AI Industry Faces Environmental Crisis as New Study Reveals Massive Water and Carbon Impact Through 2030

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

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. 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, idle power and utilization rate, 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 and industrial resources, combined with the collected average climate data for each state. 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 model, using its default scenario data. 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 data, 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. The energy consumption of AI servers is projected to be driven predominantly by top-tier models designed for large-scale generative AI computing. 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 process. 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 years. 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 markets. Considering the influencing factors for the total AI server capacity shipments and existing uncertainties, we generate distinct scenarios as follows: 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 and represent the annually projected shipments and average rated power of the top-tier AI servers. 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. 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. is the projected CoWoS capacity within each scenario. is the number of graphic processor units (GPUs) per server and is set as 8, reflecting the configuration of most commonly used AI server systems. In addition, , and 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. The applied AI server electricity usage model is a utilization-based approach initially derived from CPU (central processing unit)-dominant servers 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 cases. However, several recent studies have shown a strong correlation between GPU utilization and overall server power consumption when dealing with AI workloads, indicating that GPUs are the dominant contributors to energy use in AI servers. 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 and idle power 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. The sensitivity analysis was conducted to quantify the uncertainty, as shown in Supplementary Fig. 6. Moreover, the GPU processor utilization is calculated as the following: where and represent the average processor utilization of active GPUs and the ratio of active GPUs to total GPUs, respectively. Note that the and commonly have higher values during training compared with inference. Specifically, we use currently available AI traces, including Philly trace, Helios trace, PAI trace and Acme trace, to determine the 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 is set as 50% and 90% for inference and training, respectively. Moreover, the values are further determined on the basis of recent experimental studies. 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, 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. 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 centres). 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, 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 model, 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 model. 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. 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. 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: 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 footprints. 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. Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

AI power use forecast finds the industry far off track to net zero

Several large tech firms that are active in AI have set goals to hit net zero by 2030, but a new forecast of the energy and water required to run large data centres shows they're unlikely to meet those targets As the AI industry rapidly expands, questions about the environmental impact of data centres are coming to the forefront - and a new forecast warns the industry is unlikely to meet net zero targets by 2030. Fengqi You at Cornell University in New York and his colleagues modelled how much energy, water and carbon today's leading AI servers could use by 2030, taking into account different growth scenarios and possible data centre locations within the United States. They combined projected chip supply, server power usage and cooling efficiency with state-by-state electrical grid data to conduct their analysis. While not every AI company has set a net zero target, some larger tech firms that are active in AI, such as Google, Microsoft and Meta have set goals with a deadline of 2030. "The rapid growth of AI computing is basically reshaping everything," says You. "We're trying to understand how, as a sector grows, what's going to be the impact?" Their estimates suggest US AI server buildout will require between 731 million and 1.125 billion additional cubic metres of water by 2030, while emitting the equivalent of between 24 and 44 million tonnes of carbon dioxide a year. The forecast depends on how fast AI demand grows, how many high-end servers can actually be built and where new US data centres are located. The researchers modelled five scenarios based on the speed of growth, and identified various ways to reduce the impact. "Number one is location, location, location," says You. Placing data centres in Midwestern states, where water is more available and the energy grid is powered by a higher proportion of renewables, can reduce the impact. The team also pinpoints decarbonising energy supplies and improving the efficiency of data centre computing and cooling processes as major ways to limit the impact. Collectively, those three approaches could cut the industry's emissions by 73 per cent and its water footprint by 86 per cent. But the group's projections could also be scuppered by public opposition to data centre installations because of their potentially extractive impact on the environment. In Virginia, which hosts about one-eighth of global data centre capacity, residents have begun lodging opposition to further planned construction, citing the impact on their water reserves and the wider environment. Similar petitions against data centres have been lodged in Pennsylvania, Texas, Arizona, California and Oregon. Figures from Data Center Watch, a research firm tracking data centre development, suggests local opposition has stymied $64 billion worth of projects. However, it is unclear, even in places that have successfully rejected data centres, just how much power and water they may use. That is why the new findings have been welcomed - albeit cautiously - by those who have attempted to study and quantify AI's environmental impact. "AI is such a fast-moving field that it's really hard to make any kind of meaningful future projections," says Sasha Luccioni at AI company Hugging Face. "As the authors themselves say, the breakthroughs in the industry could fundamentally alter computing and energy requirements, like what we've seen with DeepSeek", which used different techniques to reduce brute-force computation. Chris Preist at the University of Bristol in the UK says, "the authors are right to point out the need to invest in additional renewable energy capacity", and adds data centre location matters. "I think their assumptions regarding water use to directly cool AI data centres are pretty pessimistic," he says, suggesting the model's "best case" scenario is more like "business as usual" for data centres these days. Luccioni believes the paper highlights what is missing in the AI world: "more transparency". She explains that could be fixed by "requiring model developers to track and report their compute and energy use, and to provide this information to users and policymakers and to make firm commitments to reduce their overall environmental impacts, including emissions".