12 Min Read
July 30, 2024
By
Thomas Davis, an Associate Professor at Ohio State University, specializes in energy, environmental humanities, and literary history. Author of, The Extinct Scene and is working on Forming Attachments: Aesthetic Education and Ecological Crisis. Scott Stoneman teaches Communication Studies at Mount Saint Vincent University. He is the host of Pretty Heady Stuff, producer of Volatile Trajectories, and co-authored Widening Scripts: Cultivating Feminist Care in Academic Labour.
In Google’s widely-covered 2024 Environmental Report, the company disclosed that its greenhouse gas emissions are up by 48% from 2019, reaching 14.3 million metric tons of carbon dioxide equivalent in 2023. Google expects their emissions to keep rising and blames artificial intelligence for the spike: “Reducing emissions may be challenging due to increasing energy demands from the greater intensity of AI comput[ing] and the emissions associated with the expected increases in our technical infrastructure investment.”1 Note that the inevitability of these rising emissions and investments is presented as a necessary consequence of profiting from AI.
OpenAI CEO Sam Altman has acknowledged the extraordinary energy resources AI will require. He explained in Davos in January 2024 that AI would need an energy “breakthrough,” and he had his eye on nuclear fusion as the climate-friendly solution to AI’s energy needs.2 A few months later, Altman told an audience at Harvard that the “huge amount of energy” required by AI is irrelevant because the technology is destined to “figure out” solutions that will offset its carbon intensity. With the swagger of a snake oil salesman, he declared: “We’re going to power up stuff… in all of these fantastic ways.”3
Discussions of AI’s utopian and apocalyptic potential often sound like speculative fiction, but its material needs and ecological consequences are very real. It is all too common to read anxious statements from those within the AI world about its risk of undermining human civilization, but much rarer to see serious consideration of the natural resources that are driving the current expansion of AI, even as it becomes clear that the energy and water consumption required to “train” and operate this technology is making it a major contributor to our ecological crisis.
In this brief essay, we want to look at the discourse around the accelerated development of AI, some of the speculative imaginings that guide it, and, most importantly, the underappreciated fact of its materiality. Altman may condemn us for having a “deeply unproductive streak,” but we are invested in foregrounding Big Tech’s destructive ecological effects and opening up a conversation about how to model an anticapitalist de-escalation of the current AI arms race.4 Technophiles, or as Yannis Varoufakis would say, technofeudalists like Sam Altman, present AI as both extraordinarily powerful and technologically novel. Yet, what we find emerging in the reports and scholarship on AI’s materiality strikes us not as something radically new but as familiar contradictions, even if in amplified form, of the accumulation of capital, specifically in its extractive relation to nonhuman nature and the exploitation of human labor in the Global South.
We can begin with AI’s energy problem as one instance where its promise to reshape the world runs aground. As with any profit-driven “post-industrial” project, the competition to dominate the AI market is ridiculously resource-intensive.5 In a stunning investigative report, Kate Crawford notes that while the “full planetary costs of generative AI are closely guarded corporate secrets,” projections suggest that “the demand for water for AI could be half that of the United Kingdom by 2027” and that very soon “large AI systems are likely to need as much energy as entire nations.”6 ChatGPT alone “is already consuming the energy of 33,000 homes.”7 According to one analysis, the initial work of “training a single big language model is equal to around 300,000 kg of carbon dioxide emissions,” or roughly “125 round-trip flights between New York and Beijing.”8 And that’s just to get one piece of AI software up and running. Steven Gonzalez Monserrate reports that one “data center can consume the equivalent electricity of 50,000 homes.” Globally, data centers use “200 terawatt hours (TWh) annually,” meaning they collectively use more electricity than most countries.9
Despite the airy promises that AI will enable decarbonization and increase energy efficiency, AI infrastructure currently needs so much energy that it is extending the life of coal plants slated for decommission and inspiring Microsoft to move aggressively toward small modular reactors to power its investments in this new frontier of digital domination. AI is also being used extensively by the fossil fuel industry to locate new sites for resource extraction, optimize existing sites, and market its products. In fact, some reports “estimate that AI and other digital technologies could generate as much as USD 425 billion in value for the oil and gas sector by 2025.”10
Although calculating the total costs of AI’s energy consumption has been difficult because many of the larger companies have not made that data available, their water consumption is harder to hide. A recent AP News story noted that Microsoft disclosed a 34% increase in its water consumption (“nearly 1.7 billion gallons”) in a year, which researchers conclude is due primarily to the growth of AI.11 This finding, from a research team based at UC-Riverside, offers some clarification about the water consumption: “ChatGPT gulps up 500 milliliters of water (close to what’s in a 16-ounce water bottle) every time you ask it a series of between 5 to 50 prompts or questions.”12
The data centers swallowing all of this water are located in Iowa, which Microsoft and OpenAI both note is a strategic location because of the mild weather that assists with cooling the data centers and the large amounts of available water. Those data centers, in the month before GPT-4 was completed, used 6% of the district’s water, amounting to 11.5 million gallons.
Our point is not to dismiss the possible uses of AI but to emphasize that as long as AI operates within the capitalist world-system, it will replay and exacerbate the contradictions of capital that have made life unlivable for so many for so long. So what role, if any, might AI have in a post-carbon or a post-work future?
Beyond resource consumption, the social and economic impacts of AI are equally concerning. While Silicon Valley has taught us to accept that “digital solutions and technological management are the best and only responses to climate change,” the Critical Carbon Computing Collective (4C) examines how the AI hype cycle “conceal[s] who wins, loses, and profits from… new tech.”13 According to 4C, climate projects that require extensive computing are “prone to manipulations” due to their perceived technical complexity.14 For example, Bill Gates’ claims that machine learning can help address the climate crisis by genetically modifying beef cows illustrates how the tech industry uses the mystique around AI’s complexity to frame it as a kind of “new electricity.”
The bill of goods being sold to the public, as AI commandeers climate action, consists of the following sort of if-then statements: if solving the climate crisis is too complicated, then a supercomputer will serve up the answer. If we can’t figure out a political means to expedite thousands of building retrofits, then a machine learning algorithm can step in and leverage private and public data to make the built environment more legible to utilities and developers (and provide a solution to those that can afford it). If intermittent supply is an issue for renewable energy, then we can use machine learning to find a way to keep the power on. If crops are failing due to worsening drought, then we can deploy AI-driven precision agriculture to willfully boost production. If climate change is generating more unpredictable extreme weather, then AI will let us process the sublimely random sloshing of air and water around us and anticipate climate impacts with enough foresight to get people out of the way of danger. If data centers are demanding absurd amounts of energy, then we can avoid having to scale back or quash gratuitous practices (like Bitcoin mining or AI-generated media) by just asking an intelligent machine to magically abate those luxury emissions.
If we turn from AI’s energy hunger to its labor needs, we find again familiar instances of the racialization and exploitation of labor that historians and critics of capital will recognize. Artificial intelligence is emerging in a world where, as Aaron Benanav says, it is designed to “embody capitalist control” and “end our obligations to one another.”15 Like past booms, the so-called “golden decade of deep learning” aims to definitively free capital from any responsibility to labour, regardless of the resources, redundancies, or trauma required to achieve it. The energized enclosure of AI means that GAMAM’s ultrarich overlords have an opportunity to replace more workers and even abdicate responsibility for entire categories of work. Altman dismissed the negative social effects of AI in a meeting captured by Vauhini Vara:
an entrepreneur asked [Altman] when AI would start replacing human workers. Altman equivocated at first, then brought up what happened to horses when cars were invented. “For a while,” he said, “horses found slightly different jobs, and today there are no more jobs for horses.”16
But the use of human labor to train LLMs has thus far replayed the most horrific exploitations of global capital. Consider the recent reporting on the content moderators in Kenya who were hired to train OpenAI’s LLMS. These workers were tasked with sorting through and labeling text and images from nightmarish corners of the internet—content that included murder, torture, hate speech, bestiality, and child sexual abuse—to ensure that products like ChatGPT don’t reproduce it. The Kenyan workers for Sama, the firm OpenAI used to outsource this work, reported extraordinary psychological distress from the hours they spent reading and interpreting violent, traumatic content. In keeping with the extractive labor practices that have long defined capitalist exploitation in the Global South, these workers took home between $1.32 and $1.44 per hour, which ranks lower than what TIME reports as “the minimum wage for a receptionist in Nairobi at $1.52 per hour.”17 If Altman and others figure AI as a uniquely futuristic technology, its perpetuation of racial capitalism on a global scale is anything but novel.
Our point is not to dismiss the possible uses of AI but to emphasize that as long as AI operates within the capitalist world-system, it will replay and exacerbate the contradictions of capital that have made life unlivable for so many for so long. So what role, if any, might AI have in a post-carbon or a post-work future? Emma Strubell and her co-authors have pointed to the use of AI for optimizing solar energy through “nowcasting,” or short-term weather forecasts.18 They show how, by using historical data and geolocation, AI can help the production and storage of solar energy. AI can also theoretically be used at scale to detect pipeline leaks, ensure railway safety, enhance HVAC performance, and maximize energy use in buildings. But, the possibilities of AI for decarbonization depend less on the pace of technological advance than they do on the political and economic configurations in which such advances will unfold. What, for instance, would AI look like in a degrowth economy? Given that it is so incredibly energy-intensive, could AI ever be compatible with post-capitalist alternatives?
There is something seductive about AI tools becoming the literal deus ex machina solution for our crisis of energy addiction and our desires for a life emancipated from endless, exploited labor. But it pays to be critical about carbon and computing, especially when AI beckons us with visions of fully decarbonized automated luxury, yet its answers come with a steep price tag, both financially and ecologically. Achille Mbembe has observed that the reign of “algorithmic reason” in our century “goes hand in hand with the rise of mythoreligious-type reasoning.”19 By turning our attention to AI's materialities and political economy, we might find a way out of the discursive and conceptual cul-de-sacs of utopian promise and apocalyptic catastrophe that Altman and his like have steered us into. If we are to fold AI ever deeper into our everyday lives, we can and should demand more equitable, ethical, and environmentally-conscious AI infrastructure. That might include public ownership of AI infrastructure, democratic debate and control over AI’s energy requirements, and mandating open-source development across the board and across borders. Whatever the case, we should not assume that AI, or any technology crafted and deployed within the capitalist world-system, will solve the ecological and human problems that capital generates and requires for its very existence. Those assumptions are, at best, mythoreligious fantasies.
Notes:
