregated by decade. The yellow

and purple correspond respectively to H-DI and L-DI sectors. The central boxplot line corresponds to the median

value, and the blue diamond to the mean value. The vertical red dashed line sets the threshold between upward

and downward eﬀects. From top to bottom, the factors appear in the following order: Energy, Scale, Eﬃciency, and

Productivity.

Energy demand does not substantially diﬀer when comparing low and high digital intensity (L- &

H-DI) categories (DI-2), despite a very slightly lower mean value for L-DI. This apparent similarity

conceals more pronounced diﬀerences among driving factors and periods. As may be expected, the

average eﬀects are stronger for H-DI for all factors: H-DI sectors have a stronger upward scale eﬀect

and slightly stronger downward eﬃciency and productivity eﬀects. The role played by digitalisation

may thus not be visible directly in the dynamics of energy demand, but rather through its eﬀect on

value added growth.

Looking into diﬀerences across periods, we ﬁnd similar dynamics across both categories, with

subtle diﬀerences. Despite a higher mean values for the entire period, H-DI display reductions in

energy use in the ﬁnal period, alongside a progressive reduction of the scale eﬀect and a slowdown

in the rate of productivity improvements. Eﬃciency improvements however keep accelerating from

2000 onwards. With decelerating value added growth, reductions in energy demand may be achieved

through combined technical improvements, even with productivity gains slowing down. The same

trends can be observed for L-DI, but with overall lower magnitude. Productivity gains were weaker

for L-DI relative to H-DI in the 1970s and 1980s. While both groups converge to similar rates

of improvements in the 2010s, the initial diﬀerence in productivity, combined with reduced value

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

added growth, explains the weaker reductions in energy for L-DI.

Overall, the mixed understanding of energy dynamics looking at DI-2 groups masks the polar-

isation of the eﬀects observed with DI-4, as presented in Section 4.2.2 in the main article. H-DI

and L-DI have stronger growth in energy demand, and this is associated with stronger scale eﬀects

(when considering the median, which moderates the eﬀect of the South Korean MACHIN sector

that drives the mean values up for MH-DI). The intermediate DI-4 categories combine lower scale

eﬀect with either the strongest eﬃciency gains (MH-DI) or strong productivity improvements (ML-

DI), and this results in the lowest growth (or reductions) in energy demand. Considering only two

categories of digital intensive industries does not allow to observe this polarisation, and justiﬁes the

choice of the original DI-4 classiﬁcation, despite its limitations.

Table S.2: Decomposition results by DI-2 category and decade, 1971-2019

Sector Period Energy Scale Eﬃciency Productivity

Avg. Med. Avg. Med. Avg. Med. Avg. Med.

Tot. Tot. 3.82 1.07 30.73 4.24 0.935 0.880 0.560 0.280

L Tot. 2.54 1.04 16.56 3.54 0.955 0.890 0.592 0.283

L 1971-80 1.96 1.21 3.49 2.31 0.957 0.952 0.915 0.582

L 1980-90 1.56 1.05 2.54 2.06 1.02 0.965 0.902 0.535

L 1990-00 1.17 1.03 1.54 1.31 0.965 0.954 0.940 0.856

L 2000-10 1.04 0.913 1.78 1.32 1.01 0.982 0.811 0.681

L 2010-19 1.06 0.964 1.33 1.21 0.990 0.965 0.906 0.817

H Tot. 6.03 1.12 55.08 4.71 0.901 0.870 0.504 0.269

H 1971-80 1.83 1.21 3.41 2.39 1.00 1.02 0.724 0.492

H 1980-90 1.41 1.05 3.30 2.69 1.05 1.03 0.493 0.406

H 1990-00 1.73 1.13 2.23 1.63 0.980 0.956 0.887 0.644

H 2000-10 1.15 1.02 1.61 1.34 0.985 0.973 0.855 0.738

H 2010-19 1.01 0.949 1.28 1.23 0.943 0.935 0.921 0.843

Note: Summary statistics for the full sample (Tot.) correspond to the unweighted cross-country and cross-sector

mean (Avg.) and median (Med.) values of the cumulative decomposition results derived, where cumulative series

are aggregated over the total period. Results by digital intensity category (L-, H-DI) correspond to the unweighted

cross-country average and median values of the (total) cumulative decomposition results. For each DI-2 category, the

summary statistics are derived for the (total) cumulative decomposition results across all the sectors composing the

DI-2 category. Results by DI-2 category and period correspond to the unweighted cross-country average and median

values across sectors for each decade, where cumulative results by decade are obtained by multiplying the results for

all periods in the same decade. For cross-sector comparison, the minimum and maximum values for each factor

have been highlighted.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Figure S.2: Boxplots of cumulative (total) decomposition results by sector

PRODUCTIVITY

EFFICIENCY

SCALE

ENERGY

0 20 40

DIGITAL INTENSITY LOW HIGH

Note: The decomposition results in this Figure are cumulative and have been aggregated over the total period.

The yellow and purple correspond respectively to H-DI and L-DI sectors. The central boxplot line corresponds

to the median value, and the blue diamond to the mean value. The vertical red dashed line sets the threshold

between upward and downward eﬀects. From top to bottom, the factors appear in the following order: Energy,

Scale, Eﬃciency, and Productivity.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

S.2 Results for the Conversion eﬀect

Table S.3: Decomposition results for the Conversion eﬀect for full sample, by sector, by DI categories, and

across time, 1971-2019

Sector

Period

Tot. 1971-80 1980-90 1990-00 2000-10 2010-19

Avg. Med. Avg. Med. Avg. Med. Avg. Med. Avg. Med. Avg. Med.

Tot. 1.01 1.01

Sectors

AGRIL1.00 1.00 0.997 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00

MININGL0.992 1.00 1.00 1.00 1.01 1.00 1.01 1.00 1.00 1.00 0.981 0.997

FOODL1.03 1.02 1.00 1.00 1.01 1.01 1.01 1.00 1.01 1.00 1.01 1.00

ELECGASL1.02 1.00 1.00 1.00 1.01 1.00 1.00 1.00 1.01 1.00 1.01 1.00

CONSTRL0.996 1.00 0.989 1.00 1.00 1.00 0.999 1.00 1.00 1.00 0.998 0.998

TEXTILML 1.02 1.02 0.949 0.999 1.09 1.00 1.01 1.00 0.988 1.00 1.04 1.00

CHEMPHAML 1.04 1.03 1.01 1.00 1.00 1.00 1.02 1.01 1.02 1.01 0.996 1.00

MINERALML 0.999 0.997 0.991 0.997 1.01 1.01 1.00 1.00 0.997 0.996 1.00 0.998

METALM L 1.00 1.01 0.998 0.995 1.02 1.01 1.01 1.00 1.00 1.00 0.995 1.00

COKEML 1.01 1.01 0.998 0.999 1.00 1.00 1.00 1.00 1.01 1.00 0.996 1.00

WOODMH 0.986 0.984 0.972 0.994 0.973 0.989 0.996 0.992 0.999 0.998 0.998 0.995

PAPM H 1.00 0.991 0.991 1.00 0.997 0.999 1.00 1.00 1.00 1.00 0.999 0.999

MACHINMH 1.02 1.01 1.00 1.00 1.01 1.01 1.01 1.00 1.01 0.999 1.00 1.00

OTINDMH 1.04 1.04 1.14 1.18 0.947 0.840 1.03 1.01 1.00 1.00 1.01 1.01

TRANSPEQH1.03 1.02 0.998 1.00 1.01 1.01 1.00 1.00 1.02 1.00 1.00 1.00

COMSERH1.03 1.01 1.01 1.00 1.03 1.01 1.00 1.01 1.01 1.00 0.998 1.00

DI-4

L-DI 1.01 1.00 0.997 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.999 1.00

ML-DI 1.01 1.01 0.985 1.00 1.03 1.00 1.01 1.00 1.00 1.00 1.00 1.00

MH-DI 1.01 1.01 1.02 1.00 0.992 1.00 1.01 1.00 1.00 0.999 1.00 1.00

H-DI 1.03 1.02 1.01 1.00 1.02 1.01 1.00 1.00 1.01 1.00 0.999 1.00

Note: Summary statistics for the full sample (Tot.) correspond to the unweighted cross-country and cross-sector

mean (Avg.) and median (Med.) values of the cumulative decomposition results derived, where cumulative series are

aggregated over the entire period. Results by sector correspond to the unweighted cross-country average and median

values of the (total) cumulative decomposition results. Results by sector and period correspond to the unweighted

cross-country average and median values across sectors for each decade, where cumulative results by decade are

obtained by multiplying the results for all periods in the same decade. Results by digital intensity categories (DI-4)

correspond to the unweighted cross-country average and median values of the (total) cumulative decomposition

results. For each DI category, the summary statistics are derived for the cumulative decomposition results across

all the sectors composing the DI category. Sectors have been ordered by their digital intensity categories, as found

in Table 2, with low digital intensity in Table 3.1, medium-low digital intensity in Table 3.2, medium-high digital

intensity in Table 3.3, and high digital intensity in Table 3.4.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Figure S.3: Boxplots of cumulative (decadal) decomposition results for the Conversion eﬀect by sector and

by DI categories, 1971-2019

0.8 0.9 1.0 1.1 1.2

SECTORS AGRI

MINING FOOD

ELECGAS CONSTR

TEXTIL CHEMPHAR

MINERAL METAL

COKE WOOD

PAP MACHIN

OTIND TRANSPEQ

COMSER

0.8 0.9 1.0 1.1 1.2

DIGITAL INTENSITY LOW MEDIUM−LOW MEDIUM−HIGH HIGH

0.8 0.9 1.0 1.1 1.2

DIGITAL INTENSITY LOW HIGH

Note: The decomposition results in this Figure are cumulative and have been aggregated by decade. The yellow

and purple on the lower plot correspond respectively to H-DI and L-DI sectors. The yellow, green, blue and purple

on the middle plot correspond respectively to H-DI, MH-DI, ML-DI and L-DI sectors. All sectors in the upper plot

are distincted by their colour and are ordered by digital intensity (DI-4) groups, going from L-DI (purple) to H-DI

(yellow). The central boxplot line corresponds to the median value, and the blue diamond to the mean value. The

vertical red dashed line sets the threshold between upward and downward eﬀects.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

Figure S.4: Boxplots of cumulative (total) decomposition results for the Conversion eﬀect by sector and

by DI categories, 1971-2019

0.8 0.9 1.0 1.1 1.2

SECTORS AGRI

MINING FOOD

ELECGAS CONSTR

TEXTIL CHEMPHAR

MINERAL METAL

COKE WOOD

PAP MACHIN

OTIND TRANSPEQ

COMSER

0.8 0.9 1.0 1.1 1.2

DIGITAL INTENSITY LOW MEDIUM−LOW MEDIUM−HIGH HIGH

0.8 0.9 1.0 1.1 1.2

DIGITAL INTENSITY LOW HIGH

Note: The decomposition results in this Figure are cumulative and have been aggregated over the total period.

The yellow and purple on the lower plot correspond respectively to H-DI and L-DI sectors. The yellow, green, blue

and purple on the middle plot correspond respectively to H-DI, MH-DI, ML-DI and L-DI sectors. All sectors in the

upper plot are distincted by their colour and are ordered by digital intensity (DI-4) groups, going from L-DI (purple)

to H-DI (yellow). The central boxplot line corresponds to the median value, and the blue diamond to the mean

value. The vertical red dashed line sets the threshold between upward and downward eﬀects.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5069660

Preprint not peer reviewed

ResearchGate has not been able to resolve any citations for this publication.

AI adoption in America: Who, what, and where

Article

Full-text available

Jan 2024

We study the early adoption and diffusion of five artificial intelligence (AI)‐related technologies (automated‐guided vehicles, machine learning, machine vision, natural language processing, and voice recognition) as documented in the 2018 Annual Business Survey of 850,000 firms across the United States. We find that fewer than 6% of firms used any of the AI‐related technologies we measure, though most very large firms reported at least some AI use. Weighted by employment, average adoption was just over 18%. AI use in production, while varying considerably by industry, was found in every sector of the economy and clustered with emerging technologies, such as cloud computing and robotics. Among dynamic young firms, AI use was highest alongside more‐educated, more‐experienced, and younger owners, including owners motivated by bringing new ideas to market or helping the community. AI adoption was also more common in startups displaying indicators of high‐growth entrepreneurship, including venture capital funding, recent product and process innovation, and growth‐oriented business strategies. Early AI adoption was far from evenly distributed: a handful of “superstar” cities and emerging hubs led startups' adoption of AI. These patterns of early AI use foreshadow economic and social impacts far beyond this limited initial diffusion, with the possibility of a growing “AI divide” if early patterns persist.

CLPFUDatabase: A suite of R packages for energy conversion chain analysis

Article

Full-text available

Jan 2024

Industry 4.0 and energy in manufacturing sectors in China

Article

Full-text available

Dec 2023
RENEW SUST ENERG REV

Decomposition analysis applied to energy and emissions: A literature review

Article

Full-text available

Nov 2023

Decomposition analysis has been widely used to assess the determinants of energy and CO2 emissions in academic research and policy studies. Both the methodology and application of decomposition analysis have been largely improved in the past decades. After more than 50 years’ developments, decomposition studies have become increasingly sophisticated and diversified, and tend to converge internally and integrate with other analytical approaches externally. A good understanding of the literature and state of the art is critical to identify knowledge gaps and formulate future research agenda. To this end, this study presents a literature survey for decomposition analysis applied to energy and emission issues, with a focus on the period of 2016–2021. A review for three individual decomposition techniques is first conducted, followed by a synthesis of emerging trends and features for the decomposition analysis literature as a whole. The findings are expected to direct future research in decomposition analysis.

Earth beyond six of nine planetary boundaries

Article

Full-text available

Sep 2023

This planetary boundaries framework update finds that six of the nine boundaries are transgressed, suggesting that Earth is now well outside of the safe operating space for humanity. Ocean acidification is close to being breached, while aerosol loading regionally exceeds the boundary. Stratospheric ozone levels have slightly recovered. The transgression level has increased for all boundaries earlier identified as overstepped. As primary production drives Earth system biosphere functions, human appropriation of net primary production is proposed as a control variable for functional biosphere integrity. This boundary is also transgressed. Earth system modeling of different levels of the transgression of the climate and land system change boundaries illustrates that these anthropogenic impacts on Earth system must be considered in a systemic context.

The emergence of a ‘twin transition’ scientific knowledge base in the European regions

Article

Jun 2024

This study reveals both the spatial and the combinatory patterns of the digital and green scientific knowledge bases for the creation of new knowledge in the domain of the so-called ‘twin transition’. The recent and rapid diffusion of this twin scientific knowledge in the European regions has not adhered to clearly defined spatial patterns and has been characterised by a dynamic process of actor reconfiguration. However, regions with a strong green and digital science base have a greater propensity to produce more, better quality and more visible twin knowledge. Among the most prevalent twin knowledge fields emerging today are artificial intelligence (AI)- and Internet of Things (IoT)-powered applications for energy storage, distribution and consumption; environmental monitoring and modelling; and urban planning.

A Country-Level Primary-Final-Useful (CL-PFU) energy and exergy database v1.2, 1960-2020

Research

Feb 2024

Selected datasets produced using the {CLPFUDatabase} R package and the International Energy Agencies World Extended Energy Balances to calculate aggregate country-level primary, final, and useful exergy consumption, sectoral final and useful exergy consumption, aggregate primary-final, primary-useful, and final-useful exergy efficiency, and sectoral final-useful exergy efficiency.

Degrowth vs. Green Growth. A computational review and interdisciplinary research agenda

Article

Dec 2023
ECOL ECON

The concepts of Green Growth and Degrowth represent two major narratives in the public and academic debate on socio-ecological transformation. Using a computational literature review based on the Latent Dirichlet Allocation algorithm, we systematically review 1449 journal articles on Green Growth and Degrowth published between 1972 and 2020. Based on a comprehensive full-text analysis, we uncover the hidden topic structure in the literature and identify a total of 50 topics. Using a comparative analysis of the two research fields, we confirm several findings from previous reviews, including that Green Growth research is highly policy oriented, focuses on practical implementation and builds on empirical research methods, while Degrowth research is highly theory-driven, focuses on the analysis of complex human-nature interrelationships and builds on a sound theoretical and conceptual foundation. Our analysis further reveals that Green Growth and Degrowth are two rather isolated fields of research with little exchange and mutual reference on related topics. However, the topic of social change and transition represents an intersecting and promising future research area. We derive four generalizable recommendations that address current shortcomings and interdisciplinary potentials of Degrowth and Green Growth research and contribute to shaping the future of social-ecological transformation research.

Inconsistent definitions of GDP: Implications for estimates of decoupling

Article

Sep 2023
ECOL ECON

Gregor Semieniuk

Efforts to assess the possibilities for decoupling economic growth from resource use and negative environmental impacts have examined their historical relationship, with varying and inconclusive results. This paper shows that ambiguities in the historical measurement arising from definitional changes to GDP are sufficiently large to affect the results. I review the history of structural revisions to GDP using the example of the United States, and on international comparisons of purchasing power parity, and compare decoupling results using GDP vintages reported between 1994 and 2021 for most countries. Between vintages, 10-15% of countries switch between relative decoupling and recoupling from energy or materials on decadal intervals, and up to as many countries as decouple absolutely in an older vintage stop or newly start absolutely decoupling in the newer vintage. GDP vintages also affect environmental Kuznets curve results on absolute decoupling in Grossman and Krueger's seminal paper and accelerate the International Energy Agency's annual global decline in energy intensity by up to −0.2 percentage points. Inconsistencies in economic measurement introduce ambiguity into historical decoupling evidence and model projections into the future. To advance debate, rigorous reporting and sharing of data vintage for subsequent comparison and replication are urgently needed.

Digitalization in Europe: A potential driver of energy efficiency for the twin transition policy strategy

Article