The climate impact of digital technologies remains insufficiently measured. Using global input-output data, we calculate the embodied –i.e., supply chain-related– greenhouse gas emissions of digital industries (hardware, IT services, and communications). We find that the total embodied emissions of digital industries in 2021 are 4.1% of global emissions, with 77–87% being accounted for upstream (i.e., under Scope 3 of the Greenhouse Gas Protocol). We show that 42% of digital emissions are ultimately accounted for within non-digital industries. Hardware accounts for the largest share of digital industries’ embodied emissions, while increasing demand for IT services has driven emissions growth over the past decade. Our findings highlight the need to reduce digital emissions across all industries’ value chains. This includes accounting for embedded digital inputs, adopting circular economy principles in hardware manufacturing, and limiting embodied emissions from IT services, such as artificial intelligence.
To calculate the total intermediate footprint including all digital industries, we have to add up several intermediate inputs from different industries (and countries), for instance, from several software and hardware industries. This is not trivial, as double-counting is a serious problem that occurs when simply adding up all intermediate inputs. For example, when calculating the intermediate input of the software industry in the automobile industry, the emissions embodied in the hardware on which the software runs are already included in the embodied emissions of the software industry. Therefore, if one wants to calculate the joint share of intermediate inputs of two industries in the Leontief column of a third industry, one has to take into account that embodied emissions can be attributed to two different industries if they are in the same value chain. To solve this problem, we subtract the intermediate inputs of industry i from the Leontief inverse as follows:
Then, we add the intermediate inputs of industry k, which are not included in Li, to the elements of Li. This is done by replacing L with L−i and i with k in Equation (7).
Matrix Li,k now contains the inputs of all industries that the industries i and k require to produce input for one unit of industry j.
To add up the intermediate input of all digital industries i,..., m, we iterate the procedure defined by Eqs. (9) and (10) up to industry m:
Moreover, we can verify that adding up all intermediate inputs from all industries of the economy to industry j equals the Leontief column of industry j, i.e. \({l}_{j}^{1,\ldots ,n}={l}_{j}\). In other words, a Leontief column is the sum of all intermediate inputs corrected for double counting. One can simply verify this property by showing that L1,…,n − L = 0.
Finally, total embodied emissions of all digital industries EICT, including intermediate demand in non-digital industries, are fully accounted for in the following matrix:
This matrix is analogous to matrix E of Eq. (4), with the Leontief inverse L replaced by the ICT-adjusted Leontief inverse Li,…,m. This new matrix, denoted by EICT, decomposes the ICT total embodied emissions into elementary contributions of the form \({c}_{i}\,{l}_{ij}^{i,\ldots ,m}\,{y}_{j}\) representing the emissions of the i-th sector that are required by the ICT industry to satisfy the final demand yj of the j-th sector. The sum of row i adds up to the direct emissions of industry i required for its production, while the sum of column j adds up to the embodied emissions of ICT production required by the final demand of industry j, and the sum of all the coefficients adds up to the total embodied emissions TICT, which are always larger than (or can in theory be equal to) FICT. Using the ICT-adjusted Leontief inverse, we obtain T ICT as:
Additionally, industries can be aggregated by simply aggregating the row sums for the contributing emissions, or by simply aggregating column sums for the final demand’s embodied emissions to conveniently present results such as Sankey diagrams. For instance, T ICT − FICT represents 42% of the total embodied emissions of ICT, which are contained in the final demand of all non-digital industries. Moreover, with the full matrix EICT, we are able to analyse the embodied emissions of non-digital industries at a granular, sector-by-sector level.
Our approach leads numerically to the same results as the method of Cabernard et al., which decomposes the total embodied emissions of an arbitrary group of sectors—referred to as ‘target sectors’ in their paper—into a matrix structure similar to our matrix EICT13. However, we introduce the concept of an adjusted Leontief inverse Li,…,m, which is a reduction of the standard Leontief inverse L, designed for the analysis of target sectors (in our case, the ICT sector). Our alternative algorithm has several advantages. For instance, it does not require partitioning the economy into target sectors and the rest of the economy. Moreover, it is particularly well suited for a group of target sectors consisting of a relatively small number of individual industries (as it is the case for the ICT sector) since the proposed approach involves only a limited number of additional computations beyond the calculation of the standard Leontief inverse. Furthermore, the iterative nature of our algorithm allows intermediate embodied emissions from several industries to be aggregated or disaggregated, thereby providing novel insights into the interconnectedness of global value chains.
Also, the advantages of IO analysis over bottom-up approaches include the availability of annually updated, open-access data, which facilitates the analysis of trends over multiple years. Another strength is that the approach attempts to model the technical processes involved in specific products and industries, allowing to comprehensively capturing value chains over the analysed period.
Data
We use FIGARO (Full international and global accounts for research in IO(?)) tables (2023 edition), Eurostat’s annual IO tables covering 64 industries (NACE, rev.2, two digits), 45 countries, including 27 EU member countries, several non-member countries and a rest-of-the-world aggregate (ROW). Moreover, FIGARO distinguishes between five categories of final demand: capital formation, households, government, changes in inventories/assets, and non-profit (see Supplementary Table 1).
FIGARO is an environmentally-extended IO database as it provides air emissions as environmental extensions to the IO trade data(?). The data is derived from air emissions accounts reported by the EU27, Norway, Switzerland, Turkey, UK's Office for National Statistics, and estimates of air emissions by EUROSTAT based on the EDGAR database for the remaining countries.
The data is available for the period between 2010 and 2021. We use the 2023 edition and not the 2024 edition for the main analysis due to methodologies irregularities in the more current version, which are reported by Eurostat (conversation with Eurostat). Also, we currently cannot use the 2025 edition as the matching environmental information has not published yet.
We chose FIGARO primarily because of its recency, higher industry granularity, and reliability. For example, the latest OECD ICIO data extends only to 2020 and aggregates industries into fewer categories, whereas FIGARO offers more detailed sectoral resolution55. FIGARO is also widely used by national statistical offices (e.g. in France and the UK) to calculate consumption-based footprints and is considered in the literature as one of the more robust and reliable multi-regional IO tables.
For a comprehensive description of the data, we refer to Remond-Tiedrez & Rueda-Cantuche56.
