Abstract

In the age of machine learning, cryptocurrency mining, and seemingly infinite data storage capacity enabled by cloud computing, the environmental costs of ubiquitous computing in modern life are obscured by the sheer complexity of infrastructures and supply chains involved in even the simplest of digital transactions. How does computation contribute to the warming of the planet? As information technology (IT) capacity demands continue to trend upward, what are some of the ecological obstacles that must be overcome to accommodate an ever-expanding, carbon-hungry Cloud? How do these material impacts play out in everyday life, behind the scenes, where servers, fiber optic cables, and technicians facilitate cloud services? This case study draws on firsthand ethnographic research in data centers—sprawling libraries of computer servers that facilitate everything from email to commerce—to identify some of the far-reaching and tangled environmental impacts of computation and data-storage infrastructures. It surveys a range of empirical accounts of server technicians to illustrate on-the-ground examples of material and ecological factors that permeate everyday life in the Cloud. These examples include air conditioning and thermal management, water cycling, and the disposal of e-waste. By attending to the culture of workplace practice and the behaviors and training of technicians in data centers, this case study reveals that the Cloud is not fully automated, nor is it hyperrational; emotion, instinct, and human judgment are enlisted to keep servers running. This case study closes with a speculative vignette that scales up from various local impacts to a planetary framework, sketching some of the particular ways that computation contributes to climate change and the Anthropocene.

Keywords: climate change, Anthropocene, data centers, data storage, digital ecology, materiality of computation, sustainable computing

Steven Gonzalez MonserrateProgram in History, Anthropology, and Science, Technology, and Society, MIT

Learning Objectives

• Learners will be able to identify the various material and ecological impacts of computation and digital data storage practices.
• Learners will recognize the environmental toll of digital life and the complexity of infrastructures involved in its operation.
• Learners will understand some of the ways that user behavior and cultures of computing influence efficiency and sustainability outcomes.
• Learners will be able to apply a holistic, humanistic approach to sociotechnical challenges.

Introduction: Materializing the Immaterial

Screens brighten with the flow of words. Perhaps they are emails, hastily scrawled on smart devices, or emoji-laden messages exchanged between friends or families. On this same river of the digital, millions flock to binge their favorite television programming, to stream pornography, or enter the sprawling worlds of massively multiplayer online roleplaying games (MMORPGs), or simply to look up the meaning of an obscure word or the location of the nearest COVID-19 testing center. Whatever your query, desire, or purpose, the internet provides, and all of the complexity of everything from unboxing videos to do-it-yourself blogs are contained within infinitely complex strings of bits. As they travel across time and space at the speed of light, beneath our oceans in fiber optic cables thinner than human hairs, these dense packets of information, instructions for pixels or characters or frames encoded in ones and zeros, unravel to create the digital veneer before you now. The words you are reading are a point of entry into an ethereal realm that many call the “Cloud.”

While in technical parlance the “Cloud” might refer to the pooling of computing resources over a network, in popular culture, “Cloud” has come to signify and encompass the full gamut of infrastructures that make online activity possible, everything from Instagram to Hulu to Google Drive.1 Like a puffy cumulus drifting across a clear blue sky, refusing to maintain a solid shape or form, the Cloud of the digital is elusive, its inner workings largely mysterious to the wider public, an example of what MIT cybernetician Norbert Weiner once called a “black box.”2 But just as the clouds above us, however formless or ethereal they may appear to be, are in fact made of matter—water molecules in various states of condensation and crystallization—the Cloud of the digital is also relentlessly material.

To get at the matter of the Cloud we must unravel the coils of coaxial cables, fiber optic tubes, cellular towers, air conditioners, power distribution units, transformers, water pipes, computer servers, and more.3 We must attend to its material flows of electricity, water, air, heat, metals, minerals, and rare earth elements that undergird our digital lives. In this way, the Cloud is not only material, but is also an ecological force. As it continues to expand, its environmental impact increases, even as the engineers, technicians, and executives behind its infrastructures strive to balance profitability with sustainability. In my experience, nowhere is this dilemma more visible than in the walls of the infrastructures where the content of the Cloud lives: the factory-libraries where data is stored and computational power is pooled to keep our cloud applications afloat.4

In this case study, I bring you into the beating heart of the digital, into the blinking corridors of data centers (or server farms) that make digital industry possible. As an anthropologist, I approach the study of the Cloud holistically, taking seriously the technological and material aspects of computation and data storage, while also attending to the ways that the Cloud is a social and cultural formation. In what follows, I draw on five years of qualitative research and ethnographic fieldwork in North American data centers to illustrate some of the diverse ecological impacts of data storage and some of the sociocultural factors that influence the sustainability of digital infrastructures.5 I also provide a broad, introductory overview of some of the fast-moving and evolving literature on the material impacts of computation and data centers from a range of disciplines including computer science, engineering, media studies, and more.

Discussion Question: Where is the Cloud? Using the resources below, describe how you might “locate” the Cloud: What are the human and technological infrastructures that bring it about? What histories are relevant to narrating the “location” you sketch?

Cloud the Carbonivore

It is four in the morning when the incident occurs. At that moment, I am crouched on the floor of one of the containment aisles of the data center, computers arrayed like book stacks in a library on either side of me. The clamor of server fans makes it nearly impossible for me to hear Tom, the senior technician I am shadowing, explain to me how to pry open a faulty floor tile. With a specialized tool, I remove the white square tile from its hinges, noticing tiny perforations etched on its surface, points of ingress designed to help cool air rush up from a vast, pressurized cavity beneath us called a “plenum.” I set the tile aside, feeling a rush of cold tickle my nose as a gust of chill whips up from the exposed underfloor plenum. I go about replacing the tile, using one with more notches to improve airflow to this particular cluster of dense computing equipment. That is when I hear the alarms go off. Amid a sea of blinking green and blue lights, an entire rack of computers suddenly scintillates yellow, and then, after a few seconds, a foreboding red. In that instant, panic sweeps over Tom’s face, and he too is flush and crimson as he scrambles to contain the calamity unfolding around us.“They’re overheating,” Tom says, upon inspecting the thermal sensors, sweat dripping from his brow.I feel the heat swarming the air. The flood of warmth seeps into the servers faster than the heat sinks printed onto their circuit boards can abate, faster than the fans can expel the hot air recycling in a runaway feedback loop of warming. The automatic shutdown sequence begins, and Tom curses, reminding me that every minute of downtime, of service interruption, may cost the company many thousands of dollars. Within two minutes, however, the three massive air conditioning units that had been idling in a standby state activate to full power, flooding the room with an arctic chill and restoring order to the chaotic scene.

In the vignette above, which draws on my ethnographic fieldnotes, I recount an episode that data center technicians refer to as a “thermal runaway event,” a cascading failure of cooling systems that interrupts the functioning of the servers that process, store, and facilitate everything online (Figure 1) The molecular frictions of digital industry, as this example shows, proliferate as unruly heat. The flotsam and jetsam of our digital queries and transactions, the flurry of electrons flitting about, warm the medium of air.6 Heat is the waste product of computation, and if left unchecked, it becomes a foil to the workings of digital civilization.7 Heat must therefore be relentlessly abated to keep the engine of the digital thrumming in a constant state, twenty-four hours a day, every day.

Figure 1 A data center technician. (Original photograph by the author; artistic rendering applied by the author to preserve the technician’s anonymity.)

To quell this thermodynamic threat, data centers overwhelmingly rely on air conditioning, a mechanical process that refrigerates the gaseous medium of air, so that it can displace or lift perilous heat away from computers.8 Today, power-hungry computer room air conditioners (CRACs) or computer room air handlers (CRAHs) are staples of even the most advanced data centers. In North America, most data centers draw power from “dirty” electricity grids, especially in Virginia’s “data center alley,” the site of 70 percent of the world’s internet traffic in 2019.9 To cool, the Cloud burns carbon, what Jeffrey Moro calls an “elemental irony.”10 In most data centers today, cooling accounts for greater than 40 percent of electricity usage.11

While some of the most advanced, “hyperscale” data centers, like those maintained by Google, Facebook, and Amazon, have pledged to transition their sites to carbon-neutral via carbon offsetting and investment in renewable energy infrastructures like wind and solar, many of the smaller-scale data centers that I observed lack the resources and capital to pursue similar sustainability initiatives.12 Smaller-scale, traditional data centers have often been set up within older buildings that are not optimized for ever-changing power, cooling, and data storage capacity needs. Since the emergence of hyperscale facilities, many companies, universities, and others who operate their own small-scale data centers have begun to transfer their data to hyperscalers or cloud colocation facilities, citing energy cost reductions. According to a Lawrence Berkeley National Laboratory report, if the entire Cloud shifted to hyperscale facilities, energy usage might drop as much as 25 percent.13 Without any regulatory body or agency to incentivize or enforce such a shift in our infrastructural configuration, there are other solutions that have been proposed to curb the Cloud’s carbon problem. Some have proposed relocating data centers to Nordic countries like Iceland or Sweden, in a bid to utilize ambient, cool air to minimize carbon footprint, a technique called “free cooling.”14 However, network signal latency issues make this dream of a haven for green data centers largely untenable to meet the computing and data storage demands of the wider world.

As a result, the Cloud now has a greater carbon footprint than the airline industry.15 A single data center can consume the equivalent electricity of fifty thousand homes.16 At 200 terawatt hours (TWh) annually, data centers collectively devour more energy than some nation-states.17 Today, the electricity utilized by data centers accounts for 0.3 percent of overall carbon emissions, and if we extend our accounting to include networked devices like laptops, smartphones, and tablets, the total shifts to 2 percent of global carbon emissions.18

Why so much energy? Beyond cooling, the energy requirements of data centers are vast. To meet the pledge to customers that their data and cloud services will be available anytime, anywhere, data centers are designed to be hyper-redundant: if one system fails, another is ready to take its place at a moment’s notice, to prevent a disruption in user experiences. Like Tom’s air conditioners idling in a low-power state, ready to rev up when things get too hot, the data center is a Russian doll of redundancies: redundant power systems like diesel generators, redundant servers ready to take over computational processes should others become unexpectedly unavailable, and so forth. In some cases, only 6–12 percent of energy consumed is devoted to active computational processes.19 The remainder is allocated to cooling and maintaining chains upon chains of redundant fail-safes to prevent costly downtime.

That being said, there are two computational processes performed by servers that are particularly energy-intensive and are of increasing concern to scholars, activists, and data center industry professionals: 1) machine learning and 2) cryptocurrency mining. In a study conducted at the University of Massachusetts, Amherst, PhD candidate Emma Strubbel determined that training a handful of artificial intelligence models can emit over 626,000 pounds of carbon dioxide, as much as five American automobiles do over their lifespans.20 In this way, computation is metabolic: to maximize returns on computational processes, energy inputs must match intensity in the same way that tons of cooling (BTUs) must be matched to electricity curves (kwh) to prevent thermal runaway. Ironically, advances in machine learning have led to sustainability innovations in a number of industries and have advanced research to support environmentalist agendas.21 The question therefore becomes: how can AI reduce its ecological footprint? Can machine learning algorithms be designed to operate with greater energy efficiency? Some scholars and practitioners are working toward better understanding of precisely how machine learning contributes to greenhouse gas emissions, so that methods can be developed from a design and policy standpoint to mitigate those effects.22

Like AI and machine learning, the mining of cryptocurrency is a computationally intensive process with a growing ecological impact.23 Given the increasing computational complexity of blockchain operations, the average Bitcoin miner is no longer an MIT student experimenting with GPUs in a dorm room, but instead a person with enough resources to afford specialized, high-performance computers and the costly capital required to cool and host them.24 The energy requirements for cryptocurrency production are so high that miners only stand to profit on returns if the cost of energy where their computers are located is sufficiently cheap.25 For places like Iceland, with free cooling and a largely sustainable geothermal energy grid, high-performance computation is a “natural” fit.26 However, cheap energy is also available in places like China, where over 73 percent of electricity consumed by data centers is sourced from coal.27 Authorities on this matter debate the precise carbon footprint of cryptocurrency production, but estimates range anywhere from 20 to 115 TWh annually, around 0.33 percent of global electricity usage.28 Framed differently: the equivalent of one US dollar in Bitcoin requires over seventeen megajoules of energy to produce, which is double the amount of energy required to mine copper, gold, or platinum.29

Discussion Question: What are some of the challenges of scale (policy, design, culture) that stand in the way of bringing about a carbon-neutral Cloud? What kinds of interdisciplinary and practitioner conversations might bring about more sustainable machine learning or cryptocurrency production?

Precipitations

It is late July in Arizona. The sun is white and hot on this cloudless day. I feel it scorch the back of my neck as I follow Jeremy, a junior technician, to the backlot behind a data center, where dozens of shipping containers are arrayed in rows. Amid this 117-degree heat wave, our task is to repair an evaporative cooling system that is failing. We unfasten the screws on one of the exterior panels before entering the shipping container, which I am surprised to learn is actually a modular server cluster. Pipes snake up from tiny channels in the lot, where potable water is pumped up from the ground, to seep up into a spongey, filter media. To my eyes, this foamy material resembles a honeycomb or a wasp’s nest (Figure 2). The sediment-rich waters of the Colorado River have congealed to form an oozy soot on the porous surface that is not unlike honey. The wet tray of material evaporates quickly in the arid desert air, the roiling cloud of moisture gently cooling the loudly buzzing servers around us, Jeremy explains. This, I learn, is why the shipping container has the nickname, “The Mouth.”

Figure 2 Adiabatic cooling filter media. (Original photograph and artistic rendering by the author.)

The Cloud may be a carbonivore, but as the example of “The Mouth” shows, the Cloud is also quite thirsty. Like a pasture, server farms are irrigated. In many data centers today, chilled water is piped through the latticework of server racks to more efficiently cool the facility, liquid being a superior convective agent than air. This shift from cooling air to cooling water is an attempt to reduce carbon footprint, but it comes at a cost. Weathering historic drought and heat domes, communities in the western United States are increasingly strained for water resources. In Mesa, Arizona, where I spent six months researching the emergence of a desert data center hub, some politicians are now openly opposing the construction of data centers, framing the centers’ water usage as inessential and irresponsible given resource constraints.30 In Bluffdale, Utah, residents are suffering from water shortages and power outages, as a result of the nearby Utah Data Center, a facility of the US National Security Agency (NSA) that guzzles seven million gallons of water daily to operate.31

Figure 3 A data center in Arizona. (Photograph by the author.)

Figure 4 Noise pollution experiences: A wordcloud by the author, sampled from the author’s interview data with residents in Chandler, Arizona.

Figure 5 Computer graveyard in Agbogbloshie, Ghana. Source: Agbogbloshie Makerspace Platform. CC BY-SA 2.0

Figure 6 Data center technician. (Original photograph by the author; artistic rendering by the author to preserve the technician’s anonymity.)