Policy Research Working Paper 10905 Firm Networks and Global Technology Diffusion Paulo Bastos Katherine Stapleton Daria Taglioni Hannah Yi Wei Development Economics Development Research Group September 2024 Policy Research Working Paper 10905 Abstract This study examines the role of multinational firms and Although the locations where these technologies originate global value chain linkages in the cross-country diffusion exhibit a higher prevalence of technology job openings, this of emerging technologies. The analysis combines detailed advantage diminishes over time as diffusion accelerates in information on the near-universe of online job postings wealthier and geographically closer countries and regions. in 17 countries with data on multinational networks and The study highlights the significant role of firm-to-firm link- firm-to-firm linkages from 2014 to 2022. Online job post- ages in technology diffusion, with some linkages proving ings are utilized to investigate how jobs related to emerging more influential than others. Firms that were previously technologies spread through firm networks. The findings buyers or innovation partners of establishments in tech- show that emerging technology jobs are highly concen- nology-originating locations experienced faster growth in trated within multinational firms and their supply chains. jobs related to these technologies. Moreover, relationships Approximately one third of all emerging technology job outside corporate boundaries play a particularly critical role, postings during this period come from Fortune 500 firms, and these connections are influential beyond the factor of their affiliates, buyers, suppliers, or innovation partners. geographical distance. This paper is a product of the Development Research Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at pbastos@worldbank.org, kstapleton@worldbank.org, dtaglioni@worldbank.org, and yi.wei@economics.ox.ac.uk. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Firm networks and global technology diffusion * Paulo Bastos† , Katherine Stapleton† , Daria Taglioni † , Hannah Yi Wei ‡ Keywords: technology diffusion, multinational firms, global value chains, firm-to-firm linkages, disruptive technologies JEL Codes: O33, F23, L14, O14, F14 * The authors thank Juan Jos´e Tapia Montenegro and Juan Porras for research assistance, and Jacob Green- spon for valuable comments. They also gratefully acknowledge generous funding from the Umbrella Facility for Trade which is supported by the governments of The Netherlands, Norway, Switzerland, Sweden and the United Kingdom. † Trade and International Integration, The World Bank. Bastos email: pbastos@worldbank.org. Stapleton email: kstapleton@worldbank.org. Taglioni email: dtaglioni@worldbank.org. ‡ Department of Economics, University of Oxford. Wei email: yi.wei@economics.ox.ac.uk. 1 Introduction The speed with which new technologies spread across nations is key for explaining cross- country differences in economic performance (Comin and Hobijn, 2004, 2010; Comin and Mestieri, 2018). Multinational firms and firm-to-firm relationships through global value chains are believed to play an important role in this process of international technology diffusion (World Bank, 2020). However, understanding the role of firm-to-firm relation- ships in the cross-country spread of new technologies has proved difficult because of the lack of detailed panel data on both the adoption of new technologies across countries and firms, and on different types of international-firm-to-firm linkages. In this paper, we leverage a rich combination of granular information on online job post- ings with data on several types of international firm-to-firm linkages to study the role of multinational firms and global value chains in the cross-country diffusion of new technolo- gies. We examine the diffusion of 29 new technologies across 17 countries and 107,981 firms in the period 2014-2022. Following Bloom et al. (2021), we use the raw text of online job postings and earnings calls data to measure hiring patterns and firm discussions relat- ing to frontier technologies, using this as a proxy for technology diffusion. We combine this information with detailed data on international firm-to-firm linkages, which includes infor- mation on buyer-supplier relationships, innovation partners and affiliates in multinational firm networks. To study the process of technology diffusion, we employ various measures of dispersion used in the literature, such as the Gini coefficient and the coefficient of vari- ation, which allow us to document the intensity and spread of emerging technology jobs across countries, regions, and firms. We first establish four stylized facts about international technology transfer. First, emerging technology hiring has become more dispersed across countries and regions over time. This finding is consistent with evidence on the US only by Bloom et al. (2021). Sec- ond, emerging technology jobs are highly concentrated in multinational firms and their supply chains. Around one third of all job postings related to these 29 frontier technolo- gies in the 17 countries are posted by Fortune 500 companies and their buyers, suppliers, affiliates or innovation partners. Third, technology jobs spread faster to richer regions and countries, a finding consistent with the existing literature on cross-country technology dif- fusion. Finally, technology jobs also spread faster to closer regions and countries to the locations where the technologies emerged. Given the concentration of emerging technology jobs in multinational firms and their supply chains, we next explore the role of pre-existing firm-to-firm linkages in the diffusion of these technologies. For each technology, we use global patents data to identify the regions where the technology emerged, finding that all of these technologies emerged from cities within the USA. We then identify firms that had ex ante linkages to a firm in the 1 city and industry of the technology’s emergence, focusing on the fifteen technologies that emerged after 2005. We evaluate whether firms with such linkages before the takeoff of each technology saw relatively higher growth in job postings related to that technology in the years after the takeoff. We find that ex ante buyer and innovation partner linkages play an important role in the spread of jobs related to these technologies, with less evidence of a role of suppliers or affiliates. In addition, we find that the role of these linkages does not dissipate over time. The faster spread of technology jobs to firms with these pre-existing linkages remains after controlling for distance to the technology emergence location, and we find no evi- dence of a stronger role of these linkages for firms that are geographically proximate to the technology emergence location. We also explore the role of linkages within and outside firm boundaries. We find that the effects for firms with pre-existing buyer and innovation partner linkages are driven by firms that do not share the same ultimate owner. Finally, we show that there is no differential effect of these ex ante buyer and innovation partner linkages for US relative to internationally-linked firms, and the key results are robust to dropping technologies that emerged in Silicon Valley. Our study contributes to several strands of existing research on the diffusion of inno- vation, capabilities, and technology, particularly through firm-to-firm linkages and global value chains. A significant body of literature has examined technology diffusion across countries, often using measures of technology adoption at the country level (Comin and Hobijn, 2004, 2010; Comin and Mestieri, 2018). However, these studies have not exten- sively explored the role of international firm-to-firm linkages in this process. Our research addresses this gap by providing micro-foundational evidence on the importance of these linkages in the diffusion of specific new technologies. Our findings on the positive role of income per capita, and negative role of distance, in the spread of jobs related to emerging technologies, is consistent with this existing literature, while we also highlight the addi- tional important role of firm-to-firm linkages in the cross-country diffusion process. Our approach for identifying new technologies and pioneer firms draws on Bloom et al. (2021), who use textual analysis of patents, job postings, and earnings calls to study interactions between innovation and employment within the US.1 We extend their work by examining technology diffusion across 17 countries and highlighting the role of firm- to-firm linkages in the cross-country diffusion of new technologies. This broader scope allows us to uncover the significant impact of international buyer-supplier and innovation 1 In using online job postings to document technology diffusion, our paper is also related to earlier re- search that studies the relationship between technology and labor markets, including studies on the diffusion of a single specific technology (Autor et al., 2003; Akerman et al., 2015; Acemoglu and Restrepo, 2020; Graetz and Michaels, 2018; Artuc et al., 2023), specific innovations during important historical episodes (Griliches, 1957; Goldin and , 1998; Caprettini and Voth, 2020), and studies examining impacts of technology progress more generally (Krueger, 1993; Berman et al., 1994; Autor et al., 1998; Michaels et al., 2014). 2 partner relationships on technology spread, consistent with findings in Haddad and Harri- son (1993); Aitken and Harrison (1999); Javorcik (2004); Haskel et al. (2007); Keller and Yeaple (2009); Alfaro (2017). Additionally, our paper is related to the extensive theoretical and empirical literature on the effects of multinational activity and global production networks on firm performance and technology diffusion. Previous studies have suggested that multinational corporations (MNCs) facilitate significant technology transfer and R&D activities through their roles in Global Value Chains (GVCs) (Yeaple, 2013; Yoshino and Rangan, 1995; Alfaro et al., 2015). Engagement in GVCs often leads to enhanced skill development since it requires high levels of explicit coordination between firms and hence cooperative learning (Caliendo and Rossi-Hansberg, 2012), and increased demand for high-skill labor (Verhoogen, 2008; Brambilla et al., 2012). Moreover, domestic firms linked to MNCs frequently experience productivity and innovation spillovers (Alfaro-Ure˜ na et al., 2022), and compliance with global standards imposed by MNCs drives technological upgrades (Antr` as and Helpman, 2004; Garcia-Marin and Voigtl¨ ander, 2015). Yet, previous research was unable to measure the diffusion of specific new technologies through firm-to-firm international linkages. Our empirical findings align with these insights, showing that pre-existing relationships with buyers and innovation partners, rather than suppliers or affiliates, are critical for the rapid diffusion of new technologies; and suggesting that relationships with counterparts external to the corporate boundaries are particularly important. This evidence indicates that the nature of the firm-to-firm linkages is crucial, as only specific types of linkages foster the diffusion of emerging technologies. The direct measurement of emerging tech- nologies and different types of firm-to-firm linkages allow us to uncover these channels of technology transfer, thereby deepening our understanding of how multinational networks and GVCs facilitate the transfer of emerging technologies on a global scale. The remainder of this paper is structured as follows. In section 2, we describe the data and the main measures we employ in the empirical analysis. In Section 3, we document stylized facts about the geographic spread of new technologies across countries, and re- gions. We also investigate sector and firm concentration of such technology hiring. In Section 4, we document the role of multinational supply chains in technology diffusion. In doing so, we first investigate the technological advantage of firms located in pioneer loca- tions and its evolution over time and we then study the role of various types of firm-to-firm linkages to pioneer firms in shaping technology diffusion. Section 5 concludes the paper. 3 2 Data and measurement This paper uses data on online job postings from Lightcast to capture jobs related to emerg- ing technologies, and matches this at the firm level with data on firm-to-firm linkages from Factset. In this section, we first describe the online job postings dataset, the classification of new technologies, and the matching process with FactSet firms. We then discuss addi- tional data cleaning and complementary sources. Finally, we provide different measures of technology concentration including the Gini coefficient and the coefficient of variation. 2.1 Online job postings Our primary dataset includes 573 million scraped online job postings from Lightcast (for- merly Burning Glass Technologies). Our sample covers 17 advanced economies from 2014 to 2022.2 We use the job postings data to identify jobs relating to emerging technologies for sev- eral reasons. First, it allows us to identify jobs related to specific and granular technologies over time. In the absence of detailed data on the deployment of multiple technologies across firms, regions, and countries, real-time rich text data sources have been increasingly used as a barometer for the diffusion of new technologies, as reflected through their skills demand footprint or extent of discussion in company boardrooms (see for example Ace- moglu et al. (2022) and Bloom et al. (2021)). These datasets are particularly useful for studying rapid advances in new technologies that have not yet been clearly defined or cap- tured by statistical agencies using traditional data sources. Moreover, Lightcast provides not only the raw text of each job description, but also additional information about the in- dustry, occupation, and location of each job posting, allowing us to trace diffusion through each of these channels. The main potential shortcoming of the Lightcast data on online job postings concerns the representativeness of actual vacancies. First and most notably, it only includes jobs posted online and this subset of vacancies may not be fully representative of all vacancies, while vacancies are also a flow measure of labor demand, rather than a measure of the stock of employment. A growing literature has now shown that online job postings tend to be over-representative of higher-skilled occupations and industries, although recent re- search has generally found that in advanced economies a very high share of jobs are now posted online. For the US, Hershbein and Kahn (2018) show that the share of jobs online as captured by Lightcast is roughly 90% of the jobs in JOLTS. Bloom et al. (2021) also show that the trend in Lightcast postings after 2010 parallels that of JOLTS openings in the US. 2 The sample includes: Austria, Belgium, Denmark, France, Germany, Italy, Luxembourg, Netherlands, Norway, Sweden, Switzerland, United Kingdom, Singapore, Australia, New Zealand, Canada and the United States. 4 For the UK, Javorcik et al. (2019) show it covers just above 85 percent of all (online and of- fline) job vacancies as measured by the UK Vacancy Survey. Bastos et al. (2023), exploiting the same dataset across 35 countries, find a close relationship between total job vacancy rates as measured by official statistical agencies and each country’s ratio of Lightcast job postings to total employment. We complement the above evidence by comparing Lightcast job postings as a share of each country’s working age population, shown in Appendix Table A1. We find that Bel- gium, Switzerland and Luxembourg have the highest ratios of vacancies to working age population, while New Zealand, Norway and Italy have much lower shares. Another limi- tation of the data on online job postings is that they only measure stated but not necessarily realized demand. We therefore note that these measures of high-tech job postings serve as a proxy to track technology diffusion in the labor market, but future work expanding this with measures relating to the hires, separations, wages and internal training might provide a more complete picture. At this point,however, such consistent data is not available for such a wide sample of countries over time. 2.1.1 Identification of new technologies We use the raw text of the job postings to search for the 305 technology keywords from the classification of 29 technologies identified by Bloom et al. (2021).3 We use Google translate to translate these keywords into all of the official languages of the 17 countries. We search for the keywords in both English and all official languages of the country. We then classify each job posting as being related to a specific technology if it includes at least one keyword for that technology. This classification by Bloom et al. (2021) represents one classification of influential innovations that have been discussed prominently by investors and executives of publicly listed companies in the last two decades. Given that 53% of utility patents filed by USPTO are of foreign origin as of 2019 and that one-third of the earning calls were hosted by non US-based firms, these 29 technologies should also be suitable for studies examining diffusion outside the US. Figure 1 lists the number of job postings mentioning each of the 29 technologies. Cloud Computing and Social Networking were mentioned the most, with more than 10 million job postings identified for each of these technologies, followed by Smart Devices then Machine Learning or AI and Online Streaming. 3 For example, “cloud environment”, “cloud solution”, “cloud storage” and “cloud service” all belong to Cloud Computing technology. 5 Figure 1. Job postings by technology Note: This figure displays the total number of job postings for each of the 29 technologies across all 17 countries from 2014Q1 to 2022Q4. 2.1.2 Industry and region classifications We extract the industry code from each Lightcast job posting. As most countries in our sample are in the EU and have the 4-digit ISIC industry classification, we use crosswalks to convert all other national industry classifications to ISIC. This results in 88 industries with highly unequal distribution of technology-related hiring (see Section 3 for details). The crosswalks used are shown in Appendix A2. We adopt a similar standardised procedure for region classifications to ensure consis- tency and comparability of data across the 17 countries. For European countries, we use the Nomenclature of Territorial Units for Statistics (NUTS) 2021 classification to achieve a consistent unit of measurement. For anglophone countries, we group them into states or provinces for the United States, Canada, Australia/New Zealand, ITLS1 for the United Kingdom, and at the country level for Singapore. Our approach results in 156 regions across 17 countries. Appendix A2 provides more details for region level aggregation. 6 2.2 FactSet data on firm-to-firm relationships We use data on firm relationships from FactSet, a private company specializing in market intelligence on firm networks, including inter-firm and intra-firm linkages. FactSet collates data from various sources, including annual reports, investor presentations, press releases, and company websites. Among FactSet’s offerings, the FactSet Fundamentals dataset pro- vides consolidated financial statements and firm information for publicly traded companies, while FactSet Relationships dataset covers supply chain relationships, including suppliers, customers, and partnerships, as well as firm-to-affiliate relations. The FactSet Revere Supply Chain Relationships database is recognized as one of the most comprehensive sources for global firm supply chain information (Huang et al., 2023). This dataset integrates information from official firm filings (10-K, 10-Q, and 8-K reports submitted to the SEC), investor presentations, press releases, and company websites (Fact- Set, 2021).4 FactSet analysts normalize relationship types by cross-referencing a firm’s records with its partners’ (”reverse” links), providing a comprehensive view of supply chain linkages compared to other databases like Compustat Segment (Huang et al., 2023). As of March 2024, FactSet actively monitored over 54,000 ”source” companies, result- ing in a dataset of more than 2 million relationship links involving approximately 360,000 firms. Each link includes information on start and end dates, with historical data available from 2003 and global coverage starting in 2014, expanding to Latin America in 2016.5 The database specifies the nature of each relationship. For our study, we focus on buyer, sup- plier, and innovation partner relationships, supplemented with information on ownership ties and other firm characteristics. At any given time, we observe the parent and ultimate parent firm of a company, and any subsidiaries, if applicable.6 The dataset also includes firm name, industry affiliation, headquarters country, and entity type, which we use in our analysis. Annual sales data, converted to constant 2020 U.S. dollars, are available but not used in our study. We combine the two aforementioned FactSet data modules to construct a final dataset of 169,420 firms with bilateral relationship information from 2012 to 2022. These are defined as de jure corporate institutions, with 27% (46,000 firms) having available information from FactSet Fundamentals for at least one year within the sample period. 4 In financial reporting, a 10-Q is a company’s quarterly report with unaudited financial statements, while the 10-K is an annual report with audited financials. An 8-K updates shareholders on significant unscheduled events. U.S. listed firms must detail any customer accounting for more than 10% of revenue (Gofman and Wu, 2022), with some firms voluntarily disclosing additional customers (Huang et al., 2023). 5 Many Latin American records from 2014 suggest backward updating for these firms. 6 The parent or ultimate parent is identical to the company itself for self-owned firms. 7 2.2.1 Matching Lightcast job postings to Factset We identify 10 million unique Lightcast company names associated with at least one job posting spanning from 2012 to 2022 across 17 countries. The set of company names are extracted from the raw text of all job postings and are not cleaned or standardized, nor cross-referenced with other postings, allowing for several variations in the naming of a single company. To overcome this issue, we draw on the legal names of all entities in the FactSet Relationships dataset, and define a list of name variations to help us matching both Lightcast and FactSet datasets. We take these variations by converting company names to lowercase, removing punctuation and special characters, and eliminating company suffixes (e.g., Ltd., Inc., LLC, etc.). The specific list of variations is in Appendix A4. We then use both exact and fuzzy matching approaches to match the Factset firm names to company names in Lightcast. As a first step, we narrow down both matching approaches to the country where each company’s headquarters are legally registered, according to Fact- Set. Using the different company name variations, we proceed to search for each FactSet company name within Lightcast, delimited by the headquarters’ country. This sequential exact matching allow us to reduce the list of Lightcast company names by eliminating those already exactly matched with FactSet before the fuzzy matching procedure. The additional fuzzy matching approach was considered necessary due to potential variations in spelling and punctuation in company names that could not be captured by the exact matching, lead- ing to inaccurate matching rates. We utilize lowercase company names without punctua- tion marks to conduct the fuzzy matching approach, employing the term frequency-inverse document frequency (TF-IDF) method as the primary fuzzy matching technique, comple- mented by the Levenshtein distance method for additional verification of name matches. Merging the results from both exact and fuzzy matching, we then obtain an initial subset of 80,000 matched FactSet firms out of the total 169,420. As a second step, we expand the FactSet Relationships dataset by integrating ownership and structural information, enabling us to search for affiliates of the 169,420 firms as well. For instance, if the FactSet firm were ‘Microsoft’, using the relationship structure data we can access the entire family of affiliates, including ‘GitHub’, ‘LinkedIn’ and ‘Skype Technolo- gies’. This integration yielded a newly expanded dataset of nearly 2.5 million Factset enti- ties, including parent companies from the original sample. Subsequently, we conduct a sec- ond exact and fuzzy matching exercise searching for this expanded list of company names in the Lightcast data without restricting the matching process to the headquarters country. This procedure allow us to identify 440,000 affiliated entities from FactSet in the Lightcast data, which are associated with 108,000 firms in the FactSet Relationships dataset. After consolidating the outcomes from both steps and removal of duplicate matches7 , our final 7 The duplicate-elimination criteria is detailed in Appendix A4. 8 dataset comprises 109,000 FactSet firms (representing 65% of the 169,420 firms from the supply chain data) linked with nearly 1.5 million company names in Lightcast. To identify the FactSet firms with potential operations in the 17 Lightcast countries, we refined the original sample, keeping only those firms with at least one affiliate in any of the 17 Lightcast countries, according to information reported by FactSet or Lightcast data. For instance, although Samsung’s headquarters may not be in any of the 17 Lightcast countries, FactSet might report it has an affiliate in the US, while Lightcast might include job postings an affiliate in the UK. However, certain FactSet firms, primarily located in Asia, do not have an affiliate in any of the 17 Lightcast countries and are therefore excluded, resulting in a sample of 128,764 firms. Then, we removed every affiliate without location information on the subnational region, leaving us with 126,358 firms. Out of this 126,358 firms, we successfully identified 107,981 in Lightcast, resulting in a final match rate of around 85%. Our final matched dataset is aggregated at what we define as the establishment level, with establishments firm-country-subnational region triplets (subnational regions are typ- ically the highest-level administrative subdivision in each country8 ). Using this definition, the total number of establishments is 1,476,936, belonging to the 107,981 matched firms in the 17 countries. 2.3 Measuring technology pioneer locations and additional datasets 2.3.1 Pioneer regions and firms To measure technology diffusion globally, it is important to identify where the technolo- gies first emerged. For each of the 29 technologies identified, Bloom et al. (2021) define the ‘pioneer’ locations for these technologies as the core-based statistical areas (CBSA)s that collectively accounted for 50% of the global citation-weighted patent grants for that technology during the ten years before its emergence year. While their analysis of pioneer locations is focused only on the US, a robustness check repeating their exercise using global patents data demonstrates that all pioneer locations for these technologies globally are in- deed in the US. All the regions globally with the highest number of patent citations for these technologies are located within the U.S., as shown in Figure 2. We hence proceeded using the US CBSA pioneer locations for each technology. 8 This follows Acemoglu et al. (2020) who define establishments of a firm as the collection of firm job vacancies in each US commuting zone. Sub-national regions are generally defined at Territorial Level 2 (TL2) in OECD Regional Statistics, the ”first government layer after the national or federal one.” https: //www.oecd.org/regional/regional-statistics/geographical-definitions.htm 9 Figure 2. Pioneer locations for 29 technologies using global patents data Note: This figure shows the pioneer locations covering all 29 technologies, identified based on global patent data (not limited to the US). Specifically, the pioneer locations are defined as those which collectively ac- counted for 50% of the global citation-weighted patent grants for that technology during the ten years before its emergence year. To identify pioneer firms for each technology, we first limit firms that are located in the pioneer CBSAs using the zip code of the firms. Next, from this sample we select firms that belong to a pioneer industry based on the 4-digit NAICS code.9 Finally, we label the firm as a pioneer if it was in the pioneer location at least 1 year before the technology emerged, with the year of emergence as defined in Bloom et al. (2021). We identify in total 5031 pioneer firms for the 29 technologies using this method. Figure 3 plots the distribution of pioneer firms for selected new technologies that emerged after 2005.10 9 The pioneer industry is identified by Bloom et al. (2021) following the same approach for pioneer locations. 10 We do not use the pioneer firms as identified in Bloom et al. (2021) due to the small number of firms, which, once matched with the Factset data sample would result in a sample size that is too small. 10 Figure 3. Geographic location of pioneer firms by technology Note: This figure shows the geographic locations of pioneer firms for each technology that emerged after 2005. The size of the red circles is proportional to the average normalized share of technology-related job adverts. 2.3.2 Distance to pioneer regions and firms We are interested in the role that distance plays in technology diffusion. Using the re- gion of each Lightcast job posting, we retrieve the latitude and longitude parameters of the most populous city in every macro-region or US state to build the distance-to-pioneer- region dataset for the 29 new technologies using information from Simplemaps.com and World Cities Database. Likewise, for the distance to pioneer firms, we use the exact latitude/longitude-coordinates from FactSet. 2.3.3 Earnings calls transcripts We also use data from Refinitiv Eikon on 214,685 transcripts from quarterly shareholder earnings calls meetings of publicly listed firms. This type of data has been increasingly used by economists to measure firm uncertainty (e.g. Hassan et al. (2019)), firm sentiment (e.g. Hassan et al. (2020)) and proxies for whether new technologies are ‘disruptive’ in terms of being widely discussed in boardrooms (e.g. Bloom et al. (2021)). Our sample covers 6,525 11 firms in the same 17 countries and years for which there is Lightcast data. Over two-thirds of these earning calls were held by US-based firms (62% percent of all earnings calls since 2014). Figure 4 plots the time series for both the share of Lightcast job postings and the share of earnings calls transcripts mentioning each of the 29 technologies across all 17 countries from 2014 to 2022. We see that growth in technology hiring is highly correlated with growth in earnings calls mentions – the correlation is statistically significant at the 5% level with a coefficient of 0.825. Figure 4. Share of technology-related earnings calls and job postings Note: This figure displays the share of technology related job postings and the share of quarterly earnings calls in all 17 countries, from 2014Q1 to 2022Q4. 2.4 Concentration measures Our key measure of labor-market diffusion is the ratio of technology-related job postings to total job postings at the country, region, and firm levels. As the Lightcast data offers variation at multiple dimensions, we are also interested in the distribution of technology postings across countries, regions and firms. For example, are technology-related postings 12 concentrated within a few countries or more dispersed across countries? Is technology hiring distributed equally across firms? To answer these questions, we follow Goldfarb et al. (2019) and construct the Gini Coefficient at the country, region, and firm levels across the 29 technologies. As a robustness check, we also compute the Coefficient of Variation (CV), which is the measure of dispersion employed by Bloom et al. (2021).11 To account for the fact that different technologies may accumulate different scales of job postings, and that different regions, countries or firms may also have systematically different size of postings, normalization provides a robust measure to gauge the extent to which technology-related hiring for each firm, region and country compares to their overall distribution. For each technology τ and year t, we construct the normalized share of job postings by each region, country, or firm - denoted by i, using the equation below: share jobs exposedi,τ,t Normalized sharei , τ, t = (2) share jobs exposedτ,t 3 Stylized facts about the global dispersion of new tech- nologies 3.1 Technology hiring has become more dispersed across countries and regions over time We begin by establishing the extent to which jobs related to these technologies have spread across countries, by measuring the dispersion of the technology share of job postings across 17 countries and 156 regions, for each of the 29 technologies from 2014 to 2022.12 As described above, we use the Gini Coefficient and CV as our concentration measures. T For each technology T and year t, we regress the concentration measures Yj, t on the years since the emergence of a technology across the country, region and firm dimensions denoted by j, accounting for technology and year fixed effects: T T Yj, t = α + β ∗ Years since emergencet + µT + µt + ϵj, t, T (3) where Years since emergence is defined as the number of years since the year when a tech- 11 The CV for each technology T and year t is calculated as: sd(YtT ) CVtT = (1) mean(YtT ) 12 As described above, we use state as region for the US and Australia, province for the EU and Canada, and city for the UK. There is only one region for Singapore, New Zealand and Luxembourg. 13 nology first appeared in the earnings calls transcripts following the definition by Bloom et al. (2021). Consistent with expectations, hiring related to these technologies has become more dis- persed across countries and regions over the past decade. The negative coefficients (hence the decline in the values of Gini and CV) across both panels, displayed in Figure 1, demon- strate that technology hiring has become more dispersed over time both across countries and country-region pairs. The stronger coefficients for the country-region pair regressions suggests greater diffusion across country-regions over time than across countries alone. In terms of magnitude, for the country-level regression, an increase in 1 year since the emer- gence of the technology is associated with a 4.9% decline in the coefficient of variation, and a 3.4% decline in the Gini. Table 1. Dispersion of technology hiring across countries and regions Coefficient of Variation Gini (1) (2) Panel (a): Across countries Years since emergence -0.0160* -0.0279* (0.0106) (0.0048) Adjusted R-squared 0.8408 0.7933 Observations 232 232 Tech FE Yes Yes Year FE Yes Yes Panel (b): Across country-regions Years since emergence -0.1747*** -0.0872*** (0.0506) (0.0190) Adjusted R-squared 0.8168 0.8365 Observations 232 232 Tech FE Yes Yes Year FE Yes Yes Notes: This table displays regression results of the two concentration measures in columns 1-2 on the years since the technology’s first emergence across 17 countries and 156 regions, controlling for technology and year fixed effects. Standard errors are clustered by technology. The coefficient of variation is defined as the standard deviation of normalized technology share of job adverts divided by the mean. The Gini is defined as the inequality of the distribution of tech share of job adverts, computed using Stata’s ineqdeco command. *** p<0.01, ** p<0.05, * p<0.1 3.2 Technology hiring is highly concentrated in multinational firms and their supply chains We proceed by using the FactSet-Lightcast matched dataset to explore the role of multina- tionals and their supply chains in technology hiring. While our matched dataset of 107,981 firms accounts for only just above 20 percent of all online job postings in these 17 coun- tries, they account for nearly 40 percent of all technology postings. In addition, Fortune 500 firms alone account for around 7 percent of all technology job postings, while Fortune 14 500 and their affiliates, buyers, suppliers and innovation partners account for nearly 30 percent of all technology-related job postings, as is shown in Figure 5. Figure 5. Technology job postings in multinationals and their supply chains 15 3.3 Technology hiring intensity is higher in richer countries Figure 6. Income and technology job posting share in 2022 Note: This figure plots the number of technology-related job postings as a share of total job posting in 2022, and the log GDP per capita in PPP (2015 dollars) in the same year for each of the 17 countries. Consistent with a broad literature on the patterns of technology diffusion (see, e.g, Comin and Mestieri (2018)), we also find that the share of all hiring related to these new tech- nologies is higher in richer countries. Figure 6 shows that countries are far from displaying similar levels of technology hiring. The figure plots the relationship between the technol- ogy share of job adverts and the log of GDP per capita on a PPP basis in 2022, displaying a strong positive correlation. The correlation is statistically significant at the 5% level with a coefficient of 0.6259. The countries with the highest technology-share of job adverts are Singapore, Luxembourg, the US, UK and Canada, which is consistent with the findings of Autor et al. (2006) and Aghion et al. (2018). The relationship between income and technology hiring is further explored in Section 3.4. 16 3.4 Geographic proximity matters for the diffusion of technology jobs We next evaluate the role of distance in shaping technology diffusion across countries and regions. We begin by constructing distance to the pioneer region for each technology and country. As discussed in Section 2.3.2, pioneer locations for each technology are defined as the US CBSAs that jointly accounted for 50 percent of citation-weighted patents during the ten years before the technology emerged. To examine the role of distance from pioneers in technology diffusion, we run the fol- lowing specification: T T T Yc, t = α + βdc, t + γxc, t + µT + µt + ϵc, t, T (4) We use the normalized share of technology-related job postings for technology T in T country c and year t, as the dependent variable, Yc, t . The log of distance to the pioneer region for technology T in country c and year t is represented by dT T c, t . Then xc, t is the additional control measuring the log of income per capita (PPP in 2015 US dollars). We also include year, technology and country-technology fixed effect dummies – µt and µT respectively. Results for the baseline distance regression are displayed in Table 2. The significant and positive coefficient on income suggests that the share of technology-related job postings is positively associated with a country’s income per capita. This is in line with the positive correlation between income per capita and the share of technology-related job adverts discussed above in Section 3.3. The distance coefficients across all specifications are also statistically significant and negative. These results suggest that emerging technology jobs spread faster to closer locations, in line with much of the previous literature on technology diffusion. 17 Table 2. Regression of diffusion on distance, across 17 countries Dep variable: Normalized share of tech-related postings (1) (2) Mean distance from pioneer, logs -0.121** -0.147*** (0.052) (0.050) Income per capita, logs 0.940*** (0.230) Technology FE Yes Yes Year FE Yes Yes Adjusted R-squared 0.067 0.111 Observations 4,176 4,176 Note: This table displays country-technology level regression results of the normalized share of technology job postings for each technology on log of mean distance to pioneer cities for that technology and log of income per capita from 2014 to 2022, controlling for technology and year fixed effects. Mean nearest distance calculated over the regions within the 17 countries. Income per capita is PPP in 2015 US dollors. Robust standard errors are clustered by technology. *** p<0.01, ** p<0.05, * p<0.1. We next run cross-sectional regressions to estimate how the role of distance has changed over time from 2014 to 2022. Additionally, to account for non-linear distance effects, we also include the square of distance in the specification and control for technology fixed effects as is shown below, where Yc,T is the share of job postings for technology T in country c, dc,T and d2 c,T are (log) distance and (log) square of distance for technology T in country T, and xc is the log of income, for each country c: Yc,T = α + βdc,T + θd2 c,T + γxc,T + µT + ϵc,T (5) We plot the distance coefficients over time following specification (5), shown in Figure 7. In line with our findings so far, we observe that the coefficient on distance has been fairly steady over time, although became larger in 2022. The coefficient on income per capita is larger in magnitude and has a higher statistical significance in most years, although the coefficient dropped in magnitude and significance in 2022. 18 Figure 7. Distance and income effects from 2014 to 2022 2 1.5 1 Coefficient .5 0 -.5 14 15 16 17 18 19 20 21 22 20 20 20 20 20 20 20 20 20 Mean distance from pioneer, logs Income per capita, logs Note: This figure displays the coefficients on distance and income per capita (and corresponding standard errors at the 10% level) from 2014 to 2022 following the cross-sectional regression in equation 5. These regressions include all 17 countries. For comparison, and considering that all pioneer locations are in the US, we proceed by exploring the distance effect within the US alone. We apply the same specification as equation (4), for each of the 29 technologies across 51 US states from 2014 to 2022. Again, while there is evidence of a negative association between distance from pioneer locations and technology job postings within the US, the distance effect is weaker than the role of income per capita. Finally, we explore the role of distance to pioneer cities for each technology and firm. In particular, for each technology, we build the nearest distance of each firm to the tech- nology’s pioneer city, controlling for technology and year fixed effects. We also control for each firm’s sales and country fixed effects for an additional robustness check. Our results, shown in Table 4, are consistent with those from the previous country-level analysis. Across columns (1) to (4), we observe significant and negative distance coefficients and positive income effects. 19 Table 3. Distance and income effect on technology hiring within the US Dep variable: Normalized share of tech-related postings (1) (2) Mean distance from pioneer, logs -0.004** -0.004*** (0.000) (0.000) Income per capita, logs 0.012*** (0.004) Technology FE Yes Yes Year FE Yes Yes Adjusted R-squared 0.079 0.083 Observations 13,311 13,311 Notes: This table displays regression results of the share of technology-related job posting on the (log) distance and (log) median household income in the US, for technology T in state s and year t, controlling for technology, year and state-technology fixed effects. The median household income for each state is measured in 2022 US dollars with data the Federal Reserve Bank of St. Louis. Standard errors are clustered by technology. *** p<0.01, ** p<0.05, * p<0.1 Table 4. Distance and income effect on technology hiring, across 107,981 global firms Dep variable: Normalized share of tech-related postings (1) (2) (3) (4) Nearest distance from pioneer city, logs -0.075*** -0.104*** -0.087*** -0.100*** (0.016) (0.021) (0.019) (0.026) Firm sales in USD, logs 0.034** 0.030** (0.013) (0.013) Technology FE Yes Yes Yes Yes Year FE Yes Yes Yes Yes Country FE Yes Yes Adjusted R-squared 0.001 0.002 0.001 0.001 Observations 18,709,716 18,709,716 3,907,761 3,907,761 Notes: Standard errors clustered at the tech level in parentheses. The normalized shared is calculated by share jobs exposedi,τ,t Normalized sharei,τ,t = . The normalized share is capped at the 99th percentile of share jobs exposedτ,t non-zero observations. The specification excludes observations before the year of emergence of a technology. 4 Diffusion through multinational supply chains The results reported in the previous section showed that emerging technology jobs are dis- proportionately concentrated in Fortune 500 firms and their supply chains. In this section, we examine empirically the role of various types of firm-to-firm linkages in the spread of new technologies. We find that firms with pre-existing supply chain linkages to pioneer firms before technologies emerged see faster technology diffusion, with external buyers 20 and innovation partners being the most important types of linkages for technology diffu- sion. 4.1 Empirical specification In this sub-section we outline the baseline empirical specifications for examining the con- centration of technology hiring in pioneer firms and how it dissipates over time, and the role of pre-existing supply chain linkages to pioneer firms in technology diffusion. For ev- ery specification within this section, we aggregate the matched dataset of establishments to the firm level as we have supply chain linkages data only at the firm level, resulting in a final firm-technology-year panel dataset.13 4.1.1 Concentration of technology hiring in pioneer locations over time Following our definition of pioneer firms in Section 2.3.1, we define the dummy variable T P ioneeri, t as taking the value of 1 if firm i is a pioneer firm for technology T in year t, and 0 otherwise. For each technology T, firm i and year t, we run the following regression with firm, technology and year fixed effects: T T T T Yi, t = α + β1 P ioneeri, t + β2 (P ioneer ∗ years since emergence)i, t + µT + θt + δi + ϵi, t (6) T whereYi, t measures technology diffusion, proxied by the normalized share of technology- T related job postings. Apart from the dummy P ioneeri, t , we also interact the variable with the number of years since a technology emerged to test for the persistence of this hiring concentration in pioneer firms. 4.1.2 The role of supply chain linkages to pioneer firms We proceed by estimating a series of specifications that allow us to examine the role played by ex ante firm-to-firm linkages to pioneer firms in technology diffusion. In particular, we examine whether being linked to pioneer firms prior to the technology’s emergence i.e., whether being a supplier, buyer, innovation partner, or affiliate of a pioneer firm, results in faster diffusion of technology jobs after the technology emerges. Table 5 shows that in total 43% of entities in the Lightcast-FactSet matched dataset are connected to at least one firm in a pioneer city for at least one technology through supply chain relationships in the two years before the technology’s emergence. Pioneer firms for 13 Unlike the firm-level data from Factset, the establishment-level data derived from Lightcast does not contain sufficient information on exact location, industry and foundation year, which is needed to identify pioneers. 21 Machine Learning, in particular, have the highest share of connected firms through the supply chain network, shown by Table 6. Table 5. Pioneer supply chain linkages summary statistics Global US Matched Firm Type Number Percentage Matched Firm Type Number Percentage Total 107981 100% Total 43534 100% Pioneer firm 24261 22.5% Pioneer firm 24261 56% Supplier 7806 7.2% Supplier 4846 20.0% Buyer 6202 5.7% Buyer 4020 16.6% Affiliate 18709 17.3% Affiliate 18240 75.2% Innovation partner 4815 4.5% Innovation partner 3052 12.6% Competitor 9109 8.4% Competitor 6169 25.4% Notes: The table on the left displays the number and percentage of firms in the pioneer cities, and the sup- pliers, buyers, affiliates, innovation partners and competitors that are linked to them, as a result of matching the Lightcast data with FactSet, across 17 countries from 2014 to 2022. The table on the right lists the same variables for the US alone. Table 6. Percentage of all firms linked to a pioneer firm by technology and network Single linkage Multiple linkages Innovation Technology Buyers Suppliers Affiliates 2 linkages 3 linkages 4 linkages partners Autonomous Cars 0.55% 0.21% 0.13% 0.13% 0.11% 0.03% 0.00% Bispecific Monoclal Antibody 0.64% 0.33% 0.42% 0.59% 0.27% 0.18% 0.03% Cloud Computing 0.53% 1.25% 0.62% 0.72% 0.36% 0.22% 0.03% Computer Vision 0.81% 1.68% 1.03% 0.99% 0.55% 0.33% 0.08% 3D Printing 0.44% 0.68% 0.15% 0.45% 0.24% 0.12% 0.01% Extremely High Frequency 1.00% 1.36% 0.45% 0.88% 0.53% 0.29% 0.04% Fingerprint Sensor 0.78% 1.76% 0.92% 0.99% 0.51% 0.33% 0.07% Fracking 0.07% 0.10% 0.06% 0.04% 0.03% 0.00% 0.00% Hybrid Vehicle\Electric Car 0.20% 0.03% 0.02% 0.01% 0.01% 0.00% 0.00% Machine Learning\AI 1.17% 2.52% 1.64% 1.29% 0.73% 0.46% 0.06% Mobile Payment 0.61% 0.99% 0.49% 0.63% 0.35% 0.22% 0.04% Smart Devices 0.40% 0.66% 0.47% 0.42% 0.24% 0.13% 0.02% Social Networking 0.44% 0.85% 0.41% 0.49% 0.25% 0.16% 0.02% Virtual Reality 1.02% 2.25% 1.20% 1.32% 0.69% 0.43% 0.09% Wireless Charging 0.84% 1.36% 0.39% 0.97% 0.51% 0.31% 0.05% Notes: The table displays the percentage of firms that are linked to a pioneer firm for each technology that emerged since 2005 through supply chain relations. Percentage is calculated over the total number of firms (107,981). Multiple linkages are allowed since firms can have multiple networks with pioneer firms at the same time and for the same technology (e.g., buyer and supplier; supplier, buyer and innovation partner; etc.). Since the Factset dataset begins only in 2003, we restrict the supply chain analysis to 15 technologies that emerged since 2005, and use the 2 years prior to the year of emergence of each technology to capture pre-existing firm-to-firm linkages.14 Similarly to the pioneer 14 The resulting 15 technologies are: Social Networking, Cloud Computing, Smart Devices, Machine Learn- 22 regressions, we control for firm, year and technology fixed effects. We consider several alternative specifications, including first the dummy variables for buyer, supplier, affiliate and partner individually, and then testing them jointly in various combinations. For exam- ple, for the pioneer and buyer’s regression, we have the following equation: T T T Yi, t = α + β1 P ioneeri, t + β2 (P ioneer ∗ years since emergence)i, t T T +γ1 Buyeri, t + γ2 (Buyer ∗ years since emergence)i, t (7) +µT + θt + δi + ϵT i, t T where Buyeri, t takes the value of 1 if the firm was initially a buyer of a pioneer firm for technology t and 0 otherwise. All the remaining variables have the meaning defined above. We adopt similar specifications for the other types of pre-existing linkages individually, as well as a specification that considers all types of linkages at the same time. 4.2 Results 4.2.1 Diffusion from pioneer firms Table 7 reveals that pioneer firms see a statistically significantly higher normalized share of job postings for the technology for which they are pioneers, as shown by the positive and significant coefficients of the pioneer dummy variable in columns (1) and (2). Moreover, in column (2) we observe that this higher share of postings for the technology declines over time, indicative of the spread of technology hiring to firms in different industries and locations. ing/AI, Hybrid Vehicle Electric Car, Virtual Reality, Autonomous Cars, Computer Vision, 3D Printing, Mobile Payment, Fracking, Extremely High Frequency, Bispecific Monoclonal Antibody, Wireless Charging, Finger- print Censor. 23 Table 7. Pioneer firms and technology diffusion Dep variable: Normalized share of technology-related job adverts (1) (2) Pioneer 3.131*** 4.630*** (0.226) (0.398) Pioneer × years since emergence -0.202*** (0.036) Technology FE Yes Yes Year FE Yes Yes Firms FE Yes Yes Adjusted R-squared 0.040 0.041 Observations 10,072,709 10,072,709 Notes: Standard errors clustered at the firm level in parentheses. The normalized shared is cal- share jobs exposedi,τ,t culated by Normalized sharei,τ,t = . The normalized share is capped at the share jobs exposedτ,t 99th percentile of non-zero observations. The specification excludes observations before the year of emergence of a technology. 4.2.2 Faster diffusion for firms with supply chain linkages We now present the results relating to the role of various types of pre-existing linkages to pioneer firms in shaping the spread of new technologies. The results in columns (2) to (4) of Table 8 show that firms that were buyers, suppliers or innovation partners of pioneers before the technologies emerged subsequently observe greater hiring related to these tech- nologies after their emergence. There is less evidence of a decline in the importance of these linkages over time, as reflected by the weak significance of the negative coefficients of the linkage dummies interacted with years since emergence. The results in column (5) examine the extent to which there is also greater diffusion among firms that were affiliates of pioneer firms. In this case the coefficient of interest is also positive but not statistically significant. Since the same entity can simultaneously have different types of pre-existing linkages with pioneer firms, in column (6) we include simultaneously all these linkage variables. The results reveal that the diffusion of new technologies occurs mainly through buyers and innovation partners of pioneer firms. The coefficients associated with these variables remain positive and statistically significant. In contrast, the coefficient on suppliers of pioneers is now close to zero and no longer statistically significant. Overall, these results suggest that, when accounting for a wide range of pre-existing firm-to-firm linkages, the diffusion of new technologies through global value chains is mainly observed among buyers and innovation partners of pioneer firms. 24 Table 8. Supply chain linkages and their role in technology diffusion Dep variable: Normalized share of technology-related job adverts (1) (2) (3) (4) (5) (6) Pioneer 4.630*** 4.468*** 4.574*** 4.474*** 3.928*** 3.893*** (0.398) (0.395) (0.396) (0.395) (0.774) (0.775) Pioneer × years since emergence -0.202*** -0.191*** -0.199*** -0.193*** -0.186*** -0.184** (0.036) (0.036) (0.036) (0.036) (0.072) (0.072) Linkages of pioneer firms Buyer 1.878*** 1.490*** (0.300) (0.319) Supplier 0.702*** 0.074 (0.173) (0.170) Innovation partner 1.658*** 1.056*** (0.275) (0.291) Affiliate 0.791 0.564 (0.771) (0.770) Linkages of pioneer firms over time Buyer × years since emergence -0.052* -0.053* (0.029) (0.030) Supplier × years since emergence -0.006 0.025 (0.018) (0.018) Innovation partner × years since emergence -0.037 -0.018 (0.026) (0.027) Affiliate × years since emergence -0.014 -0.001 (0.072) (0.072) Technology FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Firms FE Yes Yes Yes Yes Yes Yes Adjusted R-squared 0.041 0.041 0.041 0.041 0.041 0.041 Observations 10,072,709 10,072,709 10,072,709 10,072,709 10,072,709 10,072,709 Notes: Standard errors clustered at the firm level in parentheses.The normalized share is capped at the 99th percentile of non-zero observations. *** p<0.01, ** p<0.05, * p<0.1 4.2.3 Diffusion through supply chain linkages controlling for distance A key question from the results above is whether firms that are more geographically prox- imate to pioneer firms are more likely to have supply chain linkages. To test whether the results are driven by distance, for each firm-technology pair we also consider the distance to pioneer firms for that technology as an additional control, given the negative and sig- nificant distance coefficients in 2.3.2. Results from Table 9 show not only negative and significant coefficients on distance across all specifications, but also consistent linkage co- efficients with similar significance level to those in the baseline, with buyer and innovation partner linkages remaining the most important. This suggest that firm-to-firm linkages matter above and beyond the role of distance. 25 Table 9. Diffusion through supply chain linkages controlling for distance Dep variable: Normalized share of tech-related job adverts (1) (2) (3) (4) (5) (6) Pioneer 4.496*** 4.336*** 4.440*** 4.341*** 3.800*** 3.769*** (0.398) (0.395) (0.396) (0.395) (0.777) (0.777) Pioneer × years since emergence -0.204*** -0.193*** -0.201*** -0.196*** -0.189*** -0.187*** (0.036) (0.036) (0.036) (0.036) (0.072) (0.072) Nearest distance from pioneer city, logs -0.071*** -0.070*** -0.071*** -0.071*** -0.071*** -0.070*** (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) Linkages of pioneer firms Buyer 1.869*** 1.480*** (0.300) (0.319) Supplier 0.701*** 0.074 (0.173) (0.170) Innovation partner 1.656*** 1.056*** (0.275) (0.291) Affiliate 0.783 0.557 (0.774) (0.773) Linkages of pioneer firms over time Buyer × years since emergence -0.053* -0.054* (0.029) (0.030) Supplier × years since emergence -0.006 0.024 (0.018) (0.018) Innovation partner × years since emergence -0.037 -0.018 (0.026) (0.027) Affiliate × years since emergence -0.014 -0.001 (0.072) (0.072) Technology FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Firms FE Yes Yes Yes Yes Yes Yes Adjusted R-squared 0.041 0.041 0.041 0.041 0.041 0.041 Observations 9,834,969 9,834,969 9,834,969 9,834,969 9,834,969 9,834,969 Notes: Standard errors clustered at the firm level in parentheses.The normalized share is capped at the 99th percentile of non-zero observations. *** p<0.01, ** p<0.05, * p<0.1 4.2.4 Linkages within and outside firm boundaries We also examine whether the patterns of technology hiring documented in the previous section are mainly driven by linkages within or outside firm boundaries. In particular, we separate buyers, suppliers and innovation partners that belong to the same ultimate owner as the pioneer firms (defined as those within the firm boundary), from those whose ultimate owner differs from the pioneer’s (defined as those outside the firm boundary). Only under 1 percent of our linked firms are within the firm boundary, while the vast majority are outside the firm boundary. The results shown in Table 10 reveal that the role of firm-to- firm linkages is driven by effects for firms outside the firm boundary, while linkages within the firm boundary are not statistically significant. 26 Table 10. Diffusion through supply chain linkages within and outside firm boundaries Dep variable: Normalized share of technology-related job adverts (1) (2) (3) (4) (5) Pioneer 4.630*** 4.448*** 4.533*** 4.453*** 4.362*** (0.398) (0.394) (0.391) (0.394) (0.390) Pioneer × years since emergence -0.202*** -0.190*** -0.196*** -0.193*** -0.187*** (0.036) (0.036) (0.035) (0.036) (0.035) Linkages of pioneer firms Buyer within org. 3.201 0.158 (3.630) (3.841) Buyer outside org. 1.868*** 1.470*** (0.299) (0.319) Supplier within org. 6.733 4.383 (4.736) (6.414) Supplier outside org. 0.686*** 0.076 (0.171) (0.170) Innovation partner within org. 4.241 2.141 (2.874) (3.455) Innovation partner outside org. 1.617*** 1.023*** (0.274) (0.292) Linkages of pioneer firms over time Buyer within org. × years since emergence 0.146 0.217 (0.266) (0.266) Buyer outside org.× years since emergence -0.054* -0.054* (0.029) (0.030) Supplier within org. × years since emergence -0.376 -0.404 (0.307) (0.488) Supplier outside org. × years since emergence -0.003 0.025 (0.018) (0.018) Innovation partner within org. × years since emergence -0.019 0.089 (0.190) (0.299) Innovation partner outside org. × years since emergence -0.034 -0.015 (0.026) (0.027) Technology FE Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Firms FE Yes Yes Yes Yes Yes Adjusted R-squared 0.041 0.041 0.041 0.041 0.041 Observations 10,072,709 10,072,709 10,072,709 10,072,709 10,072,709 Notes: Standard errors clustered at the firm level in parentheses. The normalized share is capped at the 99th percentile of non-zero observations. *** p<0.01, ** p<0.05, * p<0.1 4.2.5 Considering multiple linkages Our analysis so far has focused on firms with a single type of linkage, whether as buyers, suppliers, innovation partners, or affiliates of a pioneer firm. Next, we expand the scope to examine firms with multiple linkages - i.e. firms that are both a supplier and a buyer, or an innovation partner and a supplier, or an innovation partner and a supplier of a pioneer firm. Based on our findings that results are driven primarily by external linkages, this section focuses on firms with external linkages. The estimates in columns (1)-(5) of Table 11 indicate that all linkages combinations considered are positively associated with the normalized share of tech-related job postings. We further observe that the combinations of buyers-suppliers, supplier and innovation partner, and innovation partners-buyers display positive and significant coefficients when controlling for other linkage combinations. 27 Table 11. Technology diffusion and multiple supply chain linkages outside firm boundaries Dep variable: Normalized share of technology-related job adverts (1) (2) (3) (4) (5) (6) Pioneer 4.630*** 4.479*** 4.501*** 4.476*** 4.487*** 4.427*** (0.398) (0.393) (0.394) (0.395) (0.394) (0.393) Pioneer × years since emergence -0.202*** -0.193*** -0.195*** -0.194*** -0.194*** -0.191*** (0.036) (0.035) (0.035) (0.036) (0.035) (0.035) Linkages of pioneer firms outside firm boundary Buyer and supplier 2.311*** 1.257* (0.475) (0.763) Supplier and innovation partner 1.947*** 1.029** (0.371) (0.442) Innovation partner and buyer 2.350*** 1.402* (0.457) (0.728) Innovation partner and buyer and supplier 2.717*** -0.713 (0.581) (1.272) Linkages of pioneer firms outside firm boundary over time Buyer and supplier × years since emerg. -0.055 -0.073 (0.048) (0.065) Supplier and innovation partner × years since emerg. -0.032 -0.006 (0.037) (0.041) Innovation partner and buyer× years since emerg. -0.040 -0.015 (0.046) (0.072) Innovation partner and buyer and supplier × years since emerg. -0.043 0.051 (0.060) (0.122) Technology FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Firms FE Yes Yes Yes Yes Yes Yes Adjusted R-squared 0.041 0.041 0.041 0.041 0.041 0.041 Observations 10,072,709 10,072,709 10,072,709 10,072,709 10,072,709 10,072,709 Notes: Standard errors clustered at the firm level in parentheses.The normalized share is capped at the 99th percentile of non-zero observations. *** p<0.01, ** p<0.05, * p<0.1 4.3 Robustness checks 4.3.1 Studying international linkages only A central objective of the analysis above was to document the role of pre-existing firm- to-firm linkages to pioneer firms in shaping the cross-country spread of new technologies. Given that almost 40% of matched firms in our dataset are based in the US, and that all identified pioneer cities are also within the US, one potential concern about the baseline results is whether they are driven by firm-to-firm linkages within the US (and not by inter- national linkages). In this section, we first explore whether there is a differential effect for firms within and outside the US. We take the baseline specifications defined above and add further interaction terms with a US dummy for each supplier, buyer, innovation partner and affiliate indicator. Con- sider the pioneer and buyer’s regression again, for example, we now adopt the following specification: T T T Yi, t = α + β1 P ioneeri, t + β2 (P ioneer ∗ years)i, t T T +γ1 Buyeri, t + γ2 (Buyer ∗ years)i, t (8) T T +λ1 U S ∗ Buyeri, t + λ2 (U S ∗ Buyer ∗ years)i, t +µT + θt + δi + ϵT i, t 28 We also adopt analogous specifications for the other types of linkages. The results from these specifications are reported in Table 12. Overall, we do not observe a differential effect for supply-chain-linked firms within the US relative to those in other countries. Once we add these interaction terms, the coefficients for buyers and innovation partners remain positive and statistically significant in Column (5) as before, while the coefficients on the interaction terms are not statistically significant. This evidence therefore suggests that our results are not driven solely by firm-to-firm linkages within the US, but hold equally for the impacts of linkages across borders. Table 12. Firm-to-firm linkages within and outside the US Dep variable: Normalized share of technology-related job adverts (1) (2) (3) (4) (5) (6) Pioneer 4.630*** 4.487*** 4.569*** 4.486*** 3.920*** 3.903*** (0.398) (0.394) (0.394) (0.394) (0.852) (0.853) Pioneer × years since emergence -0.202*** -0.192*** -0.200*** -0.195*** -0.177** -0.177** (0.036) (0.035) (0.035) (0.036) (0.079) (0.079) Linkages of pioneer firms Buyer 2.312*** 1.857*** (0.568) (0.622) Buyer × US -0.637 -0.544 (0.666) (0.723) Supplier 0.610** -0.098 (0.242) (0.303) Supplier × US 0.159 0.281 (0.335) (0.364) Innovation partner 1.833*** 1.245** (0.434) (0.496) Innovation partner × US -0.260 -0.286 (0.557) (0.613) Affiliate 0.600 0.500 (0.674) (0.679) Affiliate × US 0.203 0.076 (1.103) (1.105) Linkages of pioneer firms over time Buyer × years since emergence -0.061 -0.041 (0.048) (0.054) Buyer × years since emergence × US 0.015 -0.015 (0.059) (0.064) Supplier × years since emergence -0.029 0.009 (0.023) (0.028) Supplier × years since emergence × US 0.036 0.023 (0.034) (0.036) Innovation partner × years since emergence -0.072** -0.051 (0.033) (0.041) Innovation partner × years since emergence × US 0.053 0.049 (0.048) (0.054) Affiliate × years since emergence 0.111* 0.118* (0.065) (0.065) Affiliate × years since emergence × US -0.135 -0.131 (0.104) (0.104) Technology FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Firms FE Yes Yes Yes Yes Yes Yes Adjusted R-squared 0.041 0.041 0.041 0.041 0.041 0.041 Observations 10,072,709 10,072,709 10,072,709 10,072,709 10,072,709 10,072,709 Notes: Standard errors clustered at the firm level in parentheses.The normalized share is capped at the 99th percentile of non-zero observations. *** p<0.01, ** p<0.05, * p<0.1 29 4.3.2 Excluding pioneers from Silicon Valley Another potential concern about the results from the baseline analysis is that these pio- neers might exhibit specific features, both with regard to their innovative capacity and the network to which they belong. For example, one might be concerned that a few leading technology firms in Silicon Valley might be driving the overall patterns observed in the data, and that the results for these firms might not extend to those of other innovative re- gions or technologies. To address this concern, in Table 13 we exclude pioneer firms located within Silicon Valley from our linkage sample. We observe that the empirical results remain similar to those in the baseline analysis, suggesting that our findings apply more broadly to other pioneer regions as well. In particular, the coefficients on pre-existing buyers and innovation partners of pioneers in Column (5) remain positive and statistically significant. The main qualitative difference relative to the baseline analysis is that the coefficient on affiliates in Column (5) is now also positive and statistically significant. Table 13. Firm-to-firm linkages excluding Silicon Valley pioneers Dep variable: Normalized share of technology-related job adverts (1) (2) (3) (4) (5) (6) Pioneer 4.269*** 4.135*** 4.223*** 4.140*** 2.761*** 2.743*** (0.420) (0.417) (0.418) (0.418) (0.482) (0.483) Pioneer × years since emergence -0.177*** -0.168*** -0.175*** -0.170*** -0.093* -0.093* (0.038) (0.038) (0.038) (0.038) (0.050) (0.050) Linkages of pioneer firms Buyer 1.921*** 1.539*** (0.299) (0.319) Supplier 0.690*** 0.065 (0.168) (0.166) Innovation partner 1.654*** 1.038*** (0.268) (0.284) Affiliate 1.722*** 1.523*** (0.519) (0.519) Linkages of pioneer firms over time Buyer × years since emergence -0.052* -0.052* (0.029) (0.030) Supplier × years since emergence -0.003 0.027 (0.018) (0.018) Innovation partner × years since emergence -0.037 -0.019 (0.026) (0.026) Affiliate × years since emergence -0.092* -0.081 (0.053) (0.053) Technology FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Firms FE Yes Yes Yes Yes Yes Yes Adjusted R-squared 0.040 0.040 0.040 0.040 0.040 0.040 Observations 10,057,442 10,057,442 10,057,442 10,057,442 10,057,442 10,057,442 Notes: Standard errors clustered at the firm level in parentheses. The normalized share is capped at the 99th percentile of non-zero observations. *** p<0.01, ** p<0.05, * p<0.1 30 4.3.3 Interacting linkages with affiliates of pioneer firms The results above show no significant effect for firms affiliated with pioneer firms. This lack of significance is evident both in the baseline regression (Table 8) and when controlling for distance (Table 9). To further understand this null affiliate estimate, we also interact all other linkage variables with an affiliate dummy. Consistent with earlier results, all linkage effects lose significance when interacted with the affiliate dummy, as shown in Table 14. This further solidifies the evidence that it is linkages outside the firm boundary that are important for the spread of new technologies. Table 14. Single linkage interacted with affiliate dummy Dep variable: Normalized share of technology-related job adverts (1) (2) (3) (4) (5) Pioneer 4.630*** 3.883*** 3.908*** 3.923*** 3.878*** (0.398) (0.774) (0.775) (0.776) (0.775) Pioneer × years since emergence -0.202*** -0.185** -0.187*** -0.188*** -0.185*** (0.036) (0.072) (0.072) (0.072) (0.072) Linkages of pioneer firms Buyer 1.731*** 1.416*** (0.291) (0.311) Buyer × affiliate 1.855 0.297 (1.693) (2.337) Supplier 0.545*** -0.027 (0.156) (0.161) Supplier × affiliate 3.222* 2.765 (1.809) (2.041) Innovation partner 1.517*** 1.031*** (0.258) (0.273) Innovation partner × affiliate 1.632 -0.439 (1.594) (2.234) Linkages of pioneer firms over time Buyer × years since emergence -0.048 -0.051* (0.029) (0.030) Buyer × years since emergence × affiliate -0.069 0.044 (0.137) (0.179) Supplier × years since emergence 0.003 0.031* (0.017) (0.018) Supplier × years since emergence × affiliate -0.189 -0.205 (0.156) (0.163) Innovation partner × years since emergence -0.030 -0.017 (0.026) (0.026) Innovation partner × years since emergence × affiliate -0.092 0.018 (0.129) (0.169) Technology FE Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Firms FE Yes Yes Yes Yes Yes Adjusted R-squared 0.041 0.041 0.041 0.041 0.041 Observations 10,072,709 10,072,709 10,072,709 10,072,709 10,072,709 Notes: Standard errors clustered at the firm level in parentheses.The normalized share is capped at the 99th percentile of non-zero observations. *** p<0.01, ** p<0.05, * p<0.1 31 4.3.4 Interacting linkages with distance While we have controlled for distance separately in Table 9, as an additional robustness check, we now also include both the linkage dummies and their interaction with distance and time. Table 15 shows the linkage coefficients and the linkage dummies interacted with both distance and the emergence year for each technology. In line with our previous findings, the coefficients for all linkage types (except for affiliates) remain positive and statistically significant across all specifications in Columns (2) to (5). These findings prevail in column (6) when controlling for other types of linkages, with the exception of innovation partners, whose coefficient remains positive, but loses precision. Meanwhile, there is no differential effect of linkages for firms by distance to the pioneer location in columns (2) to (6), except for suppliers, which now see a positive linkage effect that is diminishing with distance. 32 Table 15. Interacting linkages with distance Dep variable: Normalized share of technology-related job adverts (1) (2) (3) (4) (5) (6) Pioneer 4.723*** 4.561*** 4.596*** 4.556*** 3.822*** 3.800*** (0.412) (0.402) (0.402) (0.402) (0.921) (0.921) Pioneer × distance -0.420** -0.445** -0.419** -0.440** -0.331 -0.363* (0.209) (0.210) (0.209) (0.209) (0.215) (0.215) Pioneer × distance × years since emergence 0.047** 0.049** 0.047** 0.048** 0.035* 0.037* (0.020) (0.020) (0.020) (0.020) (0.020) (0.020) Linkages of pioneer firms Buyer 2.132*** 1.550** (0.626) (0.702) Supplier 1.581*** 0.893** (0.477) (0.442) Innovation partner 1.926*** 0.704 (0.610) (0.717) Affiliate 1.023 0.761 (0.975) (0.970) Linkages of pioneer firms in space Buyer × distance -0.041 -0.011 (0.094) (0.104) Supplier × distance -0.139** -0.130** (0.064) (0.063) Innovation partner × distance -0.041 0.060 (0.087) (0.103) Affiliate × distance -0.109 -0.104 (0.130) (0.130) Linkages of pioneer firms in space over time Buyer × distance × years since emergence 0.001 0.002 (0.008) (0.009) Supplier × distance × years since emergence 0.001 0.001 (0.006) (0.006) Innovation partner × distance × years since emergence -0.003 -0.006 (0.008) (0.009) Affiliate × distance × years since emergence 0.013 0.014 (0.012) (0.012) Technology FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Firms FE Yes Yes Yes Yes Yes Yes Adjusted R-squared 0.041 0.041 0.041 0.041 0.041 0.041 Observations 9,834,969 9,834,969 9,834,969 9,834,969 9,834,969 9,834,969 Notes: Standard errors clustered at the firm level in parentheses. The normalized share is capped at the 99th percentile of non-zero observations. *** p<0.01, ** p<0.05, * p<0.1. Distance refers to the distance to the nearest pioneer city, in logs. 4.3.5 Technologies that emerged after the Lightcast data start date The analysis above considered technologies that emerged after 2005, such that their link- ages could be identified through the Factset data. However, our Lightcast panel data of job postings begins from 2014 only. Therefore there are several technologies for which in the early years of emergence there is no data in Lightcast. Our coefficients on the years since emergence variables may be disproportionately stemming from the impacts during the later years after emergence. To consider whether this is the case, we also run a robust- ness check keeping only the technologies that emerged from 2014 onwards, for which there is Lightcast data for all the years directly after emergence. This results in a significantly 33 diminished sample to only 3 very new technologies: machine learning, autonomous cars and extremely high frequency. The estimates based on this restricted sample are presented in Table 16. Power is lost in this sample and so coefficients on the linkages variables are weaker and only the buyer coefficient remains statistically significant. However, the results remain qualitatively similar to those in the baseline analysis, with pioneer locations seeing higher technology postings and this advantage diminishing over time. This suggests that the early years after emergence are similar to the later years in terms of the extent to which technologies diffuse from pioneer firms. Table 16. Keeping technologies that emerged after 2014 Dep variable: Normalized share (1) (2) (3) (4) (5) (6) Pioneer 7.949*** 7.949*** 7.990*** 7.974*** 8.445*** 8.442*** (0.860) (0.861) (0.860) (0.861) (1.927) (1.927) Pioneer × since emergence -0.732*** -0.755*** -0.744*** -0.753*** -0.939*** -0.945*** (0.145) (0.145) (0.145) (0.145) (0.320) (0.320) Linkages of pioneer firms Buyer 0.942 1.106* (0.593) (0.648) Supplier -0.434 -0.487 (0.342) (0.326) Innovation partner 0.068 0.146 (0.561) (0.595) Affiliate -0.556 -0.535 (1.823) (1.823) Linkages of pioneer firms over time Buyer × since emergence 0.228** 0.160* (0.089) (0.095) Supplier × since emergence 0.130** 0.035 (0.053) (0.044) Innovation partner × since emergence 0.240** 0.137 (0.096) (0.098) Affiliate × since emergence 0.236 0.207 (0.304) (0.304) Technology FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Firms FE Yes Yes Yes Yes Yes Yes Adjusted R-squared 0.143 0.143 0.143 0.143 0.143 0.143 Observations 1,963,325 1,963,325 1,963,325 1,963,325 1,963,325 1,963,325 Notes: Standard errors clustered at the firm level in parentheses.The normalized share is capped at the 99th percentile of non-zero observations. *** p<0.01, ** p<0.05, * p<0.1 34 5 Conclusion Cross country differences in economic development are strongly influenced by the extent to which new technologies diffuse across nations. While it is often hypothesized that multina- tional corporations and international firm-to-firm relationships through global value chains play an important role in this process of international technology diffusion, understand- ing their role in the spread of new technologies has proved difficult because of the lack of detailed panel data on both the adoption of new technologies and various types of international-firm-to-firm linkages. In this paper, we have assembled an unusually rich combination of granular information on online job postings with data on several types of international firm-to-firm linkages to study the role of multinational firms and global value chains in the cross-country diffusion of new technologies. Following a growing literature, we used the online job postings data to construct proxies for the spread of new technologies. We then examined the spread of 29 new technologies across 17 countries and 107,981 firms over the period 2014-2022. We combined this information with detailed panel data on international firm-to-firm link- ages, which includes information on buyer-supplier relationships, innovation partner and affiliate. This made it possible to examine, for the first time, the role of these linkages in shaping the intensity and spread of technologies across countries, regions, and firms. We first established the following stylized facts about the geographic and firm-to-firm diffusion of new technologies. First, we find that technology hiring has become more dispersed across countries and regions over time. Second, while technology hiring has disseminated more across firms over time, it remains highly clustered in multinational firms and their value chains. Third, the share of technology hiring is still significantly higher in richer countries. Fourth, there is evidence that proximity to invention sites (pioneer cities) contributes to accelerate technology diffusion across countries and regions. We have then explicitly examined the role of different types of firm-to-firm linkages in the cross-country diffusion of new technologies. The results point to a dual trend. On the one hand, there is a strong technological advantage for firms located in pioneer cities in the US, as revealed by the fact that technology postings tend to be concentrated in these firms for several years. However, we also document that the technology advantage of these firms tends to diminish over time. On the other hand, we provide evidence that firm-to-firm linkages play an important role in the international diffusion of technologies. In particular, the results reveal that buyers and innovation partners of pioneer firms experience the most significant technological benefits. In contrast, being an affiliate of a pioneer firm does not emerge as a statistically significant driver of the diffusion of new technologies across countries. Once we control for firm ownership, we find that external firm relationships drive the positive results. 35 References Acemoglu, Daron and Pascual Restrepo, “Robots and jobs: Evidence from US labor mar- kets,” Journal of Political Economy, 2020, 128, 2188–2244. , David Autor, Jonathon Hazell, and Pascual Restrepo, “AI and jobs: Evidence from online vacancies,” Technical Report, National Bureau of Economic Research 2020. , , , and , “Artificial Intelligence and Jobs: Evidence from Online Vacancies,” Journal of Labor Economics, 2022, 40 (S1), 293–340. Aghion, Philippe, Ufuk Akcigit, Antonin Bergeaud, Richard Blundell, and David He- mous, “Innovation and Top Income Inequality,” The Review of Economic Studies, 06 2018, 86 (1), 1–45. Aitken, Brian J. and Ann E. Harrison, “Do Domestic Firms Benefit from Direct Foreign Investment? Evidence from Venezuela,” American Economic Review, 1999, 89, 605–618. Akerman, Anders, Ingvil Gaarder, and Magne Mogstad, “The skill complementarity of broadband internet,” Quarterly Journal of Economics, 2015, 130, 1781–1824. Alfaro, L., P. Antras, D. Chor, and P. Conconi, “Internalizing Global Value Chains: A Firm-Level Analysis,” National Bureau of Economic Research, Inc, 2015, (21582). Alfaro, Laura, “Gains from Foreign Direct Investment: Macro and Micro Approaches,” World Bank Economic Review, 2017, 30, S2–S15. na, A., I. Manelici, and J. P. Vasquez, “The Effects of Joining Multinational Alfaro-Ure˜ Supply Chains: New Evidence from Firm-to-Firm Linkages,” The Quarterly Journal of Economics, 2022, 137 (3), 1495–1552. as, P. and E. Helpman, “Global Sourcing,” Journal of Political Economy, 2004, 112 Antr` (3), 552–580. Artuc, Erhan, Paulo Bastos, and Bob Rijkers, “Robots, tasks, and trade,” Journal of Inter- national Economics, 2023, 145, 103828. Autor, David H., Frank Levy, and Richard J. Murnane, “The skill content of recent tech- nological change: An empirical exploration,” Quarterly Journal of Economics, 2003, 118, 1279–1334. Autor, David, Lawrence Katz, and Alan Krueger, “Computing inequality: Have comput- ers changed the labor market?,” Quarterly Journal of Economics, 1998, 113, 1169–1213. , , and Melissa Kearney, “The Polarization of the U.S. Labor Market,” American Eco- nomic Review, May 2006, 96 (2), 189–194. Bastos, Paulo, Jacob Greenspon, Katherine Stapleton, and Daria Taglioni, “Did the 2022 global energy crisis accelerate the diffusion of low-carbon technologies?,” 2023. 36 Berman, Eli, John Bound, and Zvi Griliches, “Changes in the demand for skilled labor within U.S. manufacturing: Evidence from the Annual Survey of Manufacturers,” Quar- terly Journal of Economics, 1994, 109, 367–397. Bloom, Nicholas, Tarek Alexander Hassan, Aakash Kalyani, Josh Lerner, and Ahmed Tahoun, “The Diffusion of New Technologies,” Working Paper 28999, National Bureau of Economic Research July 2021. Brambilla, I., D. Lederman, and G. Porto, “Exports, Export Destinations, and Skills,” American Economic Review, 2012, 102 (7), 3406–3438. Caliendo, L. and E. Rossi-Hansberg, “The Impact of Trade on Organization and Produc- tivity,” The Quarterly Journal of Economics, 2012, 127 (3), 1393–1467. Caprettini, Bruno and Hans-Joachim Voth, “Rage against the machines: Labor-saving technology and unrest in industrializing England,” American Economic Review: Insights, 2020, 2, 305–320. Comin, Diego and Bart Hobijn, “Cross-country technology adoption: Making the theories face the facts.,” Journal of Monetary Economics, 2004, 51, 39–83. and , “An exploration of technology diffusion,” American Economic Review, 2010, 100, 2031–59. and Marti Mestieri, “If technology has arrived everywhere, why has income diverged?,” American Economic Journal: Macroeconomics, 2018, 10 (3), 137–178. FactSet, “Supply Chain Relationships - Data and Methodology Guide,” 2021. https://www.factset.com/marketplace/catalog/product/ factset-supply-chain-relationships, documentation accessed on 2023-02-06. , “FactSet Supply Chain Relationships [Dataset],” 2023. https://www.factset. com/marketplace/catalog/product/factset-supply-chain-relationships, data ac- cessed from server on: 2024-03-20. Garcia-Marin, A. and N. Voigtl¨ ander, “Exporting and Plant-Level Efficiency Gains: It’s in the Measure,” Journal of Political Economy, 2015. Forthcoming. Gofman, Michael and Youchang Wu, “Trade credit and profitability in production net- works,” Journal of Financial Economics, 2022, 143 (1), 593–618. Goldfarb, Avi, Bledi Taska, and Florenta Teodoridis, “Could Machine Learning Be a General-Purpose Technology? Evidence from Online Job Postings,” SSRN Electronic Jour- nal, 2019. Goldin, Claudia D. and Lawrence F. Katz (1998), “The origins of technology-skill com- plementarity.,” Quarterly Journal of Economics, 1998, 113, 683–732. Graetz, Georg and Guy Michaels, “Robots at work,” Review of Economics and Statistics, 2018, 100 (5), 753–768. 37 Griliches, Zvi, “Hybrid corn: An exploration in the economics of technological change,” Econometrica, 1957, 25, 501–22. Haddad, Mona and Ann Harrison, “Are There Positive Spillovers from Direct Foreign Investment?: Evidence from Panel Data for Morocco,” Journal of Development Economics, 1993, 42, 51–74. Haskel, Jonathan E., Sonia C. Pereira, and Matthew J. Slaughter, “Does Inward Foreign Direct Investment Boost the Productivity of Domestic Firms?,” Review of Economics and Statistics, 2007, 89, 482–496. Hassan, Tarek A, Stephan Hollander, Laurence van Lent, and Ahmed Tahoun, “Firm- Level Political Risk: Measurement and Effects*,” The Quarterly Journal of Economics, 08 2019, 134 (4), 2135–2202. Hassan, Tarek Alexander, Stephan Hollander, Laurence van Lent, and Ahmed Tahoun, “The Global Impact of Brexit Uncertainty,” January 2020, (26609). Hershbein, Brad and Lisa B. Kahn, “Do Recessions Accelerate Routine-Biased Technolog- ical Change? Evidence from Vacancy Postings,” American Economic Review, July 2018, 108 (7), 1737–72. Huang, Yi, Chen Lin, Sibo Liu, and Heiwai Tang, “Trade networks and firm value: Evidence from the US-China trade war,” Journal of International Economics, 2023, p. 103811. Javorcik, Beata, ben Kett, Katherine Stapleton, and Layla O’Kane, “Unravelling Trade Integration: Local Labour Market Effects of the Brexit Vote,” CEPR Discussion Papers, 2019, (No. DP14222). Javorcik, Beata S., “Does Foreign Direct Investment Increase the Productivity of Domestic Firms? In Search of Spillovers through Backward Linkages,” American Economic Review, 2004, 94, 605–627. Keller, Wolfgang and Stephen R. Yeaple, “Multinational Enterprises, International Trade, and Productivity Growth: Firm-Level Evidence from the United States,” Review of Eco- nomics and Statistics, 2009, 91, 821–831. Krueger, Alan B., “How computers have changed the wage structure: Evidence from mi- crodata, 1984-1989,” Quarterly Journal of Economics, 1993, 108, 33–60. Michaels, Guy, Ashwini Natraj, and John Van Reenen, “Has ICT polarized skill demand? Evidence from eleven countries over 25 years,” Review of Economics and Statistics, 2014, 96, 60–77. Verhoogen, Eric, “Trade, Quality Upgrading, and Wage Inequality in the Mexican Manu- facturign Sector,” Quarterly Journal of Economics, 2008, 123, 489–530. World Bank, “Trading for Development in the Age of Global Value Chains,” World Develop- ment Report, 2020. 38 Yeaple, Stephen Ross, “The Multinational Firm,” Annual Review of Economics, 2013, 5, 193–217. Yoshino, Michael Y. and U. Srinivasa Rangan, Strategic Alliances: Co-operation and Com- petition, Boston, MA: Harvard Business School Press, 1995. 39 6 Appendix A1 Lightcast data coverage Figure A1. Lightcast job adverts data as a share of working age population by country Note: The bottom figure shows the total number of Lightcast online job postings for 17 countries from 2014 to 2022: Belgium, Switzerland, Luxembourg, Germany, United Kingdom, Sweden, the Netherlands, France, Singapore, Austria, United States, Denmark, Australia, Canada, New Zealand, Norway and Italy. The top panel shows the ratio of Lightcast job postings to the working age population (extracted from the International Labour Organisation) for each country. 40 Table A1. Lightcast job adverts by country, from 2014 to 2022 Country Code Start quarter Technology postings Total postings Austria AT 2014Q1 266951 5946319 Australia AUS 2014Q1 417346 5261358 Belgium BE 2014Q1 928509 18379669 Canada CAN 2014Q1 833773 9027658 Switzerland CH 2016Q1 752679 13112599 Germany DE 2014Q1 2921739 82051043 Denmark DK 2014Q1 192309 3591168 France FR 2014Q1 3108584 54427353 Italy IT 2014Q1 256944 10129354 Luxembourg LU 2014Q1 97669 860942 The Netherlands NL 2014Q1 827063 13713021 Norway NO 2014Q1 38947 1015097 New Zealand NZL 2014Q1 69767 1136948 Sweden SE 2014Q1 223996 5191277 Singapore SGP 2014Q1 1034600 3419776 United Kingdom UK 2014Q1 4855523 40936420 United States US 2014Q1 23646858 225977968 Note: This table aggregates the number of technology-related job postings and total Lightcast job adverts by country from 2014 to 2022. 41 A2 Additional Data Processing To assess the relevance of disruptive technologies in the labor market, we conduct a tex- tual analysis of job adverts provided by Lightcast, following the method used by Bloom at al. (2021). The Lightcast dataset offers broad coverage of job postings across countries. However, due to its multilingual nature, we have to match the keywords/bigrams in each country’s official language. This involves translating the Lightcast scripts and the technical bigrams in different languages to English for cross-country comparison. Our data clean- ing involves identifying the number of job adverts mentioning disruptive technologies (we count a posting being 1 even if the keywords are mentioned multiple times), resulting in a unique dataset for each country. The information from each country is then disaggregated using available data on year, industry, region, and firm, allowing us to gain insights about the demand for specific technologies-related skills across different dimensions. Furthermore, we adopted a standardized approach to aggregate the information of job postings at the geographic level to ensure consistency and comparability of data across 17 countries, given the heterogeneity of the geographical information. For European coun- tries, we utilized the Nomenclature of Territorial Units for Statistics (NUTS) 2021 classifi- cation15 to achieve a consistent unit of measurement. While in most European countries we aggregated the data at the NUTS1 level (major socio-economic regions), we had to use the NUTS0 level (country as a whole) or Switzerland, Denmark, Luxembourg, and Norway since these countries don’t have a NUTS1 classification. On the other hand, for anglophone countries, we grouped them into states or provinces for the United States, Canada, Aus- tralia/New Zealand, ITLS1 for the United Kingdom, and at the country level for Singapore. Our approach thus ensures a standardized unit of measurement across countries and al- lows for meaningful comparison of data. Table A2 below lists the region level aggregation for each of the 17 Lightcast countries. 15 For more info about the NUTS Classification, see: https://ec.europa.eu/eurostat/web/nuts 42 Table A2. Regional level equivalences across countries BGT Countries Region Level Austria NUTS1 - 2021 Belgium NUTS1 - 2021 Switzerland Country (NUTS0 - 2021) GermanY NUTS1 - 2021 Denmark Country (NUTS0 - 2021) France NUTS1 - 2021 Italy NUTS1 - 2021 Luxembourg Country (NUTS0 - 2021) Netherlands NUTS1 - 2021 Norway Country (NUTS0 - 2021) Sweden NUTS1 - 2021 United States State and Capitol Territory United Kingdom NUTS1 / ITLS1 Australia/New Zealand State and Capitol Territory Canada Provinces Singapore Country 43 A3 Technologies and Keywords Table A3. Top Technical Bigram in Job Adverts and Earning Transcripts Bigram Technology Job Posting Earning Call social media Social Networking 9302561 9902 smart phone Smart Devices 3568299 12067 saas Cloud Computing 2552279 6807 mobile devices Smart Devices 2409096 6858 machine learning Machine Learning AI 2070152 5087 cloud based Cloud Computing 1923934 7994 cloud computing Cloud Computing 1242289 2176 cloud service Cloud Computing 1153677 3878 live stream Online Streaming 1149664 559 cloud platform Cloud Computing 992014 2256 enterprise applications Cloud Computing 987698 710 search engine Search Engine 924375 1887 artificial intelligence Machine Learning AI 862863 5471 social network Social Networking 725431 0 cloud infrastructure Cloud Computing 673007 1534 wifi Wifi 659055 2279 cloud environments Cloud Computing 649078 1389 cloud solution Cloud Computing 616987 2404 paas Cloud Computing 596042 273 public cloud Cloud Computing 589913 2292 iaas Cloud Computing 509521 142 cloud storage Cloud Computing 473538 444 social networking Social Networking 453572 507 video game Electronic Gaming 438451 827 cloud security Cloud Computing 402859 619 video conferencing Online Streaming 383378 748 hybrid cloud Cloud Computing 302718 1367 wireless networks Wifi 300529 1164 deep learning Machine Learning AI 295306 548 enterprise network Cloud Computing 286231 336 Note: This table presents the top 30 technical bigrams ordered by the number of job postings. The technical bigrams are adopted from those identified by Bloom et al. (2021) using patent citations and earnings calls transcripts. 44 Table A4. Total Job Postings Compared with Bloom et al. (2021) Global Diffusion Bloom et al. (US only) Technology Job Posting Technology Job Posting Social Networking 11004493 Cloud computing 3684901 Cloud Computing 10822170 Social Networking 3457390 Smart Devices 7026763 Smart devices 2376510 Machine Learning Ai 3187308 Machine Learning AI 679776 Online Streaming 2943758 Search Engine 535784 Wifi 982736 Online streaming 487731 Search Engine 767763 Wi-Fi 388844 Solar Power 675212 Electronic gaming 247201 Electronic Gaming 425334 Solar Power 201296 Hybrid Vehicle Electric Car 424437 Hybrid vehicle/Electric car 118550 Virtual Reality 356612 Touch Screen 109538 Autonomous Cars 334835 Rfid 80894 Computer Vision 312653 Computer vision 76350 D Printing 242706 GPS 65922 Lithium Battery 235658 Mobile payment 65482 Rfid Tags 203742 Virtual Reality 61102 Gps 132183 3d printing 57904 Mobile Payment 128069 Autonomous Cars 52974 Touch Screen 68299 Lane departure warning 32107 Oled Display 53467 Lithium battery 16926 Lane Departure Warning 36467 Software defined radio 14187 Software Defined Radio 33065 Drug conjugates 10603 Drug Conjugates 27928 Fracking 8966 Fracking 18360 Millimeter wave 6161 Extremely High Frequency 11274 Oled display 5528 Bispecific Monoclonal Antibody 9521 Bispecific monoclonal antibody 2702 Wireless Charging 4850 Wireless charging 1649 Stent Graft 2158 Stent graft 1270 Fingerprint Sensor 1436 Fingerprint sensor 711 Year coverage: 2014Q1 - 2022Q4. Year coverage: 2007-2019. Note: This table lists the number of job postings associated with each of the 29 technologies across the 17 countries our dataset covers from 2014Q1 to 2022Q4 (left) and for the US alone (right) from 2007 to 2019 as presented by Bloom at al. (2021). Comparing with tech postings in the US, both the top 4 and the last 4 technologies follow the same ranking. 45 Figure A2. Share of Job Postings and Earning Call Transcripts for the 29 Technologies, Global (left) vs the US (right) Note: The bar graph plots highly similar shares of job postings and earning call transcripts aggregating all technology-related counts from 17 countries from 2014 to 2022, comparing to the shares within the US. Both graphs are ranked in descending order by the share of Lightcast postings. 46 A4 Lightcast-FactSet matching procedure Prior to the matching process, we established four distinct variations of the company names present in both the FactSet and Lightcast datasets, taking into account any possible dif- ferences that could hinder company identification. Subsequently, we conducted an exact matching analysis using these different name variations in a predetermined order. Specifi- cally, we first matched company names using Variation N°1, followed by Variation N°2, and so on. It’s worth noting that we removed previously matched companies from the Lightcast dataset before moving on to the next variation to avoid matching the same company twice. Table A5. Variations of company names to conduct matching Variation Description 1 Lower-case company name. No changes or deletion of any special character or symbol. 2 Lower-case company name. Punctuation marks and special symbols dropped. 3 Punctuation marks-free, lower-case company name. Company suffixes (Ltd., Corp., LLC, etc.) removed (whole word and abbreviation). 4 Punctuation marks-free, lower-case company name. International company suffixes (AG, B.V., Oy, etc.) removed (whole word and abbreviation). Additionally, we also considered that a fuzzy matching was necessary even after con- ducting the exact matching because of variations in spelling or punctuation that could cause an exact match to fail, leading to an incomplete or inaccurate match. Thus, to en- sure a more thorough and comprehensive analysis, a fuzzy matching was performed using company names with Variation N°2, utilizing the TF-IDF and Levenshtein distance methods to account for such variations in the text. Once the results of the two matching processes were obtained and put together, we dropped the duplicated matches from the combined results based on entity characteristics from FactSet using the following criteria: 1. Keep the FactSet Entity that is incorporated in the same country as the Lightcast country, if there is only one of these among all of the possible FactSet ID values for the duplicated Lightcast firm-country pair. 2. For still unidentified duplicates, we keep the FactSet Entity of a duplicated company with the highest average annual sales (if info available). 3. For still unidentified duplicates, we keep the FactSet Entity of a duplicated company with the oldest founding year. 4. For still unidentified duplicates, we keep the FactSet Entity that is not a subsidiary. 47 5. If after all of these steps, we still couldn’t identify the correct duplicate, we proceed to identify those duplicated Lightcast firm-country pairs that belong to the same en- tity from the relationship dataset (i.e., the parent company), and create an artificial affiliate that captures the duplicated entities within a country. For example, if there is a ‘Google Germany’ company name in Lightcast, and it has a match with ‘Google Germany Ltd.’ and ‘Google Germany Inc.’ in FactSet, we create the artificial affiliate as ‘Google Germany’ in FactSet, and keep that instead of ‘Google Germany Ltd.’ and ‘Google Germany Inc.’. 6. We drop every duplicate we cannot identify using this criteria (≈ 1% of the dataset). 48 A5 Pioneer location and industry by technology Table A6. Pioneer location and industry by technology Technology Pioneer Location Pioneer Industry 3d printing Riverside-San Bernardino-Ontario, CA Industrial Machinery Manufacturing bispecific monoclonal antibody Boston-Cambridge-Newton, MA-NH Pharmaceutical and Medicine Manufacturing cloud computing San Francisco-Oakland-Hayward, CA Computer and Peripheral Equipment Manu- facturing computer vision Boston-Cambridge-Newton, MA-NH Navigational, Measuring, Electromedical, and Control Instruments Manufacturing drug conjugates Philadelphia-Camden-Wilmington, PA-NJ-DE-MD Pharmaceutical and Medicine Manufacturing electronic gaming Las Vegas-Henderson-Paradise, NV Motion Picture and Video Industries extremely high frequency Denver-Aurora-Lakewood, CO Aerospace Product and Parts Manufacturing fingerprint sensor Reno, NV Semiconductor and Other Electronic Compo- nent Manufacturing fracking Houston-The Woodlands-Sugar Land, TX Utility System Construction gps Grand Rapids-Wyoming, MI Communications Equipment Manufacturing hybrid vehicle electric car Santa Cruz-Watsonville, CA Motor Vehicle Manufacturing lithium battery Atlanta-Sandy Springs-Roswell, GA Medical Equipment and Supplies Manufac- turing machine learning ai San Francisco-Oakland-Hayward, CA Computing Infrastructure Providers, Data Processing, Web Hosting, and Related Ser- vices mobile payment San Jose-Sunnyvale-Santa Clara, CA Computer Systems Design and Related Ser- vices oled display Trenton, NJ Motor Vehicle Parts Manufacturing online streaming Portland-Vancouver-Hillsboro, OR-WA Motion Picture and Video Industries rfid tags Detroit-Warren-Dearborn, MI Motor Vehicle Parts Manufacturing search engine New York-Newark-Jersey City, NY-NJ-PA Computer and Peripheral Equipment Manu- facturing smart devices Dallas-Fort Worth-Arlington, TX Computer and Peripheral Equipment Manu- facturing social networking Boston-Cambridge-Newton, MA-NH Motion Picture and Video Industries software defined radio Denver-Aurora-Lakewood, CO Communications Equipment Manufacturing solar power Chicago-Naperville-Elgin, IL-IN-WI Glass and Glass Product Manufacturing stent graft Minneapolis-St. Paul-Bloomington, MN-WI Medical Equipment and Supplies Manufac- turing touch screen Washington-Arlington-Alexandria, DC-VA-MD-WV Audio and Video Equipment Manufacturing virtual reality Seattle-Tacoma-Bellevue, WA Computer and Peripheral Equipment Manu- facturing wifi Los Angeles-Long Beach-Anaheim, CA Computer and Peripheral Equipment Manu- facturing wireless charging Miami-Fort Lauderdale-West Palm Beach, FL Computer and Peripheral Equipment Manu- facturing autonomous cars Boston-Cambridge-Newton, MA-NH Computing Infrastructure Providers, Data Processing, Web Hosting, and Related Ser- vices lane departure warning Chicago-Naperville-Elgin, IL-IN-WI Motor Vehicle Manufacturing Note: This table lists the pioneer location and pioneer industry for each of the 29 technologies, identified by Bloom et al. (2021). For technologies that have more than one pioneer location or industry, we list the first that appeared in the data for illustration. 49