Click the star to add/remove an item to/from your individual schedule.
You need to be logged in to avail of this functionality.
Log in
- Convenors:
-
Jean-Christophe Plantin
(London School of Economics and Political Science)
Devika Narayan (University of Bristol)
Send message to Convenors
- Format:
- Traditional Open Panel
Short Abstract:
This panel uncovers the set of technical, organizational, and discursive practices and resources that technology companies (including and beyond tech giants) mobilize to make scalability possible in the context of large computing networks.
Long Abstract:
Scholarship and public debates use proxies to convey the “hyperscale” of the infrastructure that supports large tech companies e.g., billions of users, petabytes of data, or millions square feet of data centers. However, such accounts tend to essentialize large scale by presenting it as an achieved and static situation. This panel aims instead to uncover the myriad of technical, organizational, and discursive practices and resources that tech-intensive firms mobilize to make scalability possible in the context of large computing networks.
It is crucial to interrogate these scale-making practices to first, critically demystify the common trope of rapid “scaling up” as an imperative for tech companies. Second, access to scalable computing resources via cloud computing permits rapid expansion of business models and firm practices, a phenomenon that calls for closer examination.
We build upon a strong STS tradition linking scale and power in the context of infrastructure, but also emphasizing the ideological and discursive of scale and the relevance of a multiscalar study of infrastructures. We combine this perspective with an emphasis on corporate and organizational practices as we see scale-making as a techno-organizational project shaped by the imperatives of accumulation and growth.
We invite submissions that reveal a plurality of scale-making practices in cloud computing and data centers. Topics comprise (but are not limited to):
How do organizations (tech giants and beyond) reach and maintain scale?
Who maintains computing infrastructure at scale, under what status, and for whose benefit?
How do organizations manage and utilize cloud computing and data center resources?
What are the limits and social implications of hyperscalability? What happens when scale breaks or is not possible?
Accepted papers:
Session 1Fabian Ferrari (Utrecht University)
Long abstract:
Cutting-edge AI systems like large language models demand extensive computational resources. However, the global cloud computing industry is extremely concentrated, with Amazon, Google and Microsoft accounting for two-thirds of this market. EU policymakers aim to foster the development of public alternatives to Big Tech’s computing facilities, exemplified by state-funded supercomputers. MareNostrum 5, for instance, the most capable machine of the Barcelona Supercomputing Cluster cost a total of 202 million euros of EU taxpayer money. This paper examines the political-economic context in which publicly funded AI infrastructure projects operate and compete with Big Tech. How can state-owned large-scale AI computing be theorized, and under which conditions is it feasible?
To answer this question, the key argument of this paper is to consider large-scale AI computing as a public utility – an institution that provides essential services to the public. This key argument unfolds in three steps. First, the paper dissects the relationship between applications, models, and infrastructure in the context of large language models. Second, it uses this groundwork to probe the applicability of public utility thinking on AI infrastructure, resorting to three analytical registers of public utility thinking: stifled competition, downstream effects, and state-supported innovation. Analyzing those drivers is paramount to distinguishing large-scale AI computing from other public utilities, such as electricity, water, and sewage systems. Third, the paper stresses the importance of articulating the “infrastructural state” as a normative horizon in AI governance debates, going beyond a narrow regulatory focus on surface-level problems such as misinformation and deep fakes.
Nathan Kim (University of Michigan School of Information)
Long abstract:
This paper charts the rise of the software container, or the packaging of software into deterministic and portable environments. Through containers and container management technology like Docker and Kubernetes, software at a high level can be segregated, hidden, moved, and interfaced with in a standard and replicable manner. I contend that just as shipping containers made possible a tenfold decrease in transport cost and a thirtyfold increase in global shipping and the circulation of capital, so too does the software container's offerings of standardization and modularity enable the massive and opaque level of scale that characterizes the cloud today. To articulate this claim, I trace the history of container technology development alongside the history of the cloud, beginning with the container's prehistory in the Unix operating system through its exploding popularity in the 2010s. Through an examination of development work in open-source communities like the Cloud Native Computing Foundation and the Linux kernel that facilitate container development, I also argue that the intrinsically exclusive drive to scale is closely linked to discourses of openness and the commons, a culmination of Silicon Valley's understanding of freedom as an entrepreneurial endeavor. I conclude with a call for critical scholars of technology to not only consider the infrastructural components of platform capitalism but also its logistical aspects -- the technologies and techniques coordinating the circulation of capital.
Devika Narayan (University of Bristol)
Long abstract:
How does a cloud-based asset regime impact scaling-making? This presentation outlines two ways to answer this. First, is the speed and swiftness by which firms emerge, expand, and consolidate their position – in markets that not long ago did not even exist. That is, this asset regime is associated with the rapid capture of new markets. This is a temporal observation about rapid scaling and growth. Second, this asset regime grants firms the latitude to experiment and enact strategy pivots without prohibitive sunk costs and other risks. Leveraging this asset regime potentially mean firms can introduce new services, products, and devise new data- and platform- distribution mechanisms without long gestation periods, complex contractual linkages, or capital-intensive investment cycles. This then is an argument about accelerated industrial change. This presentation offers a dual lens through which to connect cloud computing and scale-making. First is its association with rapid growth (firm-level) and second with accelerated industrial disruption (market-level).
Boxi Wu (Oxford Internet Institute, Oxford University) Vili Lehdonvirta Zoe Hawkins (Australian National University)
Long abstract:
Besides human capital and training data, frontier AI development requires considerable "AI compute” resources produced with graphic processing units (GPUs). As a result, the location and ownership of such GPU clusters has become a geostrategic priority for governments. In this study we ask (1) how GPU clusters are distributed globally and (2) under whose ownership they are. We focus on so-called “public cloud” compute provided by US hyperscale cloud providers Amazon Web Services, Microsoft Azure, and Google Cloud, as well as their Chinese competitors Alibaba, Tencent, and Huawei. Our methodology is based on collating information from public sources, including online archives, and on interviewing experts on AI compute. We collect an original dataset on the locations of hyperscale "cloud regions" by provider, and identify the subset of these regions that feature GPUs relevant to frontier AI development. We also assess the approximate sizes of the GPU clusters belonging to different providers. We find that the geography of AI compute-enabled cloud regions is more concentrated than general cloud computing regions. The political geography is also highly polarized: a plurality of such cloud regions are located in the US, followed by China and a small number of other countries, with most countries hosting no AI compute. We discuss the potential economic and political reasons behind the observed patterns as well as their potential implications to international relations, AI governance, and digital competition.
Jacob Hellman (University of California, San Diego)
Long abstract:
“Big Cloud” — the three firms (Amazon, Microsoft, and Google) controlling two-thirds of the cloud computing market — spend annually US$20-30 billion each investing in land, buildings and hardware for data centers. Meanwhile, when in 2023 the US Federal Trade Commission solicited public comments on business practices of cloud computing providers, concern focused on anticompetitive strategies like customer lock-in fees — rather than unprecedented concentration of large-scale data center infrastructure. Yet Big Cloud’s current dominance is a result of individual long-term strategies of engineering a massive base of tangible infrastructure. Based on analysis of corporate earnings calls, this paper shows that these firms have worked to bring the cloud computing market into being, in order to dominate it. As Narayan (2022, Economy and Space) points out, for cloud-dependent companies, “scaling” crucially entails the ability to expand and contract on-demand, opening up second-order strategies of accumulation. The commercial cloud enables this by pooling users to maximize the amortization of infrastructure investments. Under the current AI cost crunch and GPU shortage, this efficiency-profitability rational has migrated from financial reporting to popular discourse about Big Cloud “making accessible” or even “democratizing” AI — for other for-profit companies, who’d otherwise find it prohibitively expensive. From an efficiency standpoint, cloud computing is a natural monopoly. Not everyone needs private racks of GPUs. Yet in lieu of socializing it, when how do we think about the implications of its inevitable concentration for sociotechnical power? This paper offers some provisional observations in light of “diseconomies” of scale.
Yung Au (University of Oxford) Srujana Katta (University of Oxford) Amelia Hassoun (University of Cambridge)
Long abstract:
In the past decade, an increasing number of public and private actors have laid claim to outer space through various speculative data projects, from Amazon’s space-based content delivery network to Thales’ plans for a lunar data centre. This paper explores these infrastructural developments and what they reveal about the incoherent fantasies of hyperscaled and hyperoptimised computing. The exaggerated and pre-emptive claims of particular data futures foreground, in particular, the continual work done to uphold specific capitalist and colonial logics underlying many expansionist projects today. These speculative infrastructural claims furthermore intersects with the longstanding discourse on the multi-scalar examination of infrastructures. The vastness of outer space and its disruption of human-centric measures of scale invites reflection on concepts such as local, regional, global, interplanetary, and intergalactic—and thus of governance and how we approach the idea of the commons. As infrastructure studies has shown, all infrastructure brings into question of who benefits and who does not; who has a say and who are excluded. Learning from infrastructural injustices in our terrestrial world, such as railways that were instruments of land grabs, to the telecommunication structures that perpetuates colonial relationalities today, this paper examines what cosmic infrastructures can tell us about the power struggles underlying hyperscale computing dreams of tomorrow.
Jean-Christophe Plantin (London School of Economics and Political Science)
Long abstract:
It is now public knowledge that US-based tech giants (e.g., Amazon, Google, Meta) have built massive networks of data centers linked by powerful subsea cables. Scholarship and the press have revealed the large number, size, and capacity of this “hyperscale” infrastructure.
In this paper, I shift the attention from size to scale and detail scale-making practices in Meta’s data centers. I reveal Meta’s efforts to scale up its computing infrastructure, starting with the smallest data center units (server), working its way up to the cabinets containing the servers, and then to the architecture of data centers.
First, Meta endorsed a widespread critique in the industry of the black-boxed model for server manufacturing from traditional firms (HP, IBM) and launched an industry consortium—open compute project (OCP)—to promote open-source hardware in the data center industry. This strategy allowed Meta and other hyperscalers to decide their standards for the design, power supply, and maintenance model for data center hardware, all designed with “hyperscale” operations in mind.
Second, the consortium designed the next steps for scaling up the adoption of such components. First, developing more powerful servers necessitated the creation of larger server cabinets. Tech giants developed and promoted via this consortium the 21-inch rack—as opposed to the industry standard of a 19-inch rack. One level up, adopting such racks (which are taller and heavier) required the design of a new architectural standard for data centers—which became the certificate OCP Ready™—allowing data center owners to adapt their architecture to accommodate such gear.
Agustin Ferrari Braun (Universiteit van Amsterdam)
Long abstract:
News media publishers are tech-intensive organisations that have adopted cloud computing as a means to scale-up both their reach and the amount of content they deliver to their audiences. While scope, business models, and providers vary, the virtual totality of the sector has integrated cloud infrastructure. Doing so required significant changes in resource allocation, ranging from hiring specialised workers and developing in-house expertise to the adaptation of workflows to the new digital environment. The implementation of cloud infrastructure in news publishers is a structural organisational change that has durable consequences for the conditions of production of journalism. Drawing from semi-structured interviews with workers of public and private Dutch news media publishers, this presentation explores the knowledge regimes framing the adoption of cloud solutions within these institutions. It argues that cloud technologies are regularly situated within scale-making discourses and practices that emerged in the 1980s in response to the liberalisation of the news market and the emergence of the 24/7 news cycle. These knowledge regimes emphasize the production, acquisition, and broadcasting of news-related content as a strategy to attract subscribers and advertisers. Rather than a transformative infrastructural development, cloud computing is presented as a new tool to reach audiences. In so doing, news media organisations foster institutional continuity while managing infrastructural disruption. However, by deploying decades-old discourses and practices, they also deprive themselves of the capacity to critically challenge scale-making assumptions logged within cloud computing solutions that were not originally designed for news media publishers.
Inga Luchs (Rijksuniversiteit Groningen)
Long abstract:
Google’s search engine, Facebook’s social network, Amazon’s marketplace – these and other platforms have been comprehensively analyzed for the socio-economic transformations they have brought about in the last decades. However, while these platforms are fundamentally powered by AI technologies, the growing platformization of AI itself has not been in focus of attention so far. Yet, it is similar discursive practices big tech is mobilizing to sustainably shape AI development. Under the discourse of a “democratization” of AI, beginners in AI development increasingly become the users of corporate cloud platforms, leading to the expansion of big tech’s infrastructural power. My contribution focuses on Google AI’s democratization discourse as offering “tools for everyone”. I detail this discourse via Google’s ML platform Vertex AI which provides users with tools for ML development and addresses advanced and less experienced developers with a low-code approach. By means of a critical discourse analysis of the platform’s documentation, I illustrate how Google presents the platform as simplifying AI development and empowering users in shaping the future of AI. I clarify how this democratization rather needs to be seen as an establishment of corporate AI products and as strategy to meet the increasing need for AI talent. Further, Google particularly enables generative AI, ensuring businesses to become reliant on its vast data and computing resources. Consequently, instead of a true participation in the design of AI applications, Google’s AI democratization only represents a commodified version of ML development which enforces capitalist logics of scale and data accumulation.
Désirée Waibel (University of Lucerne) Vera Tauber (University of Lucerne)
Long abstract:
Arguably, cloudification is regarded as a “boring” (Star 1999, 379) infrastructure yielding collaboration and synergism. In our contribution, we illustrate that cloudification creates an infrastructural “hinge” between different ecologies (Abbott 2005) on a global scale using testing as primary mode of coordination (Marres and Stark 2020). This leads to various disruptive re-configurations within librarianship.
To do so, we draw our ongoing research on the cloudification of Swiss academic libraries. Since 2020, a cloud-based library management system and a search portal are implemented by over 500 libraries via a national library network. The systems are licensed from and hosted by Ex Libris Group, the global market leader for library management systems and discovery tools. The cloud-based systems are characterized by three novel characteristics. Firstly, the systems are hosted by the global IT provider itself. Secondly, they are fluid (bugs are fixed and/or new features are introduced in monthly releases). Thirdly, they are hyper-scalable (Narayan 2022).
We start by characterizing the infrastructural linkage between Ex Libris and participating libraries of different national library ecologies, thus exemplifying how the operation of library infrastructures has evolved from an local endeavor to one of global scale. We then exemplify that testing is institutionalized as the primary mode of coordination between the ecologies and functions as an important practice for maintaining global scale. Finally, we illustrate how these multiscalar infrastructures disrupt and re-configure Swiss librarianship regarding social relations, workplace routines, and professional jurisdictions (including outsourcing of core missions to IT providers (Plantin and Thomer 2023).
Clément Marquet (Centre de Sociologie de l'Innovation, Mines Paris - PSL, i3)
Long abstract:
Scalability, security and connectivity are the main promises of the data center industry. By securing land and power, operators offer to their clients the possibility to thrive without caring for technical practicalities such as cooling of computer rooms, buying new inverters, withdrawing useless telecommunication cables or dealing with material breakdown. The efficiency of infrastructural invisibility comes at cost, which is paid locally: those facilities are more and more criticized for their power, water and land consumption, the low rate of job employment and a lack of attention of public bodies. However, far from being solely external actors’ concerns, these issues are also seen by data center operators as potential threats to their business activities and scale up promise.
This paper investigates how data accumulation challenges various actors (public services, IT corporations, colocation datacenters), questioning the material organization of their infrastructural investments. By focusing on practical operations such as cable management and server room cooling, I will show how technical and normative scaling up solutions lead a to the specialization of space and labor that progressively transformed the computer room into a modern data center facility, that is now reshaped by current hype for AI. Finally, I will claim that those scale-making practices profoundly transformed the nature data storage, challenging the relationship of the facilities with their institutional environment.
Angus Dowell (The University of Auckland)
Long abstract:
Cloud regions have emerged as dominant cloud computing configurations and are key to how big-tech companies expand and dominate markets across the globe. They are widely associated with hyperscale cloud computing, yet critical social research on its regionalised form remains scant. As researchers begin to grapple with the many scalar processes that make cloud computing powerful, the paper aims to shed light on the the significance of the 'region' as a supposedly central scale at which big-tech power operates. I do so by exploring a) what the cloud region is in practice, including its socio-technical and political economic construction and b) how the region operates in relation to wider scalar process, including economic development and security trajectories under state-tech relations. As a doctoral student, the paper is based on findings from my PhD on the making of the Amazon Web Services Hyperscale Region in Auckland, Aotearoa. I’m seeking feedback on my findings and how I might better make sense of them amid a growing academic interest in scale-making.
Andreas Baur (Universities of Amsterdam and Tübingen)
Long abstract:
The Gaia-X initiative has been launched to bolster European digital sovereignty and to contest the dominance of non-European cloud providers with their massive market influence, whilst cloud infrastructures are core resources of our societies and economies. Despite this aim, the initiative paradoxically incorporates these very dominant hyperscalers, arguing that European companies alone could not upscale enough and compete against economic and technological dependence. The initiative speaks to European attempts to challenge Big Tech, but it is not an EU project nor a regulatory initiative. Gaia-X has a unique character combining political and non-political elements and actors and aiming at infrastructuring digital sovereignty in the cloud sector through infrastructural development and governance structures.
The main questions to be answered centre around the extent Gaia-X contributes to advancing digital sovereignty in Europe. The initiative’s evolution unveils intricacies to the current landscape of technological regulation, with controversies shaping a distinct framework. This framework hinges on the interplay between territory and nationality within a context of borderless infrastructures. Here, a specific kind of hybrid governance emerges, seeking to navigate the balance between technological openness and closure, borders and nationality in a borderless infrastructure, and bridging diverse political and business interests. The study of the development of Gaia-X provides insights in the opportunities and challenges to achieve digital sovereignty and independence in a world dominated by US-American hyperscalers. This translates more generally about current and future infrastructure governance in Europe.
Alex Gekker (University of Amsterdam) Dan M. Kotliar (University of Haifa)
Long abstract:
Our paper examines the $1.2 billion tender won by Google (GCP) and Amazon (AWS) to facilitate “moving the state of Israel into the cloud” through two axes of scale. First - the national - measures the scalability of such undertakings through a recontextualisation of the state’s capacity to store, process and manipulate local data in tech giants’ clouds. Second - the global - challenges the scalability of GCP’s and AWS’s growing glocal “cloud regions” as both highly standardized and painstakingly dependent on existing historical assemblages. In doing so we reject the traditional hierarchical conceptualisation of scale in favour of a flat ontology (Marston, Jones, and Woodward 2005). By becoming “[a] node in others’ networks, both built in and left out” (Johnson 2019, 75) Nimbus defies easy categorisation in terms of power relations, environmental impact, or data extractivism. Growing its national component might reduce its effectiveness as a participant in a tech cloud region and vice versa. In such unpacking, we aim to resist the framing of scale as a teleological “scaling” used both in big tech but - perhaps more alarmingly - byits critics who nonetheless adopt such presuppositions (Hanna and Park 2020). Instead, we propose to think of scale as multidimensional and often-times necessary, as for when good public governance is required to leverage limited resources during a time of crisis.