Expanding “Green AI” Beyond Model Efficiency

Abstract

This commentary argues that the artificial intelligence (AI) boom is not immaterial but relies on energy- and resource-intensive infrastructures. While “Green AI” scholarship has advanced model-level efficiency metrics and reporting, it has largely overlooked the material circuits that enable these models. We explain how efficiency-first framings are reinforced by an ideological blend of cybertarianism and techno-nationalism, which together reframe environmental and labor externalities as acceptable costs of competition. We propose three priorities for a more adequate agenda: (1) adopt the “trash metaphor” to highlight the afterlives and externalities of AI infrastructures; (2) implement structural solutions that oversee and regulate the entire material circuits of AI; and (3) foster green citizenship through public information rights, participatory siting, and ongoing civic oversight to advocate for and sustain the necessary structural changes. Overall, we call for aligning AI's claimed climate benefits with demonstrable compliance to ecological budgets rather than aspirational efficiency narratives.

Introduction

Although the ongoing artificial intelligence (AI) boom is frequently described as a dematerialized revolution propelled by algorithms, big data, cloud computing, and large language models, in reality, its development is inextricably underpinned by the rapidly growing demand for chips, electricity, labor, land, water, and other material resources. Take electricity as an example. According to the International Energy Agency's (IEA) projection (2025), data centers’ worldwide electricity consumption will more than double to about 945 TWh by 2030, with AI model training and deployment being the main cause of this increase.

Another attention-grabbing case making the materiality of AI visible is the intensifying digital geopolitical competition between China and the United States. Since 2022, the U.S. government has imposed restrictions on the export of advanced graphics cards to China, further tightening these measures in 2025 (Kharpal, 2022; “US revokes TSMC,” 2025). As a countermeasure, the Chinese government has recently intensified regulations on rare earth elements and magnet-related technologies essential for motors, turbines, and data center equipment (Kim et al., 2025; Pallardy, 2025). When discussing the AI revolution, both sides of the ongoing techno-national rivalry frequently employ terms like “big data,” “innovation,” and “talent” to highlight intangible prowess, yet their policies target infrastructures and supply chains.

The conflicts around AI chips and rare-earth magnets reveal a straightforward reality: The current AI revolution is based on energy-intensive and extractive systems. However, the prevailing academic discourse on “green AI” predominantly focuses on model-level assessments (e.g., monitoring computational usage, reducing training emissions via hyperparameter optimization), with comparatively insufficient emphasis on the material circuits of data centers, networks, and hardware lifecycles. A recent systematic literature review corroborates this bias: The majority of green AI publications since 2020 have focused on monitoring footprints and optimizing models instead of examining upstream and downstream infrastructures (Verdecchia et al., 2023).

This commentary advocates for broadening the green AI discourse to encompass infrastructural materiality in addition to model efficiency. Following a brief review of current green AI research, we discuss the ideological blend of cybertarianism and techno-nationalism that helps to obscure the material costs of AI in both U.S. and Chinese discourses. The commentary concludes with three research and policy priorities: (1) Employ the “trash metaphor” (Pasek, 2025) to emphasize the afterlives and externalities of AI infrastructures, (2) implement structural solutions that oversee and regulate the entire material circuits of AI, and (3) foster green citizenship to advocate for and maintain the necessary structural changes. The urgency of this expanded agenda is sharpened by recent claims, notably Elon Musk's argument that future AI competition may be decided by “watts” as much as by models or chips (Lee, 2026), which rightly foreground electricity as a bottleneck but risk recasting green AI as a race for more power rather than a debate about ecological limits (Lal & You, 2025).

The state of green AI

The academic discussion of green AI is embedded in a longer period of research tradition in green communication, which emphasizes materiality, environmental and labor costs, and social governance with active public engagement. It also positions communication technologies and behaviors in the nexus of broader economic, societal, and political systems. However, as the digital infrastructure expands globally, less scholarly attention has been paid to both material and infrastructural levels due to their growing invisibility. In recent years, scholarly discussions on green AI have focused on techniques for measuring and lowering the carbon footprint of models, ranging from carbon-conscious scheduling, pruning, and distillation procedures to standardized emissions calculators and reporting. There have been calls to make energy efficiency a primary evaluation criterion, to mandate energy and carbon footprint reporting in machine learning papers, and to develop tools for carbon emissions estimation (Henderson et al., 2020; Lacoste et al., 2019; Schwartz et al., 2020). Echoing such calls, research has been undertaken to quantify the carbon footprints of large language models (e.g., Luccioni et al., 2023; Strubell et al., 2019) and to establish best practices for reducing energy consumption during machine learning training (Patterson et al., 2022). This growing body of research is essential for elucidating AI's environmental effects and promoting the assessment of AI models beyond mere accuracy and complexity.

However, even sympathetic studies now acknowledge a blind spot: The material circuits that enable the functioning of AI models. As noted in Verdecchia et al. (2023), the field of green AI is dominated by work on footprint monitoring and hyperparameter tuning, with far fewer studies attending to issues such as supply chains, land use, water consumption, and e-waste. For example, Li et al. (2025) argue that AI's water footprint, particularly during training and serving, leads to significant and frequently underreported water withdrawals and consumption, necessitating location- and time-aware accounting. Similarly, Kaack et al. (2022) distinguish computing-related, application-level, and system-level effects, warning that efficiency gains at the model level can be swamped by rebound and infrastructure growth without policy constraints. Recent energy-systems scholarship (e.g., Lal & You, 2025) echoes this point by noting that the energy and climate consequences of AI infrastructure deployment itself remain underexplored relative to both traditional data centers and model-level green AI approaches.

Taken together, the current discourse on green AI lacks a shift from evaluating per-model efficiency to emphasizing per-infrastructure accountability. Model-centric evaluations seldom incorporate the specifics of electricity and water consumption, the location and cooling of data center campuses, the extraction of minerals and components upstream, or what happens to servers and accelerators downstream. The significance of per-infrastructure accountability is pronounced at scale: regardless of the efficacy of any singular model, there will be a substantial increase in AI data centers and their corresponding carbon emissions by 2030–2035 (IEA, 2025; Zewe, 2025).

Mirroring the academic narrowness, public discourse often equates “green AI” with more energy-efficient chips and models. For example, media coverage of Google's sixth-generation Trillium chip (e.g., Cherney, 2024) claims performance enhancements, highlighting a 67% increase in energy efficiency compared to its predecessor, while the wider implications for data centers and grids are only marginally addressed. Likewise, prevalent explainers and institutional blogs (e.g., Kandemir, 2025; Zewe, 2025) tend to depict sustainability as optimizing models or improving hardware design, rather than addressing governance issues related to data centers’ siting, water consumption, and grid constraints. In short, the efficiency-centric narratives are not wrong, but they are incomplete.

Cybertarianism meets techno-nationalism

Efficiency-centric narratives about “green AI” rest on a broader ideological blend of cybertarianism and techno-nationalism, which reinforce the public's utopian perceptions of digital technologies. Cybertarianism, heir to the “Californian ideology” (Barbrook & Cameron, 1996), portrays networked media as inherently democratizing: Heroic founders disrupt the establishment; “interactive” users become sovereign co-creators; and privately owned, lightly regulated infrastructures are assumed to be progressive. Barbrook and Cameron's classic critique traces this ideology to 1990s techno-utopianism fused with market libertarianism, later expanded by environmental media scholars (e.g., Maxwell & Miller, 2012), who show how a technological sublime renders digital systems clean and immaterial, obscuring labor and ecological costs. Another line of critique against cybertarianism comes from political economy studies on “free labour” and prosumption, which demonstrate how participation is routinely appropriated as unpaid value creation under platform capitalism (Fuchs, 2014; Terranova, 2000).

Techno-nationalism, by comparison, positions digital technologies as strategic assets in international competition, thereby justifying state control over capital, standards, energy, and supply chains to gain a competitive edge (Na & Pun, 2023). Recent research demonstrates how low-carbon and high-tech ambitions are expressed through techno-nationalist narratives (Chen, 2024). Think tanks in Washington, meanwhile, have been pushing for the inclusion of computing, electricity, and data center site selection as national tech policy priorities, calling for non-intermittent power supply, tax credits, and permitting reforms to hasten the development of domestic AI clusters (e.g., Datta & Fist, 2025).

In the current AI upcycle, the most visible intersection of cybertarianism and techno-nationalism is the framing of “heroic founders” and their ventures as national champions. Both U.S. and Chinese coverage personify progress by glorifying entrepreneurs and “frontier labs,” treating national prestige as the ultimate stake. A recent example is China's celebration of models like DeepSeek as evidence of resilience against the U.S. silicon blockade (Chen et al., 2025) and a driving force to the diversification of global AI. By shifting attention from infrastructures and institutions to personalities and performance benchmarks, this personalization reinforces the cybertarian focus on exceptional innovators, while the techno-national frame converts their corporate success into a proxy for state strength.

Moreover, further revealing the ideological blend is the pairing of light-touch market rhetoric with heavy-handed statecraft. Neoliberalism promises markets to be self-correcting, yet in the case of the ongoing AI arms race, both China and the United States intervene materially. As noted in the Introduction, the United States has progressively tightened controls on advanced AI chips while exploring policies—from tax credits to financing electricity infrastructure—to secure “domestic compute” (Kharpal, 2022; “US revokes TSMC,” 2025). Meanwhile, China has pushed back by combining industrial policy with expanded controls on rare earths and magnet-related technologies essential to motors, turbines, and data center equipment (Kim et al., 2025; Pallardy, 2025). Along with this interstate rivalry, cybertarianism naturalizes private ownership and frictionless scaling of core infrastructures, while techno-nationalism normalizes reliance on state sponsorship and regulation. Together, they frame environmental and labor externalities as tolerable costs of competition rather than governable public problems—a dynamic long flagged by critics of cyber-utopianism, who warned that internet-centric narratives deflect attention from institutions, power, and material systems (Morozov, 2011).

Musk's recent intervention makes this infrastructural logic unusually explicit. In a January 2026 podcast interview (later reported in the Business Press), he argued that China's likely advantage in AI lies in its ability to bring more electricity online and that power availability, more than chip restrictions alone, will determine scaling capacity (Lee, 2026). While this watt-centric framing usefully punctures the myth of AI immateriality, it is insufficient for green AI because it treats the capacity to mobilize more electricity as the decisive metric of success, sidelining the carbon intensity, water burden, transmission bottlenecks, labor conditions, and waste streams through which those watts are produced and consumed.

Finally, as generative AI becomes integrated into everyday interactions, public narratives praise creativity and co-production while obscuring the invisible work of annotation, moderation, and the extractive dynamics of data capture (Wu et al., 2025). This simultaneous celebration of consumer interactivity and neglect of digital labor results in a cybertarian gloss—participation as empowerment—overlaid on highly capital-intensive, tightly centralized, and programmable infrastructures that concentrate control and rents. This contradiction has been highlighted in critical accounts of prosumption (e.g., Fuchs, 2014), which emphasize how seemingly participatory systems routinely appropriate user activity as unpaid value creation.

Where to go from here

Making AI's material circuits explicit is the first step toward a more accurate assessment of its environmental impact, as we have argued in this commentary. Rapidly expanding data center campuses, whose daily operations depend on optical networks, power grids, water resources, and components with geographically concentrated supply chains, are essential for training and implementing AI models. Equally vital but often overlooked are the precarious workers who annotate data, moderate content, repair hardware, and disassemble e-waste, since they are dispersed globally. These material details are becoming harder to ignore. For example, in the United States alone, electricity consumption at data centers is expected to more than double between 2024 and 2030 (Leppert, 2025). Simply put, if scholarly discussions on green AI remain confined to per-model efficiency, they will keep missing important questions such as who pays for what, when, how much energy and water are used, how network capacity is provided, which materials are extracted upstream, and what happens to servers and chips when they reach end-of-life.

China's rapid build-out of solar generation, battery storage, and domestic semiconductor capacity illustrates why more electricity and more hardware do not automatically resolve green AI concerns. On the one hand, these developments may relieve some immediate bottlenecks: China generated over 10,000 TWh of electricity in 2024, continues to expand wind and solar capacity, and is experiencing a battery-storage boom driven partly by data center growth and renewable integration (Mak, 2026). Its domestic AI chip ecosystem is also scaling quickly, with Chinese vendors shipping 41% of the country's AI accelerator cards in 2025 and broader AI demand stimulating semiconductor testing, packaging, and optical interconnect industries (Baptista, 2026; Pan & Chen, 2026). On the other hand, these same capacities can accelerate infrastructural expansion. Reuters reporting on China's “power edge” notes persistent transmission bottlenecks, renewable curtailment, underutilized western data centers, and the risk of overcapacity across the AI stack (Mak, 2026). Likewise, China's own green-transition planning acknowledges that AI data centers could consume more than 1000 TWh of electricity annually by 2030 and complicate national climate goals unless their growth is governed more tightly (You, 2026). From a green AI perspective, then, Musk is only partly right: Watts matter, but the decisive question is how those watts are generated, transmitted, allocated, and governed within ecological budgets.

Where shall green AI research go from here? In our view, there are three worthwhile research and policy priorities. First, research should treat AI as a prospective waste regime. Pasek's (2025) provocative argument that “AI is trash” offers an insightful design heuristic: It urges scholars and practitioners to consider AI's afterlives as seriously as lifetimes. Contaminated cooling water, heat and greenhouse gas emissions, obsolete chips and circuit boards, degraded batteries, and exported residues—these waste materials prompt concrete questions that model-centric accounts rarely address: Can hardware be designed for refurbishment and secondary markets rather than accelerated obsolescence? How are embodied emissions and hazardous outputs disclosed by data centers? Which communities and jurisdictions bear the material residues of AI expansion? How is the AI waste globally distributed?

Second, model-level optimization must be accompanied by structural, whole-circuit governance. Mandating campus- and workload-level disclosures (e.g., site-specific water withdrawals and consumption; periodic reporting of electricity consumption from both renewable and non-renewable sources) promotes transparency and may incentivize sustainable practices when paired with appropriate policies. These policies include, but are not limited to, tying data center approvals to clean energy supply, funding recycling for magnets and other high-value components, requiring extended producer responsibility for servers, standardizing refurbishment markets to extend hardware life, codifying labor standards for data annotation, and so on. Without this infrastructural shift, efficiency enhancements risk functioning merely as a cybertarian legitimation of the ongoing expansion of energy- and material-intensive systems.

Third, the research agenda of green AI needs green citizenship to ensure structural changes are durable. Maxwell and Miller's (2012) “greening the media” initiative is both technical and civic: It calls for the public to oversee and govern the material infrastructures of the media ecosystem. In the context of the ongoing AI revolution, this means institutionalizing public participation in the siting and permitting of data centers; conducting public consultations on site-specific energy and water demands; supporting public investment in transmission, storage, and clean-firm capacity to reduce the environmental impact of AI growth; and insisting on fair labor practices and responsible e-waste management.

More specifically, green citizenship works through at least four channels. First, it turns environmental information into democratic capacity: communities, workers, and local institutions need actionable rights to know how much electricity, water, land, and hazardous waste particular AI facilities require. Second, it expands participation from one-off consultation to ongoing oversight, so residents, labor representatives, and civil-society groups can contest permits, demand cumulative-impact assessments, and monitor compliance over time. Third, it links individual responsibility to collective institutions, echoing green-citizenship scholarship's insistence that environmental politics cannot be reduced to consumer choice alone (Gabrielson, 2008). In AI governance, this means moving beyond appeals to “use AI responsibly” toward organized claims on utilities, regulators, procurement offices, universities, and cloud providers. Fourth, it intersects with what energy-transition scholars call energy citizenship and energy democracy, where public participation and representational forms of governance help shape infrastructures rather than merely ratify them after the fact (Silvast & Valkenburg, 2023; Wahlund & Palm, 2022). Green citizenship, in this stronger sense, can improve green AI initiatives by forcing disclosure of site-specific energy and water burdens, strengthening labor and e-waste accountability across supply chains, and giving affected communities standing to slow, redirect, or condition infrastructure growth. Green citizenship, in short, provides the foundation for legitimate, democratically accountable growth of computational capacity within ecological constraints.

In conclusion, none of the proposals above deny AI's potential contributions to decarbonization and sustainability. Instead, they emphasize that AI's claimed benefits should be demonstrated within ecological limits rather than assumed. With more research focused on e-waste, governance structure, and citizenship, model efficiency could be better linked to infrastructural accountability, ensuring that green AI becomes more than just a hopeful label for future chips.