Cover: CMC Data Center in Tân Thuận, Hồ Chí Minh, Vietnam. By Linh.ttt
Te Li, Monthly Review || In spring 2023, Microsoft announced a multi-billion-dollar investment in OpenAI, framing the partnership as a leap toward a cleaner, smarter, and more efficient civilization. The imagery accompanying such announcements is invariably ethereal: luminous neural networks, weightless data streams, and algorithms dancing across frictionless digital space. Artificial intelligence (AI), in the dominant register of Silicon Valley and its media ecosystem, presents itself as the apotheosis of dematerialization—a technology so refined, so purely cognitive, that it has finally escaped the dirty, entropic world of steam engines, coal mines, and factory floors.
This article argues that such representations are ideological in the precise Marxist sense: they invert reality, presenting as immaterial a process that is profoundly and consequentially material. The training of GPT-4 is estimated to have consumed energy equivalent to the annual electricity use of thousands of households.1 A single query to a large language model requires approximately ten times the electricity of a standard Internet search.2 Microsoft’s water consumption increased by 34 percent in a single year, a surge its own environmental report attributed directly to the expansion of AI infrastructure.3 These are not incidental inefficiencies awaiting technical correction; they are structural necessities of a technology whose physical substrate—semiconductors, data centers, cooling systems, and transmission networks—is among the most resource-intensive infrastructure humanity has ever constructed.
The myth of digital dematerialization has a long genealogy. Since the 1990s, theorists of the “information economy” have argued that the shift from manufacturing to services, and from atoms to bits, would decouple economic growth from material throughput.4 The rise of AI has given this thesis a new and more powerful iteration: if previous digital technologies merely processed information, AI—so the argument runs—generates intelligence itself, a resource whose abundance does not deplete nature but transcends it. This is the vision animating the rhetoric of “AI for climate,” “AI for sustainability,” and the broader claim that computational power can substitute for natural resources in solving the ecological crisis.
The theoretical framework developed in this article challenges this vision at its foundations. I draw on the thermodynamic tradition within ecological political economy, from Nicholas Georgescu-Roegen’s foundational The Entropy Law and the Economic Process to the ecosocialist synthesis developed by John Bellamy Foster and Brett Clark, to argue that the rise of AI under capitalism represents not a transcendence of the material world but an intensification of capital’s entropic relationship with it.5 The second law of thermodynamics is unsparing in its universality: every computation is a thermodynamic event. It consumes low-entropy energy—ordered, usable, free to do work—and returns high-entropy waste to the biosphere in the form of heat, carbon dioxide, and degraded matter. No algorithm, however elegant, suspends this law. The question is not whether AI produces entropy, but how much, at what rate, and which ecosystems absorb the consequences.
Karl Marx understood production as a metabolic process—a continuous interchange between human societies and the natural world, mediated by labor and technology. In Capital, he observed that machinery does not create energy but transforms and transmits the natural forces embedded within it, and that this transformation always involves the consumption of natural substance.6 What he could not have anticipated was a form of capital accumulation in which the primary productive force is computational power, and in which the thermodynamic demands of that power would place unprecedented stress on planetary energy systems, freshwater reserves, and climate stability. The rise of AI confronts us with the need to extend Marx’s metabolic analysis into the digital domain.
The Dynamics of Accumulation—AI as a High-Entropy Engine
The ecological crisis of AI is not, in the first instance, a thermodynamic problem. It is a social and historical one. The specific processes that are generating AI’s ecological burden—the competitive arms race among a handful of technology monopolies, the imperative to scale computational capacity regardless of social utility, and the systematic externalization of ecological costs onto communities and ecosystems in the Global South—are products of a particular historical formation: capitalism in its monopoly-digital phase. Thermodynamics does not cause these processes; it registers their consequences. The second law of thermodynamics tells us that every computation degrades energy. It does not tell us why computation is organized on this scale, at this rate, for these purposes, and at whose expense. For that, we need a social and historical analysis. What the thermodynamic framework provides, and what makes it indispensable, is a precise account of why the ecological damage generated by these social processes is not incidental but structural, not correctable by efficiency improvements but compounding, and not reversible by market mechanisms but permanent.
To understand why AI is thermodynamically costly in a structurally necessary rather than contingent way, it is essential to examine the relationship between computational scaling, capital accumulation, and energy consumption. This relationship is governed by dynamics that make the ecological burden of AI under capitalism not merely large but self-reinforcing and expansionary.
The physical basis of AI computation is the processing of vast quantities of numerical operations—multiplications, additions, and comparisons—performed by specialized hardware at extraordinary speed. The energy required to perform these operations is not trivial. Emma Strubell, Ananya Ganesh, and Andrew McCallum’s landmark 2019 study estimated that training a single large natural language processing model with neural architecture search produces carbon dioxide emissions comparable to the lifetime emissions of five average U.S. automobiles.7 Subsequent model generations have been substantially larger. While technology firms have declined to publish comprehensive energy figures for their most recent systems, independent analyses suggest that training frontier models, such as GPT-4, consumed energy on the order of tens of gigawatt-hours, enough to power a small city for weeks.8
The relationship between model scale and computational demand is not linear but superlinear. Research on scaling laws in large language models has established that model performance scales approximately as a power function of the product of model parameters and training data.9 This means that incremental improvements in model capability require disproportionate increases in computational investment. To improve a model’s performance by a fixed percentage, the computational budget—and therefore, the energy consumption—must increase by a significantly larger percentage. This is the thermodynamic expression of diminishing returns: as frontier AI systems approach the limits of performance on benchmark tasks, each additional increment of capability extraction requires exponentially greater energy input. The entropic cost of intelligence, under current technological paradigms, rises faster than intelligence itself.
This dynamic is further amplified by the logic of capital accumulation. The AI industry is organized around a small number of large firms—Google, Microsoft, Amazon, Meta (Facebook), and their Chinese counterparts—whose competitive position depends on maintaining algorithmic superiority. In this context, computational capacity is not merely a production input but a strategic asset: the firm with greater computational resources can train larger models, attract more users, and accumulate more data, thereby reinforcing its market dominance. This creates what might be called a computational arms race, in which each firm is compelled to expand its AI infrastructure not because the marginal social benefit of additional computation justifies the marginal cost, but because the competitive logic of capital accumulation makes restraint tantamount to market exit.10 No individual firm can voluntarily limit its energy consumption without ceding ground to rivals. The result is a collective-action failure of historic proportions: the industry as a whole expands its energy footprint far beyond what any rational assessment of social need would require.
The mechanism through which this dynamic operates is illuminated by the Jevons Paradox, first identified by the English economist William Stanley Jevons in his 1865 analysis of British coal consumption. Jevons observed that improvements in the efficiency of steam engines—reductions in the amount of coal required to perform a given unit of work—did not reduce total coal consumption but accelerated it, because lower costs of energy use stimulated the expansion of energy-intensive economic activity.11 The paradox is not a quirk of Victorian political economy; it is a structural feature of capital accumulation, operative wherever efficiency gains lower the cost of a resource and thereby stimulate demand for its use.
In the AI sector, the Jevons Paradox operates with particular force. Successive generations of AI chips—from NVIDIA’s A100 to H100 to Blackwell architectures—have delivered dramatic improvements in computational efficiency, measured in operations per watt. Yet total energy consumption by AI infrastructure has risen continuously and steeply, because efficiency gains have reduced the cost of AI computation, stimulated the proliferation of AI applications, expanded the volume of inference operations, and accelerated the development of ever-larger models. OpenAI’s own analysis found that the computational requirements of frontier AI training runs doubled approximately every 3.4 months between 2012 and 2018—a rate of increase far exceeding the efficiency improvements delivered by hardware advances.12 The International Energy Agency projected in 2024 that data center electricity consumption could exceed 1,000 terawatt-hours annually by 2026, roughly equivalent to Japan’s entire national electricity demand.13
It is these concrete social and historical processes—the monopoly arms race, the Jevons dynamic, and the structural impossibility of voluntary restraint under competitive accumulation—that thermodynamics registers but cannot by itself explain. The non-equilibrium thermodynamics of Ilya Prigogine, developed most fully in Order Out of Chaos, provides the conceptual bridge between the social logic of capital and its physical consequences.14 Prigogine demonstrated that complex systems that are far from thermodynamic equilibrium—so-called dissipative structures—maintain their internal order by continuously importing low-entropy energy from their environment and exporting high-entropy waste. The living cell, the hurricane, and the flame are all dissipative structures in this sense: they sustain their internal complexity at the cost of increasing entropy in their surroundings. But Prigogine’s deeper insight, and the one most consequential for present purposes, is that the processes driving dissipative structures far from equilibrium are nonreversible. The entropy generated in the surroundings cannot be reclaimed; the degradation of the environment is permanent. This irreversibility is not a side effect of inefficiency, it is the thermodynamic signature of dissipative processes themselves.
The capitalist AI complex is a dissipative structure of this kind, but one for which scale, growth rate, and irreversibility are determined not by natural dynamics but by the imperatives of capital accumulation. It maintains the internal order of corporate profit, algorithmic optimization, and market dominance by continuously drawing down the low-entropy stocks of the biosphere—fossil fuels, freshwater, and mineral ores—and returning them as high-entropy waste: carbon dioxide, thermal pollution, and electronic refuse. The carbon dioxide emitted by data centers accumulates in the atmosphere on timescales of centuries. The aquifers drawn down by cooling systems replenish on timescales of millennia, if at all. The ecosystems disrupted by mineral extraction in the Congo Basin or the Atacama Desert do not return to their prior states when the mines close. What capitalism does, through the competitive logic of AI accumulation, is to drive these dissipative processes at a rate and scale that overwhelms the regenerative capacity of natural systems, locking in ecological damage that no future technological fix can undo. The order of the algorithm is purchased at the price of permanent disorder in the atmosphere, the watershed, and the soil.
This framing allows us to see what the techno-optimist narrative conceals: that the “intelligence” produced by AI systems is not a free gift of information technology but a thermodynamic product, extracted from nature at a cost that the market systematically fails to register. The financial accounts of technology firms record the revenues generated by AI services; they do not record the entropic burden imposed on ecosystems, communities, and climate systems by the energy and material flows that make those services possible. This is not an accounting error but a structural feature of capitalism’s relationship with nature—what Foster, Clark, and Richard York have called the “ecological rift”: the systematic separation of the costs of production from the sites and subjects that bear them.15
A further dimension of the thermodynamic analysis concerns the relationship between AI training and AI inference. Training, the process of optimizing a model’s parameters on large datasets, is computationally intensive but occurs once. Inference, the process of running a trained model to generate outputs, is individually less intensive but occurs continuously, billions of times per day, across the global deployment of AI systems. As AI is integrated into search engines, productivity software, health care diagnostics, legal research, financial analysis, and military targeting systems, the aggregate energy demand of inference grows in proportion to the scale of deployment. Goldman Sachs Research estimated that the energy demand of AI inference could exceed that of training within the current decade as deployment expands.16 This means that the ecological burden of AI is not a one-time cost of building the system but a continuous and growing tax on the energy and water budgets of the planet—a tax with a rate that increases with every new application, every new user, and every new round of capital accumulation in the AI sector.
The picture that emerges is one in which the ecological crisis of AI is produced not by thermodynamics as such, but by the specific social and historical processes of capital accumulation—the competitive arms race, the Jevons dynamic, and the systematic externalization of ecological costs—which together drive dissipative processes of Prigoginian nonreversibility at a planetary scale. Entropy is the measure of the damage; capital accumulation is its cause.
The Energy Rift: Power, Water, and Mineral Extraction
The thermodynamic analysis of the previous section establishes the structural logic of AI’s energy demands. This section examines the material reality of those demands across three dimensions: electricity, water, and critical minerals. Together, these three vectors of extraction constitute what we may call, adapting Foster and Clark’s concept of the metabolic rift, an “energy rift” specific to the digital age: a systematic disruption of the metabolic relationship between human technological systems and the natural cycles that sustain them, mediated by the spatial and social inequalities of global capitalism.17
Electricity: The Grid Under Siege
The most immediately visible dimension of AI’s ecological footprint is its demand for electrical power. Data centers—the physical infrastructure of AI, housing the servers that train models and process inference requests—are among the most electricity-intensive facilities in the modern economy. A large hyperscale data center of the kind operated by Google, Microsoft, or Amazon can consume between 100 and 500 megawatts of power continuously—comparable to the electricity demand of a mid-sized city. The expansion of AI has dramatically accelerated the construction of such facilities. Microsoft alone announced plans in 2024 to invest $100 billion in new data center infrastructure globally, with similar commitments from Google, Amazon, and Meta.18
The scale of this expansion is placing acute pressure on electrical grids in regions where data center construction is concentrated. In Northern Virginia, which hosts the largest concentration of data centers in the world, grid operators have warned that planned data center growth threatens to outpace the region’s electricity generation and transmission capacity, potentially requiring the construction of new fossil fuel generation to meet demand.19 In Ireland, data centers already account for approximately 21 percent of total national electricity consumption—a figure that the national grid operator projects could rise to 32 percent by 2031, crowding out renewable energy capacity intended for household and industrial decarbonization.20 In Singapore, the government imposed a moratorium on new data center construction between 2019 and 2022, citing energy constraints, before lifting it under pressure from technology firms.
The relationship between AI’s electricity demand and the energy transition is deeply contradictory. Technology firms have made high-profile commitments to powering their operations with renewable energy, and have invested substantially in power purchase agreements for wind and solar electricity. But these commitments are systematically compromised by three structural dynamics. First, the temporal mismatch between renewable energy availability—which is intermittent, dependent on wind and solar conditions—and data center demand, which is continuous and cannot be interrupted, means that renewable power purchase agreements frequently do not correspond to actual electricity consumption patterns. The electricity that flows through data center circuits at any given moment may be generated by natural gas, coal, or nuclear plants, regardless of what renewable contracts the firm has signed.21
Second, and more fundamentally, the growth in AI electricity demand is outpacing the expansion of renewable energy capacity. A 2024 analysis by the International Energy Agency found that projected data center electricity demand growth would consume a substantial share of new renewable generation in several major markets, effectively crowding out decarbonization in other sectors.22 Building renewable capacity to power AI does not add to the clean energy supply available for the broader economy; it absorbs clean energy that would otherwise displace fossil fuels elsewhere.
Third, the sheer reliability requirements of AI infrastructure have driven technology firms to seek long-term contracts for natural gas-fired electricity generation. Microsoft’s deal with Constellation Energy to reopen the Three Mile Island nuclear plant attracted considerable publicity, but less noticed was the broader pattern of technology firms signing capacity agreements with gas-fired generators to guarantee firm power supply.23 The ecological logic is stark: the expansion of AI is directly extending the economic life of fossil fuel infrastructure, locking in carbon emissions for decades to come.
Water: The Hidden Metabolism
If electricity is the visible face of AI’s ecological demands, water is its hidden metabolism. Data centers require vast quantities of freshwater for cooling, either through direct evaporative cooling systems that consume water as vapor or through the cooling of the thermoelectric power plants that supply their electricity. This water demand is structurally invisible in most public accounting of AI’s environmental impact, yet it represents one of the most serious and locally acute dimensions of the technology’s ecological footprint.
Pengfei Li and colleagues’ 2023 study provided the first systematic estimates of AI’s water consumption, calculating that training GPT-3 required approximately 700,000 liters of freshwater—enough to produce 370 BMW cars or 320 Tesla electric vehicles.24 For inference, the picture is similarly striking: the study estimated that a conversation of between twenty and fifty questions with ChatGPT consumes approximately 500 milliliters of water. Multiplied across millions of daily users, this represents an aggregate freshwater demand of extraordinary scale.
The corporate disclosure data confirm the trend. Microsoft’s 2022 environmental report revealed a 34 percent increase in global water consumption year-on-year, explicitly attributing the increase to AI infrastructure expansion.25 Google reported a 20 percent increase in water consumption over the same period.26 These are not marginal fluctuations; they represent a structural shift in the freshwater demands of the technology sector driven directly by the scaling of AI systems.
The geography of this water consumption is not neutral. Data centers are frequently sited in regions selected for cheap land, favorable tax regimes, and climatic conditions suitable for cooling, criteria that routinely lead technology firms to locate in areas of existing or emerging water stress. In the U.S. Southwest, data centers compete for water with agriculture and municipal systems in a region already facing severe drought conditions amplified by climate change. In Chile, technology firms have established data center facilities in and near the Atacama region, drawing on water resources in one of the world’s driest ecosystems—resources on which Indigenous Atacameño communities and small-scale farmers depend for survival.27 In the states of Telangana and Andhra Pradesh in India, proposed data center parks have faced local resistance over concerns about groundwater depletion in areas already experiencing agricultural water scarcity.
This spatial pattern reproduces, in the specific domain of digital infrastructure, the broader logic of what Rob Nixon calls “slow violence”—the gradual, dispersed, and temporally attenuated forms of ecological harm that do not register as events in the media or policy systems dominated by dramatic, instantaneous catastrophes.28 The drawdown of a regional aquifer by data center cooling operations occurs over years and decades, affecting communities whose water insecurity is already chronic and whose political voice is limited. It does not generate headlines. It does not appear in the sustainability reports of technology firms. But it is materially real, thermodynamically necessary, and structurally determined by the competitive logic of AI accumulation.
Minerals: The Extractive Foundation
The third dimension of AI’s energy rift is the extractive foundation of its hardware. The semiconductors, servers, storage systems, and networking equipment that constitute AI infrastructure require a complex array of critical minerals—lithium, cobalt, tantalum, neodymium, dysprosium, indium, gallium, and others—whose extraction involves severe and concentrated ecological damage, disproportionately borne by communities in the Global South.
The geography of critical mineral extraction maps almost precisely onto the geography of historical colonial extraction. The Democratic Republic of Congo supplies approximately 70 percent of global cobalt production, much of it from artisanal mines operating under conditions of severe environmental degradation and labor exploitation, including the widespread use of child labor.29 Bolivia, Chile, and Argentina—the “lithium triangle”—hold the majority of global lithium reserves, and their extraction involves the drawdown of saline aquifers in high-altitude ecosystems of exceptional ecological sensitivity. Rare earth elements, essential for the permanent magnets used in data center cooling fans and power systems, are concentrated in China, Myanmar, and the Democratic Republic of Congo, with processing operations generating radioactive and toxic waste streams.
The acceleration of AI hardware development compounds these extractive pressures through the logic of planned obsolescence. The competitive dynamics of the AI arms race require technology firms to continuously upgrade their hardware—replacing previous generations of GPUs and custom AI accelerators with newer, more powerful models on cycles of two to three years. This generates enormous quantities of electronic waste: discarded servers, GPUs, memory modules, and networking equipment containing toxic materials including lead, mercury, cadmium, and brominated flame retardants. Global e-waste generation reached 62 million metric tons in 2022 and is projected to grow to 82 million metric tons by 2030.30 A substantial proportion of this waste is exported, often in violation of the Basel Convention, to processing facilities in West Africa, South Asia, and Southeast Asia, where it is handled under conditions of severe health and environmental risk.
The concept of unequal ecological exchange has a long and contested history, rooted in the broader unequal exchange tradition and the Marxist critique of imperialism. Drawing on this rich intellectual lineage—which extends from classical theories of imperialism through dependency theory and world-systems analysis—scholars have progressively incorporated ecological dimensions into the analysis of North-South asymmetries.31 Clark and Foster’s contribution to this framework is grounded primarily in the critique of ecological imperialism: the recognition that the metabolic relationship between the Global North and the Global South is not merely an economic asymmetry but an ecological one, in which the periphery absorbs the environmental costs of core accumulation.32 This framework provides the theoretical foundation for understanding the global political economy of AI’s material metabolism.
These three vectors of extraction—electricity, water, and minerals—are not independent; they are interconnected dimensions of a single metabolic system organized by the imperatives of capital accumulation. Data centers require electricity, which requires power infrastructure, which in turn requires minerals and water. Cooling systems need water, which competes with agriculture and municipal supply, affecting food systems and human health. Hardware production requires minerals and mineral extraction, generating waste, presenting issues of disposal that create further ecological harm. The energy rift of AI is not a single break in the metabolism of nature but a cascading disruption across multiple ecological systems, coordinated by the invisible hand of capital and rendered invisible by the ideological apparatus of digital dematerialization.
The Thermodynamic Limits of Capital
The material evidence assembled in the preceding section points beyond the scale of the crisis to its structure. What the empirical record of electricity demand, water depletion, and mineral extraction reveals is not a series of independent market failures but a single systemic logic—one that requires theoretical, not merely technical, explanation.
The ecological crisis generated by AI capitalism is not reducible to a problem of rising production costs or supply-side constraints on accumulation. It represents, rather, a systematic assault on the regenerative capacities of the natural world itself. As Foster has argued, capitalism’s relationship with nature is defined by a structural antagonism: the logic of endless accumulation is irreconcilable with the finite regenerative limits of natural systems.33 Capital does not merely exploit nature as a condition of production; it metabolically ruptures the cycles and relationships through which nature reproduces itself. Paul Burkett deepens this analysis by recovering from Marx a conception of nature that refuses reduction to instrumental value.34 Natural systems possess use values that are irreducible to their role in the production process, and capitalism’s systematic destruction of these use values—its conversion of living ecosystems into inputs and waste sinks—constitutes an ecological crisis in the fullest sense: not a crisis of profitability, but a crisis of the biophysical conditions of life itself.
The AI economy represents an acute intensification of this dynamic. The data centers, cooling systems, and mineral supply chains that sustain the infrastructure of AI are not simply drawing down natural resources in the economist’s sense of raising input costs. They are participating in a cumulative and largely irreversible degradation of the water systems, energy ecologies, and extractive landscapes on which both human and nonhuman life depends. This degradation does not appear on the balance sheets of technology firms, not because it is economically marginal, but because capital’s accounting system is structurally incapable of registering the destruction of values that were never commodified. The ecological crisis of AI is therefore not a market failure awaiting a market correction; it is an expression of what capitalism does to nature when it operates without limit.
The dominant response to this contradiction within the framework of capitalist governance is the discourse of green AI and sustainable computing—the claim that the ecological crisis of AI can be resolved through technological innovation, market mechanisms, and voluntary corporate commitment. This response deserves serious analytical attention, not because it is convincing but because understanding its failure illuminates the structural character of the problem.
The green AI discourse rests on three claims: that renewable energy can supply AI’s electricity demands without net ecological harm; that hardware efficiency improvements will reduce the per-unit ecological cost of computation sufficiently to offset growth in total demand; and that AI itself will generate environmental benefits—through climate modeling, energy optimization, and materials science—that outweigh its ecological costs. Each of these claims is undermined by the structural dynamics of capital accumulation.
The renewable energy claim fails, as noted above, because AI’s electricity demand is growing faster than renewable capacity, because temporal mismatches between renewable supply and data center demand require fossil fuel backup generation, and because technology firms are actively contracting for gas-fired capacity to ensure reliability. The efficiency claim fails because of the Jevons Paradox: hardware efficiency improvements lower the cost of computation and thereby stimulate greater demand, producing higher total energy consumption rather than lower. The net benefit claim fails because it treats AI’s ecological costs and benefits as commensurable and tradeable, when in reality, the ecological costs are concentrated, local, and borne by vulnerable communities. Meanwhile, the benefits are diffuse, speculative, and appropriated by shareholders and consumers in wealthy countries. There is no market mechanism that can aggregate these asymmetrically distributed effects into a rational social accounting.35
The carbon offset and net-zero pledge mechanisms through which technology firms manage their public ecological accounting are subject to analogous critiques. Carbon offsets—payments to projects that claim to reduce emissions elsewhere, offsetting a firm’s own—are plagued by problems of additionality, permanence, and verification that render many of them ecologically fictitious.36 Net-zero pledges that depend substantially on offsets rather than absolute emissions reductions are, in thermodynamic terms, accounting maneuvers rather than physical interventions: they do not reduce the entropy generated by data center operations; they purchase claims on entropy reductions elsewhere, many of which do not materialize. As Clark and York have demonstrated in their analysis of carbon metabolism, the biospheric rift generated by fossil fuel capitalism is not an externality to be priced and managed but a structural feature of capital’s relationship with the carbon cycle—a relationship that the expansion of AI infrastructure is now deepening and accelerating.37
A more fundamental critique concerns the relationship between efficiency and scale. The history of industrial capitalism is a history of efficiency improvements that have consistently been overwhelmed by scale expansion, a history that Georgescu-Roegen analyzed as the inevitable consequence of applying thermodynamic insight to an economic system organized around unlimited growth.38 There is no efficiency improvement, however dramatic, that can make an exponentially expanding system sustainable on a finite planet with a fixed entropy budget. The question is not whether AI can be made more efficient—it can, and the improvements are real—but whether efficiency improvements can outpace the growth in demand driven by competitive accumulation. The evidence of the past decade suggests they cannot. The thermodynamic logic of capital accumulation provides the structural reason why.
This brings us to what we might call the thermodynamic limit of capital: the point at which the entropy generated by capital accumulation exceeds the capacity of the biosphere to absorb it without catastrophic disruption of the systems—climate, hydrology, biodiversity, and soil fertility—on which human civilization depends. This limit is not a precise threshold that can be identified in advance; it is a zone of deepening crisis, already entered in several dimensions (atmospheric carbon concentration, freshwater depletion, and biodiversity loss) and being approached in others. The expansion of AI under the current regime of capital accumulation is not moving civilization away from this limit but toward it, at an accelerating rate.
The political economy of this trajectory is clear. The costs of approaching the thermodynamic limit of capital are not borne by those who drive the accumulation, that is, the shareholders, executives, and institutional investors of the technology firms whose competitive dynamics determine the pace of AI expansion. They are borne by communities in water-stressed regions whose aquifers are depleted by data center cooling, by workers in artisanal mines whose health is destroyed by mineral extraction, by populations in climate-vulnerable countries whose food security is threatened by carbon emissions, and by future generations who will inherit a planet with a diminished capacity for ecological self-regulation. This is the political economy of entropy: the privatization of the benefits of low-entropy consumption and the socialization of the costs of high-entropy waste.39
No technical innovation can resolve this political economy, because it is not a technical problem. It is a problem of power—of who controls the means of computation, who determines the purposes for which computational capacity is deployed, and who bears the ecological costs of that deployment. Addressing it requires not better algorithms or more efficient chips but a fundamental transformation of the social relations of production in the digital economy. It requires, in short, a politics adequate to the thermodynamic stakes of the present moment.
Conclusion
The social and historical analysis developed in this article leads to a conclusion that the dominant discourse on AI and sustainability systematically evades: the ecological crisis of AI is not a problem of insufficient innovation or inadequate corporate responsibility, but a structural expression of capitalism’s irresolvable tension with the biophysical limits of the planet. The specific processes driving this crisis—the monopoly arms race, the Jevons dynamic, and the systematic displacement of ecological costs onto the Global South—are not technical malfunctions awaiting engineering solutions. They are the normal operations of capital accumulation in its monopoly-digital phase, registered in thermodynamic terms as dissipative processes of Prigoginian nonreversibility: permanent, compounding, and beyond the reach of market correction.
The ecosocialist tradition offers the most theoretically coherent starting point for an alternative. As Foster has argued, the metabolic rift between capital and nature cannot be repaired within the institutional framework of capitalism itself; it requires a fundamental reorganization of the social relations of production—one that subordinates the imperatives of accumulation to the regenerative limits of the natural world. An ecosocialist logic of computing would rest on three foundational commitments. First, it would rely on democratic control of computational infrastructure: data centers, AI platforms, and the networks connecting them constitute critical social infrastructure whose governance cannot be left to the competitive imperatives of private capital. Like electrical grids and water systems, they require democratic accountability: forms of social control that allow communities to determine the purposes for which computational capacity is used and the terms on which its ecological costs are distributed. Second, it would necessitate a reorientation of research and development priorities away from profit-maximizing applications—advertising optimization, financial trading, and labor surveillance—toward applications that genuinely serve social needs. This includes renewable energy management, public health, ecological monitoring, and education. Third, and most fundamentally, it would require acceptance that the scale of computational activity must be constrained by ecological limits. Sufficiency, by which is meant computing enough, rather than computing more, must become an organizing principle, replacing the growth imperative that drives the current AI arms race.
None of these transformations is imminent, and none can be achieved by technical means alone. The irreversibility that Prigogine identified in dissipative systems has its social analogue in the path dependencies of capitalist infrastructure: the data centers already built, the fossil fuel contracts already signed, and the extraction landscapes already degraded. What ecosocialist politics can accomplish is not the reversal of past damage but the interruption of the processes generating future damage—a break in the social logic of accumulation that thermodynamics registers but cannot by itself produce. The question before us is not whether the limits of capital will assert themselves, but whether they will be confronted on terms set by democratic societies committed to ecological survival, or on terms imposed by the cascading crises of a biosphere pushed beyond its regenerative capacity. The algorithm does not decide. Politics does.
Notes
↩ John Bellamy Foster, Marx’s Ecology: Materialism and Nature (New York: Monthly Review Press, 2000), 141–77.
↩ Sasha Luccioni, Alexandre Viguier, and Anne-Laure Ligozat, “Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model,” Journal of Machine Learning Research 24, no. 253 (2023): 1–15.
↩ Goldman Sachs Research, “AI Is Poised to Drive 160% Increase in Data Center Power Demand,” May 14, 2024.
↩ Microsoft, 2022 Environmental Sustainability Report (Redmond, Washington: Microsoft Corporation, 2022), 17.
↩ Jeremy Rifkin, The Zero Marginal Cost Society: The Internet of Things, the Collaborative Commons, and the Eclipse of Capitalism (New York: Palgrave Macmillan, 2014), 11–14.
↩ Nicholas Georgescu-Roegen, The Entropy Law and the Economic Process (Cambridge: Harvard University Press, 1971), 3–4; John Bellamy Foster, Brett Clark, and Richard York, The Ecological Rift: Capitalism’s War on the Earth (New York: Monthly Review Press, 2010), 54–76.
↩ Karl Marx, Capital: A Critique of Political Economy, vol. 1, trans. Ben Fowkes (London: Penguin, 1976), 493–94.
↩ Emma Strubell, Ananya Ganesh, and Andrew McCallum, “Energy and Policy Considerations for Deep Learning in NLP,” Proceedings of the Fifty-Seventh Annual Meeting of the Association for Computational Linguistics (2019): 3645–50.
↩ David Patterson et al., “Carbon Emissions and Large Neural Network Training,” arXiv preprint, arXiv:2104.10350 (2021), 1–9.
↩ Jordan Hoffmann et al., “Training Compute-Optimal Large Language Models,” arXiv preprint, arXiv:2203.15556 (2022), 1–19.
↩ John Bellamy Foster, “The Ecology of Marxian Political Economy,” Monthly Review 63, no. 4 (September 2011): 1–16.
↩ William Stanley Jevons, The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal Mines (London: Macmillan, 1865), 152–53.
↩ “AI and Compute,” OpenAI (blog), openai.com/blog/ai-and-compute.
↩ International Energy Agency, Electricity 2024: Analysis and Forecast to 2026 (Paris: IEA, 2024), 14.
↩ Ilya Prigogine and Isabelle Stengers, Order Out of Chaos: Man’s New Dialogue with Nature (New York: Bantam Books, 1984), 143–45.
↩ Foster, Clark, and York, The Ecological Rift, 73–76.
↩ Goldman Sachs Research, “AI Is Poised to Drive 160% Increase in Data Center Power Demand.”
↩ Brett Clark and John Bellamy Foster, “Ecological Imperialism and the Global Metabolic Rift: Unequal Exchange and the Guano/Nitrates Trade,” International Journal of Comparative Sociology 50, no. 3–4 (2009): 311–34.
↩ Arsheeya Bajwa, “Microsoft, OpenAI Plan $100 Billion Data-Center Project, Media Report Says,” Reuters, March 29, 2024.
↩ International Energy Agency, Electricity 2024, 27.
↩ EirGrid, Tomorrow’s Energy Scenarios 2023 (Dublin: EirGrid, 2023), 34.
↩ Benjamin K. Sovacool et al., “Sustainable Minerals and Metals for a Low-Carbon Future,” Science 367, no. 6473 (2020): 30–33.
↩ International Energy Agency, Electricity 2024, 27–29.
↩ Kim Crawford and Vladan Joler, “Anatomy of an AI System,” 2018, anatomyof.ai.
↩ Pengfei Li et al., “Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models,” arXiv preprint, arXiv:2304.03271 (2023), 1–10.
↩ Microsoft, 2022 Environmental Sustainability Report, 17.
↩ Google, 2023 Environmental Report (Mountain View, California: Google, 2023), 22.
↩ Peter Dauvergne, AI in the Wild: Sustainability in the Age of Artificial Intelligence (Cambridge: MIT Press, 2020), 78–79.
↩ Rob Nixon, Slow Violence and the Environmentalism of the Poor (Cambridge, Massachusetts: Harvard University Press, 2011), 2–3.
↩ Guillaume Pitron, The Rare Metals War: The Dark Side of the Clean Energy and Digital Technologies, trans. Bianca Jacobsohn (London: Scribe, 2023), 45–67.
↩ United Nations Institute for Training and Research, The Global E-waste Monitor 2024 (Bonn: UNITAR, 2024), 3.
↩ John Bellamy Foster and Hannah Holleman, “The Theory of Unequal Ecological Exchange: A Marx-Odum Dialectic,” Journal of Peasant Studies 41, no. 2 (2014): 199–233.
↩ John Bellamy Foster and Brett Clark, “Introduction to the Updated Edition of Arghiri Emmanuel’s Unequal Exchange,” Monthly Review 77, no. 8 (January 2026): 1–19.
↩ John Bellamy Foster, “Capitalism and Ecology: The Nature of the Contradiction” Monthly Review 54, no. 4 (September 2002): 6–16.
↩ Paul Burkett, “Fusing Red and Green” Monthly Review 50, no. 9 (February 1999): 47–56; Paul Burkett, Marx and Nature: A Red and Green Perspective (New York: St. Martin’s Press: 1999).
↩ John Bellamy Foster and Brett Clark, “The Robbery of Nature: Capitalism and the Metabolic Rift,” Monthly Review 70, no. 3 (July–August 2018): 1–20.
↩ Benjamin K. Sovacool et al., “Sustainable Minerals and Metals for a Low-Carbon Future,” Science 367, no. 6473 (2020): 30–33.
↩ Brett Clark and Richard York, “Carbon Metabolism: Global Capitalism, Climate Change, and the Biospheric Rift,” Theory and Society 34, no. 4 (2005): 391–428.
↩ Georgescu-Roegen, The Entropy Law and the Economic Process, 276–78.
Discover more from Class Struggle Ecology
Subscribe to get the latest posts sent to your email.
