Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA

The accelerating deployment of artificial intelligence (AI) servers is being fuelled by the growing demand for generative AI applications, ignited by milestones such as the release of ChatGPT in 2022¹. Projections signal even greater impact, exemplified by the recent Blackwell platform, which analysts herald as a new Moore’s Law era². Pronouncements from influential AI industry figures³ on both demand and supply sides are seen as transformative shifts for the entire data-centre industry. Whereas AI has been employed in various fields to advance sustainability^4,5, the remarkable energy requirements of AI itself raise concerns regarding not only energy provisioning challenges⁶ but also water scarcity and climate change issues stemming from the energy–water–climate nexus of AI data centres^1,7,8. However, the holistic energy–water–climate implications of AI computing are largely unknown, constrained by untransparent industry reports and limited data.

The climate impact of AI servers will stem primarily from their operations (Scopes 1 and 2)⁹ and supply-chain activities (Scope 3), including manufacturing and end-of-life treatment^10,11. The Scope 2 emissions from indirect energy purchases are expected to constitute a substantial portion and indicate a high dependency on the increase of AI server energy consumption. According to the International Energy Agency¹², 0.6% of global total carbon emissions comes from the data centres and data transmission networks due to their electricity consumption. The industry energy consumption could double by 2026, motivated by AI and other sectors¹³, threatening decarbonization targets under the Paris Agreement^14,15, which include a 53% reduction in data-centre emissions by 2030 and net-zero goals for the AI sector. Further, growing AI server energy consumption implies an increasing water footprint through Scope 1 (direct cooling) and Scope 2 (indirect procurement) water use^8,16. Centralized installation in water-stressed regions may perturb local water balance and threaten supply for millions^17,18.

Previous research has developed bottom-up and top-down methods to assess energy–water–carbon outcomes of servers^19,20,21, but these approaches face challenges with the rise of AI servers. Top-down approaches based on activity indices, such as data traffic and computing instances, fail to accurately represent AI-driven workloads^22,23. Detailed bottom-up approaches suffer from limited data availability and lack of industry insight⁷. Assumptions valid for traditional collocation centres often fail for AI data centres, which differ in installation and operation²⁴. Recent studies have explored computing task-based analyses to better quantify AI-related energy and resource consumption, providing insights but lacking systematic policy guidance^25,26. Two notable contributions are the 2024 data-centre report by Lawrence Berkeley National Laboratory²⁷ and the 2025 Energy and AI Report by the International Energy Agency²⁸, which present scenarios estimating US and global data-centre energy use under highly uncertain AI growth. Although use of confidential commercial data may limit reproducibility, they offer important benchmarks. Our study extends this foundation by (1) developing an open-source, bottleneck-based modelling approach with comprehensive uncertainty analysis; (2) systematically assessing energy, water and carbon impacts, incorporating dynamic interactions with local energy systems; and (3) proposing actionable mitigation strategies for potential net-zero trajectories across different AI server deployment scenarios.

Here we analyse the combined energy–water–climate impact of operational AI servers in the United States between 2024 and 2030, balancing importance and future uncertainties and addressing a series of fundamental questions. (1) What are the magnitude and spatiotemporal distributions of energy consumption, water footprint and climate impact from AI server deployment? We address this using temporal projection models and regional frameworks, assuming deployment mirrors current large-scale AI data-centre patterns. (2) What are the prospects for near-term net-zero pathways? We evaluate this by analysing best- and worst-case scenarios of key drivers, including industry efficiency improvements, server location distribution and grid decarbonization. The results of these analyses are presented in the following sections.

This section provides a base depiction of the AI servers’ energy, water, and carbon impacts in the United States, emphasizing the spatiotemporal characteristics of the system. Figure 1a shows the projected cumulative AI server installations in the United States from 2024 to 2030 under five scenarios: low demand, low power, mid-case, high application and high demand. The mid-case serves as the base scenario, while the low and high demand set projection bounds. Low power assumes server efficiency gains, and high application accounts for increased adoption driven by efficiency. Figure 1b illustrates the state-level allocation of AI servers, showing power usage effectiveness (PUE), projected grid carbon intensity (carbon emissions per unit of electricity), water usage effectiveness (WUE) and projected grid water intensity (water footprint per unit of electricity) for each state. Southern states such as Florida exhibit higher PUE and WUE than northern states such as Washington, reflecting climate impacts. Moreover, the grid factors demonstrate notable sensitivity to location, implying the importance of the local grid for the AI servers’ environmental impacts. Figure 1c–g shows energy, water and carbon results. Figure 1c illustrates AI server energy predominates over infrastructure energy. Figure 1d indicates indirect water footprint contributes 71% of total, with direct use at 29%. Annual estimates of energy consumption, water footprint and carbon emissions of AI servers from 2024 to 2030 under each scenario are presented in Fig. 1e–g. Even the lowest scenarios outline a considerable increase in the energy, water and carbon footprints of AI servers. The highest scenario yields the highest environmental impact, largely surpassing previous forecasts for the entire US data-centre market^18,29, underscoring the environmental risks of unchecked AI server expansion.

a, The projected accumulative capacity of AI servers in the United States from 2024 to 2030 under different scenarios. b, Spatial allocation data for each state, accompanied by corresponding metrics, which include PUE, WUE, grid carbon factor (kgCO₂-equivalent kWh⁻¹), and grid water factor (L kWh⁻¹) data. The metric values are calculated as the average value from 2024 to 2030. c, The ratios between AI servers and infrastructure energy consumption. d, The ratios of the indirect and direct water footprint of AI servers. e–g, The energy consumption (e), water footprint (f) and carbon emissions (g) of AI servers from 2024 to 2030 under different scenarios. The red dashed lines in e–g denote the forecast footprint of the US data centres, based on previous literature^18,29.

Source data

Full size image

The United States is selected as the research region because of its dominant position in the global AI market. The research period is selected as 2024–2030, generated from a trade-off between importance and uncertainty. Projection of the AI server accumulative capacity in the United States is the initial step, which was generated on the basis of the forecast of AI chip manufacture capacity, AI server specifications and AI server adoption pattern. In addition, we assume the AI data centre will align with current AI company large-scale data-centre allocation ratios, presented in detail in Supplementary Fig. 5. The PUE and WUE values are derived by a hybrid statistical and thermodynamics-based model³⁰. Supplementary Table 3 lists all model inputs, and the applied values are calculated as the average of the best and worst practices. In addition, projected grid factors are calculated on the basis of the Regional Energy Deployment System (ReEDS) model³¹ by involving the projected data-centre load data and adopting regulations such as the Inflation Reduction Act. The step-by-step calculation process of the projection is summarized in Methods and sections 1–4 of Supplementary Information. Important uncertainties, such as PUE and WUE estimation, technology advancements, the spatial distribution of AI server allocation and grid development patterns, and sensitivity analysis of key parameters will be discussed in the following sections.

Over the past decade, efficiency gains in the data-centre industry have stabilized environmental costs despite a doubling of computing instances²⁰. This section examines the existing potential for further improvements through system optimization and technology adoption. Figure 2a,b illustrates the achievable PUE and WUE values of AI data centres in each state. The best-practice scenario suggests notable reduction potential: over 7% PUE reduction and over 85% WUE reduction, despite the high efficiency of AI data centres compared with the industry averages (PUE 1.58, WUE 1.8). The worst-practice scenario underscores the risk of neglecting efficiency efforts. The effects of achievable PUE and WUE values are further depicted in Fig. 2c,d, showcasing the corresponding influences on the energy, water and carbon footprints of AI servers. PUE reduction yields over 7% reductions in total energy consumption and carbon emissions. WUE reduction efforts result in over a 29% reduction in the total water footprint. Moreover, it is evident that PUE and WUE efforts can align with each other, as observed in WUE improvement results. Figure 2e delves into the potential impact of adopting advanced technologies within AI data centres, focusing on advanced liquid cooling (ALC) and server utilization optimization (SUO) adoption. The results show that the best ALC adoption can reduce about 1.7% of energy consumption, 2.4% of water footprint and 1.6% of carbon emissions of AI servers by 2030. For SUO, the best-case scenario, representing total adoption by 2030, results in a 5.5% reduction in all footprint values, while the worst-case scenario, representing frozen adoption, leads to a 7.3% increase by 2030. The energy–water–carbon impacts of AI servers from 2024 to 2030 following the mid-case scenario through the worst, base and best industry practices are further presented in Fig. 2f–h. The maximum reductions of energy, water and carbon drawn from the existing potential are about 12%, 32% and 11%, respectively. These findings underscore the considerable impact of industry efforts on the environmental cost of AI servers.

a, Comparison of the best, base and worst practices regarding AI data-centre PUE. b, Comparison of the best, base and worst practices regarding AI data-centre WUE. c, Analysis of the impact of PUE practices on energy consumption, water footprint and carbon emissions. d, Analysis of the impact of WUE practices on energy consumption, water footprint and carbon emissions. e, Assessment of the effect of ALC and SUO adoption on energy consumption, water footprint and carbon emissions. f–h, The energy consumption (f), water footprint (g) and carbon emissions (h) of AI servers from 2024 to 2030 following the mid-case scenario through the worst, base and best practices of all considered industry efforts.

Source data

Full size image

Base values of PUE and WUE are calculated by averaging the best and worst practices due to the lack of cooling specifications. The worst and best values are derived by solving the corresponding optimization problem with constrained operational parameters. Specifically, the best PUE is achieved mainly by extending the free cooling period through input air set-points adjustment and enhancing facility energy efficiency. WUE is improved by reducing windage and concentration water loss, adopting air-side economizers and enabling more free cooling. The model is based on previous works³⁰, with relevant parameters detailed in section 3 of Supplementary Information. The best, base and worst practices for ALC and SUO adoption are developed from existing studies and market reports, with assumptions and calculations outlined in Methods. ALC adoption is evaluated through the increased immersion cooling in AI data centres, while SUO improvements focus on raising the active server ratio. Importantly, the advancement in AI hardware could reduce energy, water and carbon footprints by improving energy efficiency, as seen in Nvidia’s future chip structures Blackwell and Rubin. However, these gains may be offset by the rebound effect. The uncertainties of the hardware evolution and market dynamics are discussed in the definition of the scenarios, as detailed in section 1 of Supplementary Information.

The location of data centres critically shapes their environmental impact^18,32. This section presents a detailed projection of how AI server spatial distribution may influence the environmental consequences of rapid US expansion. Figure 3a,b presents the top 25%, 50% and 75% locations with the lowest projected water footprint and carbon emissions per unit of server energy. Figure 3c presents the locations with combined lowest water and carbon factors. These allocation strategies are conducted on a grid balancing-area level. Specifically, the ReEDS model is deployed to calculate the grid factors of each balancing area under different projection scenarios. Figure 3d shows that Texas plays a vital role by possessing the most balancing areas with the top 25%, 50% and 75% water and carbon factors. The other western and southern states, including Montana, Louisiana, Idaho and New Mexico, also take a large portion of areas under the combined strategies due to widely leveraged local renewables. West Coast states, such as California, Oregon and Washington, as well as New England states, are suitable for carbon reduction but also lead to higher water footprints. The primary driver behind is the adoption of hydropower, which consumes large volumes of water through evaporation, as detailed in Supplementary Fig. 6. With more hydropower applied in the grid, the Scope 2 water usage of AI servers could dramatically increase when installed in such locations. Figure 3e–g depicts the total energy consumption, water footprint and carbon emissions changes of AI servers under the 25%, 50% and 75% spatial distribution scenarios. The base scenario is applied to all simulations to identify solely the influences of spatial distribution. Installing AI servers in the top water-oriented locations results in large reductions of both water and carbon footprints. Conversely, the carbon-oriented strategies lead to higher water footprints due to the hydropower application on the West Coast and New England states. Moreover, Fig. 3f shows that the water and carbon footprints can be concurrently reduced following a combined allocations strategy.

a, The top 25%, 50% and 75% of locations with the lowest projected water footprint per unit energy (2024–2030). b, The top 25%, 50% and 75% of locations with the lowest projected carbon emissions per unit energy. c, Locations ranked in the top 25%, 50% and 75% for both lowest water and carbon per unit energy. d, State area ratios under water- and carbon-oriented allocation strategies. e, Changes in total energy–water–carbon footprints for the 25%, 50% and 75% water-oriented scenarios. f, Changes in total water–carbon footprints for the 25%, 50% and 75% carbon-oriented scenarios. g, Changes in total energy–water–carbon footprints for the 25%, 50% and 75% combined water- and carbon-oriented scenarios. h, Current water footprint and water scarcity per unit server energy by state. i, Current carbon emissions per unit server energy and renewable energy potential by state.

Source data

Full size image

To project the potential risks and benefits behind future AI server deployment, Fig. 3h,i illustrates the current water scarcity and renewable energy potential of each state. The top ten states grappling with severe water scarcity issues are California, Nevada, Arizona, Utah, Washington, New Mexico, Colorado, Wyoming, Oregon and Montana, which are concentrated primarily in the western United States. Large adoption of hydropower leads to increased unit energy–water footprint in several states such as Arizona, California, Nevada, New Mexico and Utah, exacerbating the water scarcity issue. In addition, Texas, New Mexico, Kansas, Arizona, California, Colorado, Nevada, Nebraska, Oklahoma and South Dakota are identified as the top ten states with abundant renewable energy potentials. Combined with the water- and carbon- oriented strategies, Texas, Montana, Nebraska and South Dakota, situated in the Midwestern United States, emerge as optimal candidates for AI server installation, considering both water scarcity concerns and future decarbonization efforts.

This section analyses how future grid development will affect AI server environmental impacts. Figure 4a,b illustrates the modifications of grid carbon and water factors of each state under low renewable energy cost (LRC, best-practice) and high renewable energy cost (HRC, worst-practice) scenarios, compared with the base scenario. These scenarios, extracted from the predefined cases of the ReEDS model^31,33, represent the highest and lowest levels of renewable energy penetration, respectively. The resulting total carbon emissions and water footprint of AI servers are depicted in Fig. 4c,d. The HRC scenario indicates a 20% increase in carbon emissions alongside a 2.0% increase in water footprint, while the LRC scenario suggests an over 15% reduction in carbon emissions accompanied by a 2.5% of water footprint reduction. The carbon emissions of AI servers are shown to be heavily influenced by the grid decarbonization pattern, indicating both considerable reduction potential and associated risks.

a, Modifications of the grid water factor under the best and worst scenarios compared with the base scenario. b, Modifications of the grid carbon factor under the best and worst scenarios compared with the base scenario. c,d, The water footprint (c) and carbon emissions (d) of AI servers from 2024 to 2030 following the mid-case scenario under best, base and worst scenarios, respectively. e, The changes of AI server carbon and water footprints of each state under the best (LRC) and worst (HRC) scenarios.

Source data

Full size image

Figure 4e details changes in AI server carbon emissions and water footprints by state under LRC and HRC scenarios. The development pattern, including renewable energy penetration levels and sources, notably impacts the resulting footprints of each state. States such as Georgia, Nevada, North Carolina and Tennessee show marked sensitivity to renewable cost scenarios. In addition, the Pacific states, including California, Oregon and Washington, which achieved a low grid carbon factor, slow their hydropower adoption pace under the LRC scenario, avoiding exacerbating their water scarcity issues with additional AI server installations. The wind and solar resources, which have no water consumption while generating, are further adopted in the LRC scenario to reduce carbon emissions, meanwhile alleviating the severe water scarcity challenges in these states. The presented results suggest that the impacts of grid decarbonization patterns on the environmental costs of AI servers are evident not only over time but also in the spatial variations observed among different US states. These results connect to the importance of ambitious green electricity policies in US states, which potentially could further reduce carbon emissions in our scenarios.

Building on the preceding analysis of environmental costs, efficiency, spatial distribution and grid decarbonization, this section evaluates water and carbon net-zero pathways for US AI servers. Figure 5a presents the pathways from 2024 to 2030 to achieve net-zero carbon emissions and a net-zero water footprint of AI servers under the mid-case scenario. Figure 5b presents residual emissions and water footprints across different scenarios and practices. The 2030 target aligns with major AI data-centre operator goals^14,34,35. The top and bottom 25% of locations are used to create the best and worst cases for spatial distribution. The best and worst grid development practices are modelled under LRC and HRC scenarios, respectively. These practices form the upper and lower bounds, although better solutions can be obtained through extra policies and actions beyond current considerations.

a, The contributions of each influential factor to the water footprint and carbon emissions of AI servers with best and worst practices under the mid-case scenario. The total contributions of each factor from 2024 to 2030 are also listed. The curves above the 0 level indicate the increase of carbon emissions and water footprint, while the curves below represent the reductions. The grey area represents the residual footprints that need to be reduced. b, Presentation of the capacity of residual carbon emissions and water footprints to attain net-zero targets under different temporal scenarios for each year spanning from 2024 to 2030, under both best- and worst-practice scenarios. Specifically, the top and bottom 25% of locations are used to create the best and worst cases for spatial distribution. The best grid development practice is modelled under the LRC scenario, and the worst is modelled under the HRC scenario.

Source data

Full size image

The best and worst distribution patterns of AI server deployment lead to a 49% reduction and a 90% increase in carbon emissions, and a 52% reduction and a 354% increase in water footprints, respectively. Optimistic grid decarbonization provides a further 13% reduction in carbon, while the worst case results in a 23% increase. Industry efficiency efforts are critical for water sustainability, offering over 32% reduction in water footprint under best practices. Notably, combined best practices cut residual emissions and water footprints by 73% and 86%, respectively, indicating a feasible pathway to net zero. Under the mid-case, best industry, spatial and grid scenarios each reduces over 21 Mt, 25 Mt and 92 Mt of carbon from a base of 186 Mt due to AI server installation. The best practices produce about 11 Mt residual carbon emissions by 2030, requiring 28 GW of wind or 43 GW of solar to fully offset³⁶. With over 13 GW of AI company renewables already claimed³⁷, residual emissions could remain manageable with further expansion. However, achieving this best-case scenario could be extremely challenging, particularly due to the facility constraints in deploying AI servers at optimal locations and the difficulty in reaching ideal energy and water efficiency levels within AI data centres. Conversely, the worst practices pose a risk of unachievable net-zero pathways, signifying 71 Mt annual residual carbon emissions and over 5,224 million m³ annual residual water usage by 2030 under the low-demand scenario. Such an environmental cost is nearly impossible to be fully compensated during a short period.

Projected net-zero pathways could be further influenced by other uncertainties. Figure 6 presents a sensitivity analysis of key factors: server lifetime, AI server manufacturing capacity, US allocation ratio, server idle and maximum power ratios, and training/inference distributions. Low, applied and high values of each parameter are listed in the figure; left and right halves present the effects of lower and higher values, respectively. Changes in these factors result in up to 40% variations in energy consumption, which are closely mirrored by corresponding shifts in water footprint and carbon emissions. Unmodelled uncertainties can largely be reflected by modifying these simulation factors. For example, innovations such as DeepSeek can lead to low power requirements for AI computing tasks, which then may result in a rebound effect incorporating more AI applications. This statement may imply larger chip-on-wafer-on-substrate (CoWoS) manufacturing capacity due to expanding applications as well as longer lifetime and smaller training/job ratio due to more inference-based usage. As Fig. 6 indicates, the study’s key conclusions could remain robust unless future uncertainties greatly exceed modelled ranges. The modelling approach used in this study enables future revisions with more available data. The sensitivity analysis aims to capture the potential uncertainties inherent to this problem, offering insights into the magnitude of their influence on projection.

Sensitivity analysis for AI server energy consumption, water footprint and carbon emissions considering uncertainties of server lifetime, manufacturing capacities for AI servers, US allocation ratio, server idle power ratio, server maximum power ratio and training/inference distributions. The listed numbers for each parameter from left to right represent the considered low, applied and high values.

Source data

Full size image

Investment in AI servers is accelerating, as seen in projects such as the $500 billion Stargate³. While AI advancement is a key priority, our study highlights its environmental impact. To mitigate these risks, we identify that concentrating AI server deployment in Midwestern states—Texas, Montana, Nebraska and South Dakota—is optimal, given their abundant renewables, low water scarcity and favourable projected unit water and carbon intensities. These states possess substantial untapped wind and solar resources, enabling robust green power portfolios and reducing competition with other sectors. Their lower water stress also helps ease public concerns and reduces the need for costly water-saving measures. However, several implementation challenges must be acknowledged. Texas, crucial to the optimal strategy, may need to support an additional 74–178 TWh of AI server demand, possibly exceeding its current total renewable generation of 139 TWh (ref. ³⁸). This scale-up would require substantial investment in new renewable capacity and transmission infrastructure, which is already constrained by existing congestion³⁹. Meanwhile, Montana, Nebraska and South Dakota currently host minimal data-centre capacity, indicating that most of their existing internet infrastructure may support only residential or standard industrial applications. This raises potential connectivity and security concerns for high-performance AI services. Enabling AI-grade infrastructure in these regions will require broadband and security upgrades, with associated emissions and capital costs increasing the expense of the optimal strategy. These challenges underscore the complexity of sustainable siting decisions and highlight the need for strategic coordination to ensure that AI investments support both technological leadership and long-term sustainability goals.

Expanding AI servers in the suggested regions may be further obstructed by public health concerns and other impacts on local sustainable development. Operational demands and construction of supporting facilities could generate air and water pollution through fossil fuel use, substantial water consumption and large-scale construction and transportation. To address these issues while managing economic costs, we recommend that AI companies engage in public–private partnerships with local governments. Such partnerships could alleviate governmental budget pressures, create local jobs and mitigate health concerns by funding green power upgrades and monitoring systems, such as tracking PM_2.5 (particulate matter with a diameter of 2.5 micrometres or less) levels, through combined AI investment and regulation. For the governments, implementing tax exemptions on a related facility could incentivize a win–win development process, and it is essential to ensure transparent accountability within public–private partnerships to prevent privatized gains at public cost. Beyond environmental benefits, these measures could stimulate local economic growth, as evidenced in Virginia⁴⁰. This discussion outlines policy recommendations for both industry and legislative stakeholders that balance economic and environmental impacts. They also emphasize energy market adaptations as AI-driven hyperscale loads challenge utilities‘ ‘duty to serve’. Requirements to connect large data centres could lead to server collocation with energy generation, but as shown in this study, these choices should not crowd out renewable investments to decarbonize the economy. Finally, the effectiveness of the recommended policy may be influenced by broader political and economic factors beyond the scope of this study.

The challenges associated with the optimal spatial distribution of AI servers introduce substantial uncertainties to the projected reductions of 73% and 86% in carbon and water footprints, respectively, under best-practice scenarios. These challenges are unlikely to be offset by improvements in efficiency or decarbonization efforts beyond what has already been considered. Our best-case scenario for industry efficiency approaches the physical limits of AI data centres. Moreover, projections from the US Energy Information Administration offer limited support for additional grid decarbonization compared with the considered best case⁴¹. These constraints suggest that, without additional interventions, AI data centres are likely to generate substantial environmental impacts in the coming years. Consequently, AI companies are expected to continue investing in the offset mechanisms, including power purchase agreement, carbon removal and water restoration, reaching an unprecedented level of reliance on them to accomplish their net-zero aspirations by 2030. Nevertheless, severe concerns could be raised regarding the complexity of securing long-term contracts and providing convincing reduction credits^9,42, especially bearing in mind the considerable capacity that would be needed. We argue that AI companies should shift to a more transparent approach by closely cooperating with third-party verification groups, service providers and governmental agencies. Such collaboration could reduce uncertainties through better integration of social resources and serve as a model for other electrified sectors.

Emerging high-efficiency technologies in hardware and software, exemplified by DeepSeek⁴³, may fundamentally transform AI server supply and demand. As reflected in the study’s five scenarios, these innovations may cause deviations of up to 393 million m³ in water footprints and 20 MtCO₂-equivalent in emissions between minimum and maximum impact cases, underscoring the need for tailored managements. While efficiency gains may reduce cost per computing task, they risk a rebound effect, where lower costs increase total application volume. This dynamic, as reflected by the high-application scenario, may amplify total demand and complicate AI’s environmental trajectory. To address these uncertainties, we recommend government agencies work with industry to establish real-time monitoring systems, enabling timely alerts and proactive measures before considerable environmental impacts occur. Moreover, the potential increase in total computing jobs poses both challenges and opportunities, calling for ongoing enhancements in energy and water efficiency through system optimization and adoption of strategies such as SUO and ALC to manage the added workload complexity and flexibility. Therefore, we also suggest the data-centre industry establish AI-specific benchmarks for energy, water and carbon performance, which is crucial for continuous operational efficiency gains.

The methodology framework of this study aims to achieve two goals: (1) draft the energy–water–climate impacts of AI servers in the United States from 2024 to 2030 to handle the massive concerns about AI developments, and (2) identify the best and worst practices of each influencing factor to scheme the net-zero pathways for realizing water and climate targets set for 2030. Compared with many previous climate pathway studies, which often extend predictions to 2050 for better integrating climate goals, this study focuses on the period from 2024 to 2030 due to the great uncertainties surrounding the future of AI applications and hardware development. For assessing these uncertainties, scenario-based projections are first constructed to obtain potential capacity-increasing patterns of AI servers. Technology dynamics, such as SUO and ALC adoption, are defined with best, base and worst scenarios, and a similar method is employed to capture the impact of grid decarbonization and spatial distribution. The utilized models and data required during the calculation process are illustrated in the following sections. More details on model assumptions and data generation are provided in sections 1–4 of Supplementary Information.

Data description and discussion

This section provides a comprehensive overview of the data used in this study. Historical DGX (Nvidia’s high-performance AI server line) parameters were sourced from official documentation, and future scenarios were projected on the basis of historical configurations and current industry forecasts. To attain the units of AI servers, we collected the most updated industrial report data for projecting the future manufacturing capacity of CoWoS technology, which is the bottleneck for top-tier AI server production. The data resources of the preceding process have been introduced and validated in section 1 of Supplementary Information. AI server electricity usage was assessed using recent experimental data on maximum power^44,45, idle power^44,46 and utilization rate^46,47,48,49, derived from existing AI server systems. PUE and WUE values for AI data centres across different locations were calculated using operational data from previous studies^30,50 and industrial resources⁵¹, combined with the collected average climate data for each state⁵². The allocation ratios of AI servers to each state were determined on the basis of configurations of existing and planned AI data centres, which are collected from reports of major AI companies in the United States, as data resources detailed in section 2 of Supplementary Information. In addition, projections for grid carbon and water factors were derived from the ReEDS model³¹, using its default scenario data³³. All datasets employed in this study are publicly available, with most originating from well-established sources. A key uncertainty lies in estimating the number of manufactured AI server units, as official supply-chain reports remain largely opaque. To maintain transparency and ensure reproducibility, we rely on the best available industry reports rather than commercial sources such as International Data Cooperation data⁵³, which are not granted for open access and would limit future validation despite their potential to provide better estimates. The validations of applied data are further detailed in sections 1 and 4 of Supplementary Information.

AI server power capacity projections

The energy consumption of AI servers is projected to be driven predominantly by top-tier models designed for large-scale generative AI computing^6,7. This trend is attributed to their substantial power requirements and the increasing number of units being deployed. In this study, we estimate the power capacity of these high-performance AI servers by examining a critical manufacturing bottleneck: the CoWoS process⁵⁴. This process, which is controlled nearly exclusively by the Taiwan Semiconductor Manufacturing Company, serves as a key determinant of the manufacturing capacity for AI servers in recent years⁵⁵. Our analysis uses forecast data and projection assumptions of the CoWoS process to estimate total production capacity. Other factors are integral to translating this capacity into the power capacity of AI servers: the CoWoS size of AI chips, which determines how many chips can be produced by each wafer; the rated power of future AI servers, which reflects the power demand per unit; and the adoption patterns of AI servers, which dictate the mix of various server types over time. The values of these factors are derived mainly from the DGX systems produced by Nvidia, which is the dominant product for the top-tier AI server markets⁵⁶.

Considering the influencing factors for the total AI server capacity shipments and existing uncertainties, we generate distinct scenarios as follows:

• Mid-case scenario: the CoWoS capacity is projected to slightly increase after 2026, consistent with the growth rate in 2023. Under this scenario, AI servers’ rated power is expected to have a linear relationship with the anticipated die size increase while adoption patterns remain aligned with current trajectory.

• Low-demand scenario: characterized by lower CoWoS capacity growth and lower AI server rated power compared with the mid-case scenario, this reflects a scenario of lower overall demand for AI servers.

• Low-power scenario: maintains the same assumptions as the mid-case scenario but with lower AI server rated power, representing efficiency gains in AI hardware and software development.

• High-application scenario: assumes lower AI server rated power alongside high CoWoS capacity, capturing the potential rebound effect where efficiency gains drive increased AI workload deployment.

• High-demand scenario: features higher CoWoS capacity expansion, higher AI server rated power and higher adoption of new servers compared with the mid-case scenario, reflecting a scenario of strong AI server demand.

Based on the assumptions and scenarios outlined, the annual projections for top-tier AI server shipments and their average rated power are calculated as follows:

where $N_{\mathrm{A}\mathrm{I}}$ and ${\overline{P}}_{\mathrm{A}\mathrm{I}}$ represent the annually projected shipments and average rated power of the top-tier AI servers. $R_{\mathrm{A}\mathrm{I}}$ is the ratio of CoWoS capacity allocated to top-tier AI servers and is set as 40% for 2022, 40.7% for 2023, 48.5% for 2024 and 54.3% for 2025, according to industry reports^57,58. For years beyond 2025, this ratio is assumed to remain constant at the 2025 value due to a lack of further data. The sensitivity analysis regarding this value is provided in Fig. 6. $C_{\mathrm{C}\mathrm{o}\mathrm{W}\mathrm{o}\mathrm{S}}$ is the projected CoWoS capacity within each scenario. $N_{\mathrm{G}\mathrm{P}\mathrm{U}}$ is the number of graphic processor units (GPUs) per server and is set as 8, reflecting the configuration of most commonly used AI server systems⁵⁹. In addition, $R_i$, $n_i$ and $P_i$ represent the projected adoption ratio, units yield per CoWoS wafer and rated power of the ith type of chip at each year, respectively. The details of the projections and related data resources are provided in section 1 of Supplementary Information, Supplementary Figs. 1–4 and Supplementary Table 1.

AI server electricity usage calculation

The applied AI server electricity usage model is a utilization-based approach initially derived from CPU (central processing unit)-dominant servers⁶⁰ and can be written as the following:

The preceding model assumes the total server power has a linear relationship with the processor utilization rate u. While this relationship has been well validated for CPU machines, its application to GPU utilization is less established except for a few cases⁶¹. However, several recent studies have shown a strong correlation between GPU utilization and overall server power consumption when dealing with AI workloads^44,46, indicating that GPUs are the dominant contributors to energy use in AI servers⁴⁵. Although systematic experimental validation specific to GPUs is still limited, the consistency of findings across various case studies supports the assumption that the linear relationship applies here as well. The maximum power $P_{max}$ and idle power $P_{\mathrm{i}\mathrm{d}\mathrm{l}\mathrm{e}}$ are generated on the basis of the recent DGX system experimental results, and their values are set as 23% and 88% of the server rated power, respectively^44,46. The sensitivity analysis was conducted to quantify the uncertainty, as shown in Supplementary Fig. 6. Moreover, the GPU processor utilization $u$ is calculated as the following:

where $u_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}$ and $r_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}$ represent the average processor utilization of active GPUs and the ratio of active GPUs to total GPUs, respectively. Note that the $u_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}$ and $r_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}$ commonly have higher values during training compared with inference⁶². Specifically, we use currently available AI traces, including Philly trace⁴⁷, Helios trace⁴⁶, PAI trace⁴⁸ and Acme trace⁴⁹, to determine the $r_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}$ for training and inference tasks. These traces provide comprehensive analyses on the relationship between GPU utilization rate and job characteristics. Based on the data provided in these works, the $r_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}$ is set as 50% and 90% for inference and training, respectively. Moreover, the $u_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}$ values are further determined on the basis of recent experimental studies⁴⁴. The values are set as 50% and 80% for inference and training, respectively. Therefore, the processor utilization rates for inference and training in this work are set as 25% and 72%, respectively. Following the previous works^63,64, our base estimations assume 30% of computing capacity for training and 70% for inference. A detailed sensitivity analysis on the impact of these utilization rate settings is provided in Fig. 6.

Assessment of the environmental footprints of AI servers

This study employs a state-level allocation method to evaluate the energy, water and carbon footprints of AI servers. To capture the current and future distributions of AI server capacity, we compiled data of current and in-construction large-scale data centres belonging to major purchasers of top-tier AI servers, including Google, Meta, Microsoft, AWS, XAI and Tesla. The analysis incorporates the location, building area and construction year of each data centre to calculate the state-level distribution of server capacity by annually aggregating the total building area for each state. On the basis of our calculations, no major changes in spatial distribution are projected between 2024 and 2030, even with the anticipated addition of new data centres. Therefore, we assume the current spatial distribution will remain constant from 2024 to 2030 to account for uncertainties in directly integrating the projected contributions of in-construction data centres. Further details on the methodology and spatial distribution results are provided in section 2 of Supplementary Information.

For each state, the actual energy consumption can be derived from the server electricity usage and the PUE value of AI data centres. Meanwhile, the water footprint and carbon emissions should be analysed across three scopes. Scope 1 encompasses the on-site water footprint, calculated on the basis of on-site WUE (shortened as WUE in this work) and on-site carbon emissions (typically negligible for data centres¹⁸). Scope 2 includes off-site water footprint and carbon emissions, which are contingent on the local grid power supply portfolio. Scope 3, representing embodied water footprint and carbon emissions during facility manufacturing, lies beyond the spatial scope of this study. A regional PUE and WUE model, following the idea in previous research^30,50, is applied to estimate the PUE and WUE values of AI data centres in different states. This hybrid model integrates thermodynamics and statistical data to generate estimations on the basis of local climate data. Specifically, we collected the average climate data of each state between 2024 and 2030 from an existing climate model⁵², which is then employed in calculating the PUE and WUE values of each state. Considering that the specific cooling settings for AI data centres are unknown, the base values are calculated by averaging the worst and best cases. The model parameters are detailed in Supplementary Table 2. Subsequently, the Scope 2 water footprint and carbon emissions are calculated on the basis of the grid water and carbon factors derived from the ReEDS model³¹. This approach also allows us to incorporate the projected data-centre load data, which can further interact with the grid system through services such as demand response. The validation of the ReEDS model results by using current high-resolution data is presented in Supplementary Figs. 7 and 8, and the related discussion is presented in section 4 of Supplementary Information. Optimization and analytical techniques are employed to determine optimal parameters during the simulation to generate the best and worst practices concerning industrial efficiency efforts, spatial distributions and grid decarbonization. Moreover, the water scarcity and remaining renewable energy potential data of each state are computed on the basis of the calculated environmental cost and standard data from previous literature^18,65. The preceding calculation process depends mainly on previously established approaches, and its integration into our framework is further discussed in sections 3 and 4 of Supplementary Information.

Uncertainties and limitations

There are substantial uncertainties inherent in projecting the evolution of AI servers. Our analysis presents a range of scenarios based on current data to evaluate the impacts of data-centre operational efficiency, spatial distribution and grid development. However, several key uncertainties remain an unmodelled field for this work. For a better understanding of our study and to outline future research directions, these uncertainties are categorized as follows:

• Model and algorithm innovations: the model and algorithm breakthroughs in the AI industry could fundamentally alter computing requirements.

• Supply-chain uncertainties: the complex production process of AI servers may reveal new bottlenecks beyond the current CoWoS technology, leading to varying expansion patterns.

• Hardware and facility evolutions: continued improvements in AI computing hardware and data-centre efficiency may substantially affect the environmental impact of these servers.

• Out-of-scope factors: there are other major contributors that are out of the scope of this study and would be critical to the process, such as market forces and geopolitical influences.

The impacts of these factors are multifaceted and challenging to model with existing data. For example, while the recent release of DeepSeek has been interpreted as reducing the energy demands of AI servers, it may also trigger a rebound effect by spurring increased AI computing activity, ultimately resulting in higher overall energy, water and carbon footprints⁴³. However, no fresh data have become available to simulate this complex process, based on our best knowledge when drafting. To further assess the influence of unpredictable uncertainties, we conducted a sensitivity analysis on key factors, including manufacturing capacities for AI servers, US allocation ratios, server lifetimes, idle and maximum power ratios and training/inference distributions. As shown in Fig. 6, our findings suggest that the key conclusions of this study are expected to remain robust as long as the impact of future uncertainties does not notable exceed the ranges considered. Given the highly dynamic nature of AI evolution, our modelling approach allows for future revisions as more data become available on potential shifts in industry trends.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.