Cloud adaptiveness within industry sectors – Measurement and observations

Cloud computing combines established computing technologies and outsourcing advantages into a new ICT paradigm that is generally expected to foster productivity and economic growth. However, despite a series of studies on the drivers of cloud adoption, evidence of its economic effects is lacking, possibly because many of the datasets on cloud computing are of insufficient size and often lack a time dimension as well as precise definitions of cloud computing, thus making them unsuitable for rigorous quantitative analysis. To overcome these limitations, we propose a proxy variable for cloud computing usage — cloud adaptiveness — based on survey panel data from European firms. Observations based on a descriptive analysis suggest three important aspects for further research. First, cloud studies should be conducted at the industry level as cloud computing adaptiveness differs widely across industry sectors. Second, it is important to know what firms do with cloud computing to understand the economic mechanisms and effects triggered by this innovation. And third, cloud adaptiveness is potentially correlated to a firm ’ s position in the supply chain and thus the type of output it produces as well as the market in which it operates. Our indicator can be employed to further analyze the effects of cloud computing in the context of firm heterogeneity.


Introduction
Although mention of "the Cloud," or cloud computing, is now ubiquitous in daily life, our understanding of what it actually is and how it changes private and corporate structures is surprisingly limited. Most people recognize cloud computing as a fairly recent development in information and communication technology (ICT). However, the wide range of opinions of what constitutes cloud computing and how it affects households and enterprises is a partial reflection of the many different uses of cloud computing and the resulting lack of a universally accepted and understood definition of it.
Beginning several decades ago, advances in processor and related technologies and the spread of the personal computer, as well as server structures and communication infrastructure like the Internet, helped automatize production and supply chains and facilitate management and administration. ICT as a whole was expected to have a great influence on the productivity of industries and economies. Indeed, as suggested by Cardona, Kretschmer, and Strobel (2013), ICT has some of the hallmarks of enabling or general purpose technologies that are widely adopted and induce further innovations.
Various authors identify positive productivity effects from ICT utilization (Brynjolfsson & Hitt, 2003;Bresnahan, Brynjolfsson, Erik, & Hitt, 2002;Jorgenson, 2001Jorgenson, , 2005Jorgenson, Ho, & Stiroh, 2005). These productivity effects coincided with a massive reduction in hardware prices over the last decades, which has spurred investment in IT and communication equipment (Jorgenson, 2001). Jorgenson (2005) finds that despite productivity growth rates being by far the highest in ICTproducing industries, the overall contribution of these industries to economic growth in the United States has been rather limited due to their low share of the economy. As ICT became more widely used, its main growth contribution came from total factor productivity (TFP) growth in ICT-using industries while growth rates in ICT-producing industries plateaued (Jorgenson, 2007). Similar empirical evidence is given by Brynjolfsson and Hitt (1995, 1996 and Tambe and Hitt (2012) on the effect of computers on firm-level productivity, confirming aggregate findings and painting a more nuanced picture. Cardona et al. (2013) give an overview of research on aggregate and firm-level ICT productivity effects.
"The Cloud" as a logical continuation of ICT-specific services emerged as an architectural innovation (Henderson & Clark, 1990) that was the result of isolated innovative processes and extended data transmission possibilities. Market research estimated the global private and corporate cloud computing market to have reached $56.6 billion (on public cloud services only) in 2014 and it is projected to more than double that by 2018 (IDC, 2014). Eurostat (Giannakouris & Smihily, 2014) reports that 19% of EU enterprises used cloud computing in 2014, mostly for hosting their e-mail systems and storing files in electronic form. The economic benefits of cloud computing adoption in the business segment of Europe's largest economies are estimated to have created 2.3 million net new jobs between 2010 and 2015 (Center for Economics & Business Research, 2010). Hence, major structural changes and productivity-enhancing effects are expected from the usage and diffusion of cloud computing. However, frustratingly for researchers wanting to investigate and quantify the growth impact of cloud computing, data on this phenomenon continue to be scarce.
We aim to advance the understanding of enterprise cloud computing as well as of the firms using it and the potential mechanisms triggered by implementation of this innovation. One of our contributions is that we propose an indirect measure of current or prospective cloud computing adoption that allows researchers to use existing large-scale firm-level panel datasets to analyze cloud diffusion and productivity effects via a reliable and plausible proxy. This is of particular importance as many extant surveys do not employ a precisely defined or even generally accepted measure and longitudinal studies are yet to be conducted. We utilize the widely used Harte Hanks technology database for 13 European countries and the years 2000 through 2007 to develop our cloud indicator and then merge this technology data with balance sheet information from the ORBIS database. Applying our indicator to the data, we make six observations on firm-level cloud computing regarding possible correlates of adoption and the correlation between firm productivity and cloud computing in the context of structural differences in industry sectors. These observations show how the economic effects of cloud computing could be analyzed using our indicator so as to provide initial insight into empirical cloud computing economics and shape an agenda for further research on cloud computing. We derive the following three suggestions for further research: 1. As adaptiveness of cloud computing differs widely across industry sectors, studies on cloud computing should be conducted at the industry level. For example, in our sample, services exhibit especially high adoption rates. 2. We need to understand why firms implement cloud solutions and what they actually do with the Cloud. Do they intend to increase productivity or flexibility, or both? 3. Cloud adaptiveness is potentially correlated to a firm's position in the supply chain and thus suggests a linkage of cloud adaptiveness and the type of output the firm produces as well as the market in which it operates.
We first outline the concept and market of cloud computing (Section 2). In Section 3, compile existing first steps toward a theory of cloud computing economics and review the empirical literature on the topic. Section 4 introduces the data and our measure of cloud adaptiveness; Section 5 presents six observations from descriptive analyses of this dataset. Section 6 concludes.
linked up to build a firm network, began to compete with the IBM System 360 family of mainframes (Cusumano, 2010) and eventually prevailed. PCs were easy to use and assumed some of the computing tasks, allowing for hosts to be less capable (Bresnahan & Greenstein, 1996). Operating systems and software were written and licensed by software companies, which reduced the upfront investment costs of IT-using firms and also improved the systems' agility. This type of client/server structure predominated in the 1980s and 1990s.
However, the idea of using a mainframe to centralize vital functions and capacities continued to survive. While mainframes required terminals with a command-line interface, modern thin clients could run applications and services hosted by a server with a graphical user interface. Together with grid and utility computing, this thin client network system was a precursor to cloud computing (Leavitt, 2009). From the mid-1990s onwards, the Internet spread farther and farther and became faster and faster. An early innovation quite similar to actual cloud computing involved application service providers (ASP), where thin client systems or the Internet were used to provide software services to a small number of users. However, it took the development of open-source software and the adoption of Web 2.0 standards before a system using relatively simple code and that could be widely accessed was possible (Grossman, 2009). The spread of reliable, high-speed networks further drove the development of cloud computing (OECD, 2009).

So what is cloud computing?
Cloud computing is actually a new manifestation of an "old trend," involving preexisting computing concepts and a novel combination of established components (Armbrust et al., 2009;OECD, 2009;Suciu, Halunga, Apostu, Vulpe, & Todoran, 2013). The key characteristic of cloud computing is the virtualization of resources and services. Cloud computing combines the efficiency of a mainframe with the agility of a client/server system. Firms outsource their IT systems either completely or partially, renting storage space or computing power from specialized providers. This is somewhat similar to earlier hosting services, but cloud users also benefit from additional services and scalability of capacities (Lin & Chen, 2012). Grossman (2009) identifies scale, simplicity, and pricing as the key defining features of cloud relative to conventional computing. This definition mirrors the widely accepted definition of cloud computing by Mell and Grance (2011) of the US National Institute of Standards and Technology (see Fig. 1).
As Fig. 1 shows, in contrast to simple server usage, cloud computing involves pooling IT resources such as storage or processing in a virtual system serving multiple users. Resource pooling allows for specialization and the realization of economies of scale on the provider side. Capacities are assigned dynamically according to demand; users therefore cannot locate their data in a certain geographic area. Importantly, cloud users often can purchase computing resources without any human interaction and at short notice (on-demand self-service). While traditional computing requires heavy upfront investment with fixed capacity, cloud computing allows rapid elasticity (scalability) of resources and firms order and pay for only the capacity they actually need at that specific moment. The services provided are automatically measured, which not only leads to resource optimization but also facilitates billing. Cloud computing is billed based on a pay-per-use pricing scheme. Finally, broad network access is indispensable to access and use cloud services (Armbrust et al., 2009;Mell & Grance, 2011).
There are three deployment models of cloud computing. (1) Software as a service (SaaS), where a customer purchases access to an application, such as enterprise resource planning (ERP) or customer relationship management (CRM), hosted and run in the cloud. (2) Platform as a service (PaaS) refers to access to platforms that allow customers, especially software developers, to test or deploy their own applications in the cloud. (3) Infrastructure as a service (IaaS) is a more basic service mostly offering access to storage capacities (National Institute of Standards and Technology, 2013; Suciu et al., 2013).

The cloud computing market
The cloud computing market is comprised of four major groups of actors: cloud consumers, providers, carriers, and enablers or complementors (Gerpott & May, 2014;Marston, Li, Bandyopadhyay, Zhang, & Ghalsasi, 2011;National Institute of Standards and Technology, 2013). The biggest consumer groups are firms, making cloud computing an important businessto-business (B2B) market, with products ranging from a complete cloud-based IT solution to select individual services. On the supply side of the public cloud market, we have vendors that own and operate the required data centers and platforms, including maintenance and upgrades of the system (Marston et al., 2011). Amazon (with Amazon Web Services) is the dominant provider, with Google (Google Drive), Microsoft (Windows Azure), and IBM (BlueCloud) distant followers (SearchCloudComputing, 2013). Cloud carriers provide interconnection from providers to consumers, so that most cloud carriers are telecommunication operators providing Internet access and connection (National Institute of Standards and Technology, 2013). Finally, cloud enablers "sell products and services that facilitate the delivery, adoption and use of cloud computing" (Marston et al., 2011). In other words, cloud enablers add value to bare-bones cloud services, making them cloud complementors. Cloud enablers or complementors are auditors, brokers, or additional-value service providers. A prominent example is Dropbox, which offers storage and file sharing solutions, but stores its data on Amazon's Simple Storage Service (S3) (TechTarget Glossary, 2011). Other enablers include consultancies that help firms implement cloud architecture. Moreover, cloud auditors are expected to become of increasing importance in the near future due to growing security concerns.
The cloud computing market is close to a state of maturity, with users having developed a more precise idea of their needs, and suppliers refining their business models to meet them.

Cloud computing economics
A review of the literature on cloud computing reveals that academic research is still exploratory and generalizable results few and far between. The common view in the literature is that cloud computing enables firms to reduce their fixed investments, overall costs, and risk while gaining flexibility. Cardona et al. (2013) document that IT usage, typically measured as the number of PCs in a firm, positively affects firm productivity. To our knowledge, there are no comparable data on cloud computing; prior work studies the possible economic effects using small samples or individual firms. 1

Costs, flexibility, and firm organization
One of the central economic results of cloud computing is the changing cost structure at the firm level. Cloud computing users do not have to invest in powerful personal computers and servers and hence do not incur high upfront investment and related capital costs. Instead, they incur variable expenses in the form of operating costs or pay-per-use fees (Armbrust et al., 2009;Bräuninger, Haucap, Stepping, & Stühmeier, 2012;Bayrak, Conley, John, & Wilkie, 2011;Etro, 2011;Klems, Nimis, & Tai, 2009;Yoo, 2011). This changing cost structure is considered to chiefly benefit small-and medium-sized enterprises (SME), as they have limited funds to invest in assets and suffer from unused peak capacities more than do large firms, in which tasks and peak times can be more diversified (Armbrust et al., 2009;Bayrak et al., 2011;Marston et al., 2011). Also, using cloud services avoids opportunity costs due to underutilization of local IT equipment and outdated software and security standards (Prasad, Green, & Heales, 2014).
In addition to the changing cost structure, there is consensus in the literature that cloud computing reduces total IT costs for firms or at least for SMEs. This cost advantage could originate from a specialized cloud computing vendor reaping economies of scale vis-à-vis an in-house IT solution. Hecker and Kretschmer (2010) call it "general wisdom [that] specialized outsourcing providers can produce more cost efficiently due to economies of scale, specialization," 2 and better utilization of equipment (Brumec & Vrček, 2013). Cloud users can benefit from these efficiency gains, but have to weigh them against the transaction costs of the outsourcing process (Riordan & Williamson, 1985). However, there is as yet no empirical evidence confirming this cost advantage. Brumec and Vrček (2013) model the costs of cloud computing usage and compare them with the costs of a de-novo on-premise computing system and show that leasing computing resources from Microsoft, Google, or Amazon is cost efficient for less demanding applications, but that it is still preferable to execute highly complex applications on-premises. Hence, they identify no universal cost advantage of cloud computing over on-premises IT systems.
Flexibility gains from cloud computing also affect firm organization. First, the number of IT staff on-premises can be reduced as cloud services handle many tasks previously undertaken by traditional IT staff, such as maintenance, updates, and the like. Second, the firm can react more quickly to changing conditions in its business environment. The flexibility of cloud computing even lowers entry barriers for new firms or sectors. Etro (2009) conducted a macro-simulation and finds that by lowering entry barriers, cloud computing could create up to 1 m jobs in the European Union. A third organizational and 1 In this paper, we focus on the user side of the market. On the provider side, interesting economic issues include competition and industry structure, capacity investment (Lam 2013), and pricing. 2 Hecker and Kretschmer (2010) further state that markets with high scale economies tend to concentrate, which might lead clients to change their outsourcing behavior as they are losing bargaining power. Lam (2013) studies cloud providers' capacity investment incentives according to different market structures.
potentially productivity enhancing effect of cloud computing is the alteration of business processes, allowing for changes in corporate culture, collaboration with business partners, and customer-faced services (Klems et al., 2009). Similar transformations occurred after the implementation of earlier technologies such as ERP and CRM (McAfee & Brynjolfsson, 2008). Overall, the adoption of cloud computing is expected to shift the production possibility frontier of firms outward and thus should result in higher total factor productivity.

Cloud computing and SMEs
As mentioned above, small and medium-sized enterprises (SME) are expected to benefit most from adopting cloud computing, which is why the literature has thus far focused on SMEs. Most of the studies on SMEs involve a fairly small sample and typically focus on one sector and/or one country. Alshamaila, Papagiannidis, and Li (2013) study cloud adoption in SMEs by conducting semi-structured interviews and analyzing the resulting data using the Technology-Organization-Environment (TOE) framework (explained in more detail in the following section). They find that except for competitive pressure, all factors of the TOE framework were relevant for the adoption of cloud services. Another interview-based study, this one of Irish SMEs, by Carcary, Doherty, and Conway (2013) finds that the most important reasons for cloud non-adoption are security concerns, the lack of time for implementation, and a generally low level of cloud computing in the company's sector. The study did not investigate drivers of adoption but a surprising 35% of survey participants claimed to be unaware of any cloud computing benefits. Stieninger and Nedbal (2014) find that firms are afraid that their corporate image will be negatively affected if they use cloud computing and that they are also concerned about security and privacy management. While most studies assume SMEs to be more likely to adopt cloud computing, Benlian, Hess, and Buxmann (2009) find that large firms have a head start in SaaS adoption, albeit an insignificant one. Conversely, Alshamaila et al. (2013) show that among SMEs, smaller firms are more likely than larger ones to adopt cloud computing.

The Technology-Organization-Environment (TOE) framework applied to the cloud
Based on the findings discussed above, we structure our argument along the lines of Tornatzky and Fleischer (1990), who identify three broad areas that determine a firm's adoption decision and the subsequent efficiency gains. First, the innovation has to fit the firm's existing equipment and processes, as well as its needs (technological context). If a firm already has a sufficient technological infrastructure, the firm is more likely to be able to implement the innovation successfully and to realize economic benefits. 3 An important prerequisite for cloud computing is an interconnected IT system within the firm and a broadband connection to the Internet. Second, the necessity and success of an innovation adoption depends on the firm's organizational characteristics, such as size, production processes, and so forth. It is shown that firm mechanisms and dynamics differ significantly by size (Mack & Rey, 2014). Moreover, a firm needs to be able to assimilate the innovation, so that "alignment between the objectives of an organization's IT strategy and business strategy is directly related to IT effectiveness and overall business/organizational performance" (Carcary et al., 2013). For cloud computing, the literature focuses on small firms as the decrease in upfront investment and the increase in flexibility are of particular advantage for these firms (Aljabre, 2012;Kern, Kreijger, & Willcocks, 2002;Stieninger & Nedbal, 2014). The third determinant identified by Tornatzky and Fleischer (1990) is the firm's environment, consisting of the industry, market structure and competition, regulation, and the like. The service sectors are expected to especially benefit from adopting cloud computing as they often dispose of huge amounts of data, need to exchange data with clients, or work from various or different locations. By contrast, firms with a security-sensitive environment might deliberately choose not to use cloud computing (Kshetri, 2013). Further, the firm's position in the supply chain, its market power, and the industry in which it is active all may (or may not) result in external pressure or requirements to adopt cloud computing. Finally, regulatory factors across countries may also matter for cloud adoption.
All three parts of the TOE framework seem to affect cloud computing adoption. However, none of the studies we found explicitly address heterogeneity between industries or discuss explicit productivity effects. With our indicator, first observations, and the resulting research agenda we take a first step toward closing this gap. Benlian et al. (2009) asked approximately 400 top IT executives of German firms to rate different application systems (e. g., ERM, CRM) with regard to potentially adopting them in the form of software as a service (SaaS). The survey samples of Trigueros-Preciado, Pérez-González, and Solana-González (2013) and Stieninger and Nedbal (2014) include 94 firms from Spain and nine from Austria, respectively. Unfortunately, neither study explains how cloud computing is defined. While these first exploratory results on the drivers of cloud computing adoption and economic effects are helpful, we neither know exactly what was measured nor can we say anything about diffusion patterns given these studies are in a cross-section. The same drawbacks are found in studies conducted for or published by firm research entities such as Deutsche Bank (Heng & Neitzel, 2012), KPMG (2013), and Deutsche Telekom AG (2010).

Existing microdata
When studying potential drivers of innovation adoption, surveys revealing decision makers' preferences are the correct methodological choice, especially if the surveys also include non-adopters. To identify economic effects, however, we need a variable for cloud computing usage that is consistently measured and invariant to who responds to the survey.

Our data and measure of cloud adaptiveness
We use the CI Technology Database (CITDB), which was developed by the market intelligence firm Harte Hanks and covers more than 260,000 European firm locations. Our dataset covers the years 2000 through 2007 and includes general firm characteristics like the number of employees and industry classification, as well as IT-specific information such as the number of desktop PCs, laptops, network devices, IT employees, and usage of different hardware and software. The CITDB has been used in prior studies such as Bresnahan et al. (2002), Kretschmer (2004), Forman, Goldfarb, andGreenstein (2012), and Kretschmer, Miravete, and Pernias (2012). We merged the technology data with financial data from the Bureau van Dijk ORBIS database so as to have company information on sales, assets, and so forth. The sample is largely representative, with a slight bias toward medium-sized and large enterprises.
There is information on cloud computing usage in the latest Harte Hanks data waves, but the variable's informative value is limited as there is no information on the underlying definition of cloud computing or on how the technology is understood by the person interviewed. The number of observations with cloud computing information is very low in our sample. However, CITDB contains detailed information on other IT resources and elements used in firms. With this information we construct a composite indicator of cloud adaptiveness based on the concept of architectural or combinatorial innovations (Henderson & Clark, 1990) and the TOE framework (Tornatzky & Fleischer, 1990). Loebbecke, Thomas, and Ullrich (2012) and Carcary et al. (2013) state that cloud computing is an evolutionary regrouping of earlier IT elements. We therefore construct our measure by studying the usage and combination of crucial IT resources that together lay the groundwork for cloud computing in a firm. Firms using this kind of IT structure are likely to introduce cloud computing at some point, in other words, they are cloud ready or cloud adaptive. Our definition of cloud adaptiveness matches the concept of architectural innovation developed by Henderson and Clark (1990), which is that an innovation is not always a departure from core concepts or a radical change of components' architecture; rather, "the essence of an architectural innovation [is] the reconfiguration of an established system to link together existing components in a new way" (Henderson & Clark, 1990). This is exactly how cloud computing came to be. Centralized computing and storage, interconnected IT resources, and the standardization of data were and are well-known components. However, the idea of virtualizing computer and server structures and offering software, infrastructure, and platforms as a service, changes the linkages between the components and therefore constitutes an architectural innovation. 4 We define technological and organizational readiness and will assess other organizational characteristics and the firm environment in our subsequent empirical analysis.
We build our indicator of cloud adaptiveness as a dummy variable that takes the value 1 if all four variables that describe the conditions for cloud adoption take the value 1. The variables are: (1) number of network devices per employee, (2) usage of a wide-area network (WAN), (3) share of laptops among all firm PCs, and (4) usage of groupware software (Fig. 2).
To proxy a firm's technological readiness for cloud computing adoption we use information on the number of network devices per employee and on the existence of a wide-area network (WAN) in the firm. Network devices allow direct access to internal networks or the Internet. A WAN indicates that a firm has a good network connection, most likely in the form of leased priority lines and broadband, allowing particularly fast data transfers. We assume that a firm with extensive network access possibilities and a good connection infrastructure is not only ready to use cloud computing in a next step, but also likely to be particularly open to data transfer and interconnection.
We proxy a firm's organizational readiness by its usage of groupware software and the share of laptop computers. Groupware, such as Lotus Notes, helps employees communicate or share documents, thus creating a common workspace. This means that firm employees are connected via IT and use common platforms, an important feature of cloud applications. Laptops point to flexible working patterns, including mobile access to firm data, platforms, or software. Firms using groupware and with a high share of laptops have already implemented organizational patterns compatible with a centralized and flexible IT structure. For them, adopting cloud computing is less costly and acceptance among employees likely to be high.
A firm is considered cloud adaptive if all four criteria are met (see Fig. 2). We transform the two continuous input measures-number of network devices per employee (1) and share of laptop computers (3)-into dummy variables. To do this, we compute the average across firms for each year and attribute a value of 1 to all firms with a number of network devices per employee or a share of laptops above the respective year's mean value. A cloud adaptive firm might already be using cloud computing or may adopt in the near future.
Our measure of cloud adaptiveness contributes to the literature by allowing researchers to use existing large-scale datasets for their work. Moreover, to date, the term "cloud computing" has not been well defined-neither in academic research nor by practitioners. Firm-level surveys asking about the "usage of cloud-computing" are thus of limited explanatory power. We develop a first measure of the phenomenon based on more precise firm-level data. In a survey, the questions that need to be asked to obtain our input variables can be answered clearly and unambiguously.
Clearly, a limitation of our indicator is that it does not measure cloud computing as a technology or computing paradigm, but, instead, the readiness of a firm to use it. We cannot measure the outsourcing and the scalability characteristics of a cloud and thus provide a lower bound of potential effects associated with being cloud adaptive. Further, we might be measuring a firm's technological sophistication rather than its actual cloud adaptiveness. This is a typical problem when working with proxies and cannot be entirely overcome. However, the indicator was constructed by focusing on the technological and organizational readiness for cloud computing and therefore measures a particular form of sophistication. As a robustness check, we split the sample into technologically sophisticated firms (firms whose PC intensity is above the sample median of PC intensity) and non-sophisticated firms: 99% of the non-sophisticated firms are also non-cloudadaptive, while 47% of the non-cloud-adaptive firms have a PC intensity above the median so that one does not automatically imply the other. The correlation between the sophistication dummy and the cloud dummy is 0.28. Another limitation is the equal weighting of the indicator's four input variables, implying that all four factors are equally important in classifying a firm as cloud adaptive. This is a simplification that cannot be remedied as long as there are no other studies on cloud computing and the transition from more traditional IT systems. Alternative approaches of varying weighting schemes like e.g. regression-based approaches of constructing composite indicators typically require a priori information on the selection of a dependent variable as target variable. However, as we preferably wanted to set up a general measure of cloud adaptiveness, we abstained from such an approach in our study. Our data cover the years 2000-2007, a period during which actual cloud services were not yet widely used. Given the variables and timeframe of our dataset, adaptiveness serves as a proxy for cloud computing. While the structure of our measure can be used for other datasets as well, a limitation of our analysis is that the Harte Hanks data were not collected by a research institution. This means that the sampling methodology does not guarantee representativeness and the questions are derived from practical considerations. However, the dataset has been proven useful and reliable in a number of academic studies.
We use two unbalanced panel datasets for our analysis. The first sample (adoption sample) is comprised of 73,985 observations from 25,434 companies in 13 European countries. For the paper's productivity analyses, we use a second sample (productivity sample) that is more restricted than the first sample due to the availability of variables needed to estimate the productivity measure (total factor productivity, TFP). 5 We now apply our measure of cloud adaptiveness to the data and investigate the distribution and characteristics of cloud adaptive firms, thus providing an early set of empirical findings on cloud computing at the firm level.

Observations on cloud adaptiveness
Our cloud adaptiveness measure captures a firm's technological and organizational readiness to adopt cloud computing. The descriptive observations in this section follow the TOE framework by looking at correlates of cloud adaptiveness (e.g., organizational characteristics of cloud adaptive firms, such as size) and at the firm environment (such as industry or supply chain position). Additionally, we document the productivity levels of cloud adaptive firms.

Large firms are early adopters but small firms catch up quickly
Differences in adoption behavior between smaller and larger firms are found in several empirical studies on general IT (Nguyen, 2009) and are expected in the case of cloud computing as well. However, our results over time are somewhat surprising (Fig. 3).
We would expect large firms (those with more than 249 employees) to be clearly ahead of small firms (those with less than 50 employees) when implementing productivity-enhancing IT. First, at the organizational level, large firms often have professional IT departments that keep abreast of trends and constantly work to optimize the company's IT structure. Second, large firms have the financial means to afford high investment in IT, while small firms are less able to make this type of investment and thus are not expected to be among the early adopters (Mack & Rey, 2014;Prasad et al., 2014). Further, there is empirical evidence that large firms imitate innovations more quickly than small ones (Geroski, 2000). Finally, large firms may have experience with early cloud-like structures. However, this also means that they might achieve lower cost advantages from cloud computing adoption than smaller firms as large firms might already be realizing economies of scale with their original data centers and IT networks (Marston et al., 2011;Shy, 1996). While our results match our expectations at the beginning and end of our panel period, small firms catch up within several years whereas medium-sized firms do not.
The characteristics of cloud technology differ from those of general IT and the literature contends that SMEs benefit more from cloud computing due to its pay-per-use and scalability features, allowing firms a degree of financial and capacity flexibility (Sultan, 2011) while avoiding high upfront investment. Our adaptiveness dummy cannot reflect this property.
However, it does capture whether firms have a highly interconnected IT structure and use central communication platforms. Large firms tend to have a more complex task structure than small ones and thus need various types of software and hardware (Kretschmer, 2004); standardized systems often cannot satisfy the varying needs of the different firm departments. Conversely, small firms' tasks tend to be more homogenous and generate rather uniform data. The transaction costs of becoming cloud adaptive are therefore lower for small than for large firms. Hence, the early catching up of small firms is not that surprising after all. In our panel, small firms exhibit the same level of cloud adaptiveness as large firms in 2002 and 2003. This coincides with Benlian et al.'s (2009) finding that when it comes to SaaS adoption, there is no significant difference between small and large enterprises. However, for the years 2004 through 2007, the cloud adaptiveness of small firms plateaus, whereas the share of large-firm adopters continues to rise. Later in the sample period, firms with 249 employees and less (small and medium-sized) have similar levels of adaptiveness, which is lower than that of large firms. While the different structures and challenges of small and large firms are intuitive to a large extend, the situation of medium-sized firms is ambiguous. Hence, firm size seems to be associated with cloud adaptiveness, but this finding does not lead to easily interpretable stylized facts about firm adoption behavior. Rogers (1995) states that size is "probably a surrogate measure of several dimensions that lead to innovation."

Service firms are more cloud adaptive than manufacturing firms
We expect to find heterogeneity across sectors in cloud adaptiveness. The industry in which a firm is active shapes the firm's technological needs as well as its organizational characteristics, and, almost by definition, the market in which it operates. The tertiary or services sector 6 is typically more data intensive than the manufacturing sector, which could explain this sector's high share of cloud adaptive firms (Fig. 4). For example, financial services are highly data intensive, processing research and trading operations and computational algorithms for risk management, storing and transmitting large amounts of data, and requiring time-sensitive communication with clients and trading partners. Data-intensive applications are also required in logistics and transportation, where supply-chain optimization, automated processes, and consolidation of global supply-chain providers are a source of competitive advantage. The same holds for knowledgeintensive business services (KIBS) (Mack & Rey, 2014;Musolesi & Huiban, 2010) such as consulting or IT outsourcing that create value by transferring their knowledge to clients, which requires sophisticated data management and transfer systems. In their survey of Irish SMEs, Carcary et al. (2013) find that the majority of the cloud adopters in their sample are companies active in the KIBS sector.
Traditionally, services were viewed as technologically backward and passive adopters of technology, but this view has changed dramatically with tertiarization and information technologies (Musolesi & Huiban, 2010). Scalability and mobility are crucial in services; however, in manufacturing firms producing on-premises, a traditional server structure is often more appropriate. However, this could change with the advent of "smart factories," in which production is highly interconnected and computer optimized. 7 5.3. In manufacturing, upstream capital goods industries are more cloud adaptive than downstream consumer goods industries.
In production theory, factors of production-capital, labor, and materials-serve as inputs to firms' internal value chains, which link up to form an industry-wide supply chain (Porter, 1985). In manufacturing, upstream firms provide materials and intermediate inputs to firms more downstream the supply chain, which themselves produce capital goods. These are then used by firms farther down the supply chain to produce consumer goods.
In our data, firms in the capital goods sectors such as machinery, industrial electronics, and measurement equipment show higher cloud adaptiveness than firms in the consumer and intermediate goods industries; indeed, the latter, being upstream industries, are surprisingly non-cloud-adaptive (Fig. 5). 8 A possible reason for this is that upstream sectors are not incorporated into e-business operations as much as sectors in the middle and closer to the end of the supply chain. 9 Also, intermediate goods sectors are usually more specialized and therefore do not need the same flexible interconnecting capacities as businesses in the capital or consumer goods sectors, which depend more on markets and customers. More precisely, as the capital goods sector is embedded at different stages of the value chain, mostly supplying technically advanced machinery for production, it not only has high technological requirements, but also engages in marketing and sales activities. Goods are often tailored to customer needs, which can be a highly data-intensive process. The consumer goods sector is less cloud adaptive, but still has higher rates of adoption than intermediate sectors. A survey of the fashion and apparel sector discovered that one reason for the limited use of e-business applications within the industry was the "lack of interoperability between the many systems in use" (DG Enterprise & Industry, 2012). This statement applies more generally to supply chains, particularly in sectors where product lifecycles are short. More sectoral-level research is needed to better understand the role of cloud adaptiveness in single sectors. 5.4. In services, unregulated market sectors are more cloud adaptive than nonmarket sectors; cloud adaptiveness can differ significantly within single supply chains In our sample, business and financial services, as well as wholesaling, are the most cloud adaptive service sectors, while retail trade and regulated, state-dominated industries, such as health, education, and social services, are the least likely to be cloud adaptive firms (Fig. 6). 10 This is intuitive as business and financial services are very data-intensive sectors, requiring not only the treatment but also the exchange of data permanently and in real time. Falck, Haucap, and Kühling (2013) analyze the diffusion of e-health, that is, interconnecting ICT applications in the health-care sector. They find that doctors and hospitals use some ICT applications, but that information exchange and teleconferencing are not widespread. Data security and quality of services are requirements a health-care network or cloud system needs to fulfill. Hence, the low cloud adaptiveness of these sectors could be due to insufficient service stability, a still incomplete legal framework for (international) cloud services, and the absence of widespread cloud certification services, the lack of each of which leads to a lack of trust and security in cloud computing.
Wholesale firms work closely with producers but they do not produce goods themselves. They resell to retailers, who in turn sell the goods to private consumers. A characteristic of the retail and wholesale businesses is that, in contrast to other service sectors, they have not only a flow of information, but also a flow of goods, along the supply chain (Prajogo & Olhager, 2012). In our data, the wholesale sector is one of the most cloud adaptive service industries, whereas retail shows a very  weak disposition toward cloud computing (Fig. 6). This is surprising at first sight as both are closely vertically related and basically execute very similar tasks. Their operations differ in scale and complexity though. Retail business is generally locally oriented; wholesale tends to operate at a global scale and therefore faces a much greater challenge in organizing both the flow of information and the flow of goods. A global wholesale company needs to be constantly in touch with geographically dispersed producers, not only communicating and negotiating, but also observing and analyzing the development of international markets. Looking downstream, these companies manage and maintain a large distribution network. By contrast, a locally operating retail firm requires less e-business interaction. Still, it is surprising that companies upstream and downstream the supply chain can cooperate without being fully integrated, that is, without using the same IT structure and communication channels, especially since integrated logistics system can help reduce shortages and optimize inventories (Prajogo & Olhager, 2012). Cachon and Fisher (2000) conducted an empirical study on the value of information sharing in a grocery supply chain between supplier and retailer. They find that implementing information technology such as scanners and electronic data interchange (EDI) that allows quicker order processing and sharing of demand and inventory data can reduce supply chain costs by as much as 10%. Cloud computing could further enhance these advantages. However, IT integration of a supply chain is not costless. The retail trade market is currently experiencing the emergence of large players that pressure smaller suppliers into the adoption of their specific IT systems (DG Enterprise and Industry, 2012). Standardized cloud computing solutions might be cheaper and more widely compatible. We therefore expect retail companies to catch up, with regard to cloud adaptiveness, in the near future.

In manufacturing, cloud adaptive firms are more productive
Evidence on the productivity effects of adopting cloud applications is scarce. Employing our cloud adaptiveness measure for two different types of productivity measures separated by industries provides some initial observations on the productivity of cloud adaptive and non-cloud-adaptive firms in manufacturing sectors. 11 In Fig. 7(1), cloud adaptive manufacturing firms generally exhibit higher average labor productivity 12 compared to manufacturing firms that are not as cloud adaptive, except for the machinery sector. Sectors with large differences in labor productivity are instruments, chemicals, and transport equipment. As labor productivity is only a partial productivity measure, it reflects the joint influence of a host of factors and therefore it is easily misinterpreted as technical change. An alternative productivity measure is total factor productivity (TFP), the residual in the firm's output function after having controlled for capital and labor as physical inputs. Hence, as TFP abstracts from the effect of inputs, it is a more appropriate measure of technical change, which further allows for the incorporation of spillovers and therefore proxies for cloud computing as a general purpose technology. 13 According to our estimates ( Fig. 7(2)), there are TFP differences throughout all manufacturing sectors. These findings underline the importance of accounting for the specific nature of cloud computing as it is incorporated in technical change.
Our analysis does not allow drawing causal conclusions on the productivity effects of cloud computing or cloud adaptiveness. However, there is empirical evidence that cloud computing and firm productivity are highly correlated. Not only do cloud adaptive firms achieve higher sales per employee, they also are more successful in the choice and employment of technologies, measured by total factor productivity. A causal specification needs to resolve whether these productivity differences are driven by cloud adaptiveness. IT-driven intra-industry productivity differences are found in several studies controlling for various other factors (e.g., Brynjolfsson & Hitt, 1996;Tambe & Hitt, 2012). Cardona et al. (2013) confirm that most studies use either PC intensity or IT capital as a measure for IT, which was reasonable in the 1990s, the time period most productivity studies focus on (Tambe & Hitt, 2012). We would expect similar intra-industry effects driven by modern ICT like cloud computing and its direct precursors, which can no longer be measured based on capital as services became an essential part of it. 5.6. Cloud-similar technologies are not necessarily adopted in the sectors where they allow for the highest productivity We would expect industries in which cloud adaptive firms have a large productivity advantage to also have a high adaptiveness rate. Interestingly, this is not what we find in the data (Fig. 8). The correlation between a sector's cloud adaptiveness and the productivity difference between adaptive and non-adaptive firms in this sector is slightly negative. 14 This finding shows a strong heterogeneity among a sector's cloud adaptiveness and its firm-level productivity performance. 11 We find the same qualitative results in the services sectors, but the picture is less clear. In all service sectors, cloud adaptive firms are more productive than non-cloud-adaptive ones and we also find significant mean differences for almost every sector. However, these improvements are either in labor productivity or in TFP, not necessarily both. We find no significant differences in the financial sector or in the health/education/social sector. The partly insignificant differences might be due to higher within-sector firm heterogeneity. 12 Labor productivity is measured as sales divided by employment. 13 To estimate TFP at the firm level, we follow Levinsohn and Petrin (2003). This method accounts for biased coefficients of the production function originating from unobserved productivity shocks by explicitly modeling capital and intermediates within the estimation. This method ensures consistent estimates of inputs in the production function and thus enables unbiased calculation of total factor productivity by subtracting estimated input contributions from output. For further explanation, see also Van Beveren (2012). 14 Note that in this section we consider both manufacturing and service sectors.
While e.g. a given adaptiveness in the print sector is associated with relatively low TFP differences between cloud adaptive and non-cloud-adaptive firms, the same given adaptiveness is associated with high TFP differences between firms in retail. One potential explanation why firms' productivity advantages are not necessarily associated with its cloud utilization across all sectors is reverse causality. That is, more productive firms might be more cloud adaptive in general, so that TFP differences between cloud adaptive and non-cloud-adaptive firms in a sector only weakly correlate with the adaptiveness share. Further, recall that the productivity measure of TFP is not input driven. If a firm observes a cloud productivity potential in its sector, it might take a while to catch up, become cloud adaptive and realize the productivity advantage. Another explanation of the low adaptiveness in sectors with a high productivity potential is a first-mover advantage meaning that the firms that became cloud adaptive first realize productivity advantages other firms cannot. This can be due to limited needs of this technology in the sector, bounded growth potential of the industry or particular market characteristics. However, adoption of the software and hardware that readies firms for cloud computing might be driven by factors other than direct productivity gains. One driver of adoption is flexibility gains (see Section 3). Another reason for adoption is pressure from business partners or the necessity of integrating into a supply chain (see Section 5.4). While these measures might not be productivity-enhancing in the short run, they can secure the survival of the firm in the medium or long run. Moreover, particularly at the start of an innovation adoption lifecycle, firms often choose to adopt for non-

Conclusion
Cloud computing is expected to generate productivity effects in firms and growth in the economy. We develop a measure of cloud computing that lets us examine its diffusion pattern and its association with firm productivity across industries. Our measure builds on an existing panel of firm-level data and takes into account the genesis of cloud computing as an architectural innovation. We thus observe firms' cloud computing adaptiveness over time and study adoption and productivity patterns at a stage where comprehensive firm-level panel datasets on cloud computing are not yet available. Our six empirical observations give us an opportunity to suggest an agenda for further research.
Our findings show that firm size is not a good predictor of cloud adaptiveness per se; rather, it is other firm characteristics that are correlated with the adoption decision. We also find that the service sector is more cloud adaptive than the manufacturing sector and that it is especially business services and the financial and wholesale sectors that are most cloud adaptive. Interestingly, in the manufacturing sector we find positive productivity differences between cloud adaptive and non-cloud-adaptive firms. This productivity advantage, however, does not necessarily benefit a large fraction of the firms in this sector.
With a view to the currently still poor data situation, we suggest employing our cloud adaptiveness dummy in future empirical research on cloud computing and the underlying economics of it. Our approach enables scholars to work on diffusion and productivity and to establish appropriate econometric identification strategies that help test theoretical predictions. In the long run, data-collection efforts should focus on a representative panel, be based on a precise and thorough definition of cloud computing, survey firms on their business and production processes that use a cloud service, and collect information on how cloud computing affects firm costs, communication, and organization. With the help of the TOE framework and based on this paper's observations, future work should attempt to discover the key characteristics that prompt firms to adopt cloud computing and result in successful, productivity-enhancing implementation. More specifically, studies on cloud computing should be conducted at the industry level due to massive heterogeneity. Further, given that the concept of cloud computing is very broad, care should be taken to discover which elements of cloud computing firms actually implement (e.g., IaaS, SaaS, PaaS, or even more fine-grained aspects), and what they actually do with it, to understand the underlying potentially productivity-enhancing mechanisms. Finally, cloud computing adoption varies based on a firm's position in the supply chain and thus suggests a linkage of cloud adaptiveness and the firm's type of output or its market power in addition to its industry.
By modeling and quantifying such underlying mechanisms, the economic effects of cloud computing can be understood and a consistent framework of cloud computing economics can be developed, especially at the firm and industry level.

Appendix A. Dataset
We use a dataset with data from two sources: (1) the Harte Hanks CI Technology Database (CITDB) and (2) Bureau Van Dijk's ORBIS.
Harte Hanks, a market intelligence firm, conducts annual telephone surveys to take stock of specific IT types used by individual sites (establishments) of more than 10,000 German firms. As the CITDB data are collected at the establishment level while the ORBIS database covers the company level, we aggregated the Harte Hanks dataset to the company level or extrapolated where required. The first step was the identification of all establishments that belong to one company. We matched bvd (Bureau van Dijk) company IDs to the CITDB site IDs using the company name, the zip code, and the three-digit SIC code of every observation. Next, we had two types of technology variables to aggregate: dummy variables were set to 1 at the firm level if any of the company's establishments used this technology, for example. GROUP, and 0 otherwise. Integers, such as the total number of PCs, were summed and expressed as relative numbers (per employee). As the data structure is rather complicated and sometimes misleading, the results were cross-checked carefully and in some cases weighted.
In this paper we use two different extracts from the resulting panel dataset: (1) The adoption sample for Sections 5.1-5.4 of the paper and (2) the more restricted productivity sample for Sections 5.5 and 5.6.
We chose this approach because of the high number of missing data points for the balance sheet variables required for the productivity analysis. Refer to Tables 1-4 for sample statistics.  Notes: A firm can grow or shrink throughout the panel. The service sectors include Retail, Wholesale, Business Services, Financial Sector, Transportation, Health/Education/Social, and Other Services (legal, personal, and repair services, amusement sector). "Other sectors" include the primary, construction, public utilities, and public administration sectors. Notes: A firm can grow or shrink throughout the panel. The service sectors include Retail, Wholesale, Business Services, Financial Sector, Transportation, Health/Education/Social, and Other Services (legal, personal, and repair services, amusement sector). "Other sectors" include the primary, construction, public utilities, and public administration sectors.