Cloud assets and the “unit of compute”: Market dominance in commercial computing

Introduction

There is a narrative in the political economy of digital infrastructure that the popularity of generative artificial intelligence (AI) has triggered an investment boom in the hardware underlying it. Indeed, soon after the commercial release of ChatGPT, Microsoft and Google each announced quarterly spending of about $12 billion on data centers to support AI services. ¹ Yet market analysts were “concerned” that these investments would not be profitable (Hodgson and Kinder, 2024). A few months later, however, when the firms announced revenue growth from AI, investor confidence returned, and stock prices rose. ² How did these cloud computing providers boost their revenues? It was not by simply attracting new customers. Rather, Microsoft, Google, and Amazon—defined as “Big Cloud”—drew on a particular form of techno-economic expertise they have been refining long prior to commercial AI; namely, the strategic configuration of computing infrastructure as an income-generating asset. ³ Unpacking this strategy provides important context for the subsequent and growing investment in their computing infrastructures (e.g. data centers).

We situate this recent spending boom on data centers within a longer-term and highly contingent process of asset concentration by Big Cloud (cf. Morozov, 2022). Doing so de-centers AI as the focus of attention in the political economy of large technology companies, and centers what the industry calls “compute,” the simultaneously material and immaterial resource undergirding market power in the cloud computing industry.

The magnitude of investments by Big Cloud as they compete to sell compute has been described by van der Vlist et al. (2024) as the “industrialization of AI.” The term is apt in that it captures the scale and rationalization of data centers in their materiality. But the metaphor of “industrialization” insinuates that there is something distinctly novel in Big Cloud's recent capital spending. We point out, to the contrary, that this investment is only the most visible and recent manifestation of a phenomenon ongoing for two decades: what the industry calls the “cloud migration,” in which organizations give up owning servers and become “tenants” of Big Cloud's infrastructure. Everything from Moderna's Covid vaccine discovery software to McDonald's global data repository to the US Security Exchange Commission (SEC) itself runs on Amazon Web Services (AWS), Microsoft Azure, or Google Cloud Platform (respectively). In this light, there is an analytical danger of ignoring the continuity of AI with the older and more mundane commercial cloud: complicity in the hype of AI's newness. Put differently, AI is just the most recent form taken by the established business of cloud computing. ⁴ Our empirical contribution in this paper is to trace the longer-term development of compute, as the flexible, general-purpose resource now dominated by three Big Cloud providers.

In financial terms, digital infrastructures that produce compute are generally regarded as tangible assets, while data, algorithms, and platforms are regarded as intangible assets. This distinction shapes how companies and investors understand the value of digital infrastructure. Yet it is important to recognize that “assets” here are epistemic categories (e.g. in accounting), consequential because they provide the basis for calculating a firm's financial value (i.e. market capitalization). In this paper, we take a constructivist approach to assets and assetization, tracing the specific form that has emerged to undergird Big Cloud's market power. We argue that the assets Big Cloud is investing in are neither strictly tangible nor intangible, but rather a hybrid form wherein physical infrastructure is configured through the deployment of technical expertise, accounting practices, and market devices to become valuable.

In making this argument, we build on two current but often separate debates about the market power of large technology firms. One literature stresses the role of intangible assets (e.g. intellectual property, research and development, data) in these firms’ dominance (e.g. Durand and Milberg, 2020; Klinge et al., 2023; Rikap, 2021; Sadowski, 2019); the other literature stresses the importance of tangible assets (e.g. hardware, data centers, network interconnections) (e.g. Birch et al., 2021; Edwards et al., 2024; Ferrari, 2023; Greene, 2022; Narayan, 2022, 2023). Stakes in the analytical role of tangible versus intangible assets in explaining capitalist dynamics run deep. Veblen (1908: 115) was suspicious that “intangible assets serve no materially productive work, but only a differential advantage to the owner of the industrial product.” Today, the debates touch on the technopolitics of built infrastructure vis a vis imaginaries of the post-industrial economy, and on computing's environmental footprint vis a vis the privilege of knowledge work in appropriating economic value (e.g. Bell, 1973; Crawford, 2021; Mitchell, 2002; Pistor, 2019).

Through engaging with these two literatures, we develop the concept of “cloud assets” as a hybrid category (of intangible and tangible) to understand Big Cloud's business models and accumulation strategies. Ambiguity with regards to a tangible versus intangible binary is reflected in the commonplace term “digital infrastructure” (Khanal et al., 2025). It is unclear whether this term refers to the servers, or to platforms that sit on top of them, for example, generating data and running large language models. Our argument is that neither the infrastructure nor the platforms are reducible to the other as assets. Rather, cloud assets have emerged as a historically specific asset form that has enabled Big Cloud's seemingly unassailable market positions in compute. Following from this starting point, we ask: how do these firms configure the terms and costs which govern others’ access to the resource (i.e. compute) that cloud assets both produce and enclose?

Methodologically, our arguments emerge from a qualitative and financial analysis of regulatory filings and earnings calls from the three US-headquartered Big Cloud firms, Alphabet/Google, Amazon, and Microsoft, from the years 2010 to 2022. These firms represented 67% of the cloud computing market in 2024. ⁵ By tracing data in the SEC 10K forms that these companies filed quarterly, we illustrate the significant increase in their tangible asset investment between 2010 and 2022. But tangible asset spending is not the whole story. We also manually and inductively coded annual shareholders’ reports and transcripts of quarterly earnings calls (144 in total) from these three firms using the data analysis software Atlas.ti. Initially focused on their investment in data centers, our search terms for this corpus included “tangible assets,” “hardware,” “land,” “buildings,” and “capex.” What we found led us to the more complex picture of the technical, accounting, and organizational work required to make digital infrastructure valuable. These texts reveal that Big Cloud's long-term strategy depended on identifying and configuring what we call the “unit of compute” as the basis for generating revenue across their business segments. Over time, they have sought to deliver units of compute more efficiently and at decreasing cost, leading to what we speculate is a “digital Jevon's effect,” where their investments in compute reduce the unit cost of cloud computing while increasing absolute usage. The result is an ever-deeper integration of and dependence on cloud computing within the economy at large.

In the following section, we conceptualize cloud assets by drawing on the literature on assetization, which sits at the interface of science and technology studies and political economy. Then we present data on capital expenditures by Amazon, Microsoft, and Google, and interpret their business model and investment strategy in relation to the hybridity of cloud assets. We subsequently analyze the specific epistemic devices of quantification and pricing that configure the unit of compute as a revenue-generating hybrid asset. We finish by discussing the implications of these findings for understanding power in the broader digital economy.

Theorizing cloud assets, and the conditions of their emergence

The question in this paper is the composition of Big Cloud's asset base; that is, the resources that it uses to generate revenues. Our constructivist account provides more specificity into the contingent structuring of digital infrastructure as a political-economic object—particularly regarding the interplay of the epistemic and the material technologies that facilitate market power. To do this, we turn to the literature on assetization. Assets are defined as “something that can be owned or controlled, traded, and capitalized as a revenue stream, often involving the valuation of discounted future earnings in the present” (Birch and Muniesa, 2020: 2). As a category of financial accounting, assets are specified and given value in a business's financial statements. This matters because a company that owns assets is able to borrow money against them, sell them, and generally wield them as forms of power (Nitzan and Bichler, 2009). It is therefore important to understand how epistemic framings (e.g. rules of accounting) enable assets to perform economically. This is evident in recent assetization literature concerned with how something is turned into and governed as an asset through the configuration of legal entitlements, epistemic claims, social practices, technological devices, and discursive narratives (Birch, 2017, 2023; Birch et al., 2021; Duncan et al., 2022; Langley, 2021; Williamson and Komljenovic, 2023).

Taking this assetization lens allows us to highlight an important ambiguity in the literature concerning the basis of Big Cloud's power and influence: is it intangible or tangible assets that have enabled these firms’ dominance? On the one hand, some scholars have argued that digital technology markets are structured by Big Cloud's amassing of key intangible assets like data and intellectual property (e.g. Durand and Milberg, 2020; Klinge et al., 2023; Rikap, 2021). This includes Amazon's massive portfolio of patents on robots, process optimization, and search and query (Rikap, 2022). It also includes the more amorphous “data-driven intellectual monopolies,” where Amazon, from its core position in a network of supplier and third-party firms on its platform, extracts rents from innovations at the expense of firms it collaborates with (Rikap, 2022, 2024). Other scholars extend this logic to highlight how intangible assets are a means for perpetuating colonial dominance through hierarchical value chains (Franco et al., 2024; Irani, 2018). Mainstream business and accounting scholars concur on the importance of intangible assets, but for a different reason: revenues from intangible assets can be “scaled up” almost indefinitely because they require little additional labor once created (Haskel and Westlake, 2018). Here, intangibles are aligned with policies that promote knowledge work over industrial and service labor, and depend on intellectual property regimes that favor the global north. ⁶

On the other hand, recent research has shown that Big Cloud invests more in tangible assets than intangible assets, and their tangible investments even outpace non-digital corporations in this regard (Birch et al., 2021; Morozov, 2022). This reflects a broader literature that centers attention on the materiality of cloud computing, evident in Edwards et al.'s (2024) outline of “critical data center studies.” In this vein, a number of researchers have analyzed the role of physical infrastructure in the digital economy. In particular, Narayan (2022) highlights the elasticity of cloud computing and the important scale dynamics of cloud computing regimes, enabling both the rapid expansion and contraction of compute. Greene (2022) provides a detailed history and analysis of the physical infrastructure developers and providers underlying the internet. Others like Widder et al. (2023) and Vipra and West (2023) illustrate the concentration of computing hardware in the provisioning of AI services.

To reconcile these positions, it is not sufficient to say that both tangible and intangible assets matter. Rather, we propose that a new form of hybrid asset is emerging from the historically contingent conditions of the commercial cloud computing industry. “Cloud assets” is the analytical term we use to make sense of this articulation between tangible and intangible assets. The term points to the way that digital infrastructure must get configured through technical expertise, accounting practices, and market devices in order to function as an asset. In parallel, we introduce the notion of the “unit of compute” to refer to that which is sold to generate revenue from cloud assets. Units of compute are defined and quantified through specific temporal schemes, pricing mechanisms, and configurations of computing hardware.

Cloud assets have emerged in relation to three broad shifts, both technical and organizational, in recent years. The first concerns transformations in the product lines of cloud providers. Amazon, Microsoft, and Alphabet/Google did not develop cloud computing as a commercial product from scratch, nor did they “pivot” into an existing cloud market. Rather, they constructed that market in a contingent and ongoing process, as they iteratively innovated on their own in-house data center and network engineering needs, and then realized its commercial potential. Initially, these firms developed significant internal computing capacity to run their own business operations. Then, from the mid-2000s (Amazon) and mid-2010s (Alphabet/Google, Microsoft) onwards, they sought to capitalize on the scale of their investment in this computing capacity by selling access to it to others—thus creating cloud assets. In earnings calls, these firms regularly announced that they had extended the depreciation schedule of the book value of these tangible assets, because their engineers had found ways to extract more saleable and immaterial units of compute from each generation of hardware (capital assets) before it was deemed obsolete. Such temporal durability is central to both assets (Pistor, 2019) and infrastructure (Mitchell, 2020), and in this context positively impacted corporate cash flows and by extension dividends and stock analyst sentiment. This is a key point: technical innovations were inseparable from, but not reducible to, the accounting/financial value of these assets.

Second, the rise of cloud computing reflects a transformation in the mode of accessing computing resources, from a dependence on local physical machines to the emergence of “virtual machines” (VM) owned by a third party. A VM is a software emulation of an individual computer, which appears to the user to have its own processor, storage, and network identity. In actuality, it shares underlying physical hardware with other VMs, mediated by an additional software layer. As a result of “virtualization,” multiple users can share the same server simultaneously. Once computers became virtual, the next logical step was the unbundling of data storage from processing capacity. This unbundling has been foundational for the emergence of cloud assets. Unbundling was enabled by modular access to storage and compute through platforms, plugins, and APIs (Birch and Bronson, 2022; Grabher and van Tuijl, 2020; Helmond, 2015)—all facilitated by market devices (Callon, 2021; Muniesa et al., 2007) that enable specific pricing structures. Put differently, these cloud assets are a set of modular infrastructures for separating storage from computing capacity. This entire arrangement of commercially hosted computing is referred to as “infrastructure-as-a-service,” ⁷ reflecting a shift in business model away from the one-time sale of a product (Ferrari, 2023; Sadowski, 2019). Cloud assets are constituted by this subscription model, reflecting the way that cloud contracts establish repeated revenue streams, which then underpins the value (and valuation) of these cloud assets.

Third, the embrace of (tangible-)asset-lite business models and operations by other businesses reflects the pursuit of scalar logics (Pfotenhauer et al., 2022)—such as fast and cheap scale-up—enabled by the modular configuration of computing capacity and its disaggregated pricing. That is, startups and other businesses could free up their capital for investment elsewhere (rather than on expensive servers) (Narayan, 2022). Cloud providers’ revenues from subscriptions generate increasing returns through what we speculatively define as a digital Jevon's effect (sometimes called a “paradox”). In the mid-19th century, economist Stanley Jevons noted that improvements to the efficiency of steam engines did not reduce absolute demand for coal, as a static model would predict, but rather increased coal demand in aggregate, as steam engines themselves became more relatively cost-effective (Jevons, 1865). In our case, as Big Cloud invests in cloud assets, the price of each unit of compute drops while the overall supply increases. Nonetheless, their absolute revenues from compute increase as lower prices induce other businesses and organizations to shift their own capital expenditure from owning computing infrastructure to subscribing to cloud services. Through this digital Jevon's effect, cloud assets become increasingly valuable. Meanwhile, all of this is reinforced by Big Cloud promoting ever-new reasons to consume more of their ever-cheapening unit of compute—including AI applications like large language models.

Big Cloud's investment in computing infrastructure

In this section, we present data on the scale of Big Cloud's capital expenditures on the tangible component of cloud assets. To appreciate their long-term strategy for market domination, it is instructive to consider the time horizons over which infrastructure more generally is expected to facilitate capital accumulation. “Compared to other means of capturing revenue,” Timothy Mitchell (2020) writes, “infrastructures are unusually durable.” While the commodities brought to market via railways and pipelines “create little in the way of long-term revenue streams, the life of infrastructure is typically measured in years or decades.” A strikingly similar vocabulary and temporal orientation is expressed by Big Cloud executives as they explain to market analysts their own investment strategy:

Construction of data centers is primary, it is the core infrastructure. And we have been investing in real estate … with an eye of looking for the long term, rather than just the next 12 months, because if you decide you’re going to grow in a place, then you need the capacity for multi-years. (Chief Financial Officer, Google, 2014a)

Below we outline Big Cloud's capital expenditure (CapEx), which is a measurement over time of investment in and maintenance of assets. A decade ago, Microsoft and Google were each spending under US$2 billion in CapEx every fiscal quarter. By 2022, these numbers had tripled to US$5-6b. In 2021, Amazon's US$6b spent on infrastructure for AWS outweighed the US$4.5 billion it spent on its e-commerce fulfillment centers (Amazon, 2022). Although these measurements lump together commercial office spending along with land and digital infrastructure supporting the cloud business, executives’ statements demonstrate that the latter constitute the bulk of the asset value: for example, “the majority of the CapEx this quarter was spent on production equipment—machines, data center construction” (Google, 2013); and “the primary driver of the CapEx does continue to be our expectations for compute requirements” (Google, 2019). Figures 1, 2 and 3 show total assets reported on firms’ balance sheets, focusing on the categories of “Land and Buildings” and “Property and Equipment”—that is, all the tangible components needed to build data centers. Importantly, as the figures illustrate, the firms have been markedly growing their stock of these tangible assets. Between 2005 and 2021, Microsoft's increased about 20-fold in book value, Google's about 30-fold, and Amazon's about 50-fold.

OPEN IN VIEWER

Figure 1.

Amazon's Asset Investment (2010–2021) (in millions of US dollars).

OPEN IN VIEWER

Figure 2.

Microsoft Asset Investment (in millions of US dollars).

OPEN IN VIEWER

Figure 3.

Alphabet/Google's Asset Investment (in millions of US dollars).

Big Cloud firms have refined how they monetize the physical component of the cloud asset just as much as they have the intangible component. Microsoft's CEO Nadella explained that the company moves rapidly through phases of land acquisition, data center construction, and installing equipment, adding “we have driven significant process improvement to essentially make it as efficient as one can make it” (Microsoft, 2015). Construction reflects global ambitions. The three dominant providers each operate dozens of “cloud regions,” or self-contained network infrastructure zones, in the US, Western Europe, Southeast Asia, Japan, Australia, Latin America, and South Africa. Microsoft was the first of these three to build a data center in China, in 2014; Amazon now operates there as well.

Importantly, these firms’ amassing of cloud assets is not merely an incremental response to incremental demand from corporate customers seeking to purchase off-premises computing. Rather, the investments reflect the emergence of a deliberate plan to secure market dominance by preemptively establishing an unassailable position in the cloud computing market. This much they are keen to demonstrate to market analysts who influence stock prices. On a 2014 Google earnings call, an analyst asked if the company was “buying data center sites in advance of demand?” The CFO answered affirmatively: “we have found that the option value of having more capacity on standby and available to us to grow versus not having it is actually a real strategic issue” (Google, 2015). In other words, they are grabbing land in advance of need. And the prior year, that CFO justified significant cash outlays on real-estate acquisitions as “really good opportunistic purchases….The worst thing for us would be to actually not have machines and data center capacity” (Google, 2013). Because of the high profit margins on selling compute, the cost of carrying idle infrastructure is outweighed by revenue generated from bursts in customer usage: as Amazon's CFO Brian Olsavsky revealed to analysts—and this is a key point—“every percentage utilization in our data centers is worth tens and more millions of dollars” (Amazon, 2019). Finally, all these statements on earnings calls about capital expenditure are not primarily descriptions of present fiscal reality—for that, one can read the 10Q quarterly earnings reports filed with the Securities and Exchange Commission. Earnings calls are, rather, a signal to analysts and investors that assets are being accumulated; assets which they believe will generate reliable revenue from cloud computing in coming years.

Configuring revenues in cloud computing

This section offers an account of how Big Cloud configures the intangible component of cloud assets through legal entitlements, epistemic claims, technological devices, and narratives—that is, how they turn inert data center hardware into assets producing units of compute. It proceeds in three parts, examining first the mechanism for software abstraction that isolates users from underlying digital infrastructure; second, the scaling and pricing strategies that turn in-house infrastructure into a commercial service; and third, the market devices responsible for pricing the unit of compute.

Conclusion

This paper contributes to the critical political economy of cloud computing, providing a genealogy of Big Cloud's long-term strategy to seed and then to dominate that industry. It has examined how dominant commercial cloud providers have translated their investments in infrastructure into hybrid cloud assets that bolster these firms’ financial value and market dominance. Cloud assets are our analytic lens to highlight the way that tangible computing infrastructure requires epistemic and accounting practices in order to become valuable. This is our contribution to the debate over the role of tangible assets (e.g. owning data centers) versus intangible assets (e.g. owning platforms) in cementing the power of a few large technology companies.

As the basis for our arguments, we presented data from over a decade of financial filings of Amazon, Microsoft, and Alphabet/Google, who currently each spend US$20-$30 billion per year on physical infrastructure. However, how one gets from a costly data center investment to a price for a 10-s AI-generated video (for example) is anything but straightforward or deterministic; instead, it is highly contingent. As we have conceptualized it, this translation requires a particular technique of abstraction: the construction of a generic “unit of compute.” That is, monetizing the computational power is a complex epistemological process involving metrics, standards, and codes. And we have argued that making this translation successfully entails constructing cloud assets. Raw hardware must be configured as an asset in order to generate revenue. And so, to understand economic power and financial value in cloud markets, this case suggests that it is helpful to avoid imposing a binary between tangible and intangible asset categories.

We have demonstrated the importance of durability, virtualization, pricing models, and the construction of revenue streams to the assetization of cloud computing. First, durability entails an orientation to a temporal horizon in which political-economic objects are rethought as income-generating resources with a concern for extending infrastructure lifespans (Mitchell, 2020; Pistor, 2019). Second, virtualization entails the layering of intangible assets and tangible assets. This depends upon the establishment of distinct compute instances; that is, temporally defined units of compute with distinct prices (from generic computing provision). Third, much of the infrastructure investment underpinning virtualization is, therefore, tied to pricing strategies that are built on the back of contractual arrangements (Birch and Muniesa, 2020). Storage and computing capacity are supposed to be flexible as a result of the modular techno-economic configuration of cloud assets (Birch and Bronson, 2022); however, contractual arrangements and fee structures create highly complex pricing schemes with differential access for customers. Last, revenue streams are constituted through the differentiation between types of “instances” (e.g. on-demand vs spot) and temporal commitments (e.g. reserved vs on-demand contracts). Finally, through what we call the “digital Jevons effect,” Big Cloud's large-scale investments have driven down the relative price of the unit of compute sufficiently that it opens up new use-cases for their customers and facilitates the wholesale migration onto the cloud of the computing infrastructures that these customers once owned.

Further research is required to assess what reconfigurations of power follow from consolidated ownership and control of the unit of compute. Implications of the cloud asset concept in particular pertain to how large AI companies might be regulated (Burkhardt and Rieder, 2024): that is, how to govern firms whose market power and financial value are co-produced through data center holdings, computing contracts (e.g. between Microsoft and OpenAI), and hoarding of expertise (e.g. Facebook paying $10m bonuses for LLM programmers)—which are each typically treated in isolation. Still, what is clearly at stake in the political economy of digital capitalism is not only ownership of individual consumer data, or ownership of data centers per se, but control over the supply and monetization of the unit of compute that companies, governments, universities, and other organizations increasingly depend on.