Impact Statement

In this article, we focus on the impact of growing infrastructure needs for AI, specifically data centers, and the increasing energy demand required to power them. While the energy mix that will supply these data centers is important, our main emphasis is on CO₂ emissions and their broader climate impact. In determining whether AI results in a net positive or negative contribution to emissions, it is essential to consider scenarios both with and without dedicated emissions reduction measures. Additionally, we consider the carbon benefits directly attributed to AI across various sectors of the energy industry.

2. Methodology

To accurately estimate CO₂ emissions from data centers and AI workloads, a refined methodology is necessary. This involves moving beyond simple estimations of data center energy consumption to utilizing more granular data and applying location-specific, time-varying emission factors that explicitly incorporate uncertainty. Near-term scenarios should quantify the growth of data centers and AI workloads with defined probability distributions, reflecting the inherent unpredictability. Long-term projections need to model potential efficiency gains in AI hardware and software, the indirect impact of AI in optimizing energy systems and increasing renewable penetration, and the uncertain trajectory of the energy mix. By employing MC simulations with these probabilistically defined inputs, we can generate a range of potential emission scenarios for both the near and long term. This approach allows for a more nuanced understanding of the uncertainties involved and enables sensitivity analysis based on the factors with the most significant influence on future data center and AI-related emissions. The ultimate goal is to provide robust, probabilistic insights that can inform strategies for mitigating the environmental impact of this rapidly growing sector.

3. The AI boom and its energy appetite

The AI landscape is experiencing unprecedented expansion, particularly in the areas of deep learning and large language models (LLMs). This surge in AI capabilities has been driven by breakthroughs in algorithms, the availability of large datasets, and advancements in computing hardware. Since the advent of ChatGPT, the capabilities of these LLMs have evolved from simple text generation and summarization to advanced reasoning abilities with remarkable accuracy and acceptance. These models are now routinely used on desktops (e.g., Copilot) and for content generation by creative artists. The bulk of these AI models are termed foundational models, which are trained on large datasets and are versatile and suitable for numerous downstream applications. GPT and Llama are both examples of these foundational models. Figure 1 shows the number of foundational models with more than 10 billion parameters built between 2021 and 2024 (data courtesy of EcoSystem Graphs, (n.d.). It also illustrates the increase in revenues from high-density graphics processing units (GPUs) from 2022 to 2028 (Source: ElectroIQ, n.d.). As can be seen, more investments are being made toward building GPU servers to support the development of large AI models. Although NVIDIA GPUs dominate the market for chips and low-level software libraries, there are no barriers for new entrants to join the arena of building new foundational models. Due to supply and demand mismatches and other approval process-related issues, commissioning grid infrastructure has temporarily slowed the pace of data center growth. However, the overall trend remains one of rapid expansion.

Figure 1.

The figure illustrates the growth of foundational models between 2021 and 2024 and the revenue growth of high-density GPU markets. Data courtesy of EcoSystem Graphs (n.d.).

Figure 1. Long description

The first panel on the left is a bar chart titled Number of Foundational Models by Year. The Y axis represents the Number of Foundational Models ranging from 0 to 120. The X axis represents the Year. The data shows a significant upward trend. In 2021, there are approximately 5 models. In 2022, the count jumps to approximately 65. In 2023, it reaches 90. By 2024, the number peaks at approximately 125 models.

The second panel on the right is a bar chart titled High-Density Server Graphics Processing Unit G P U Market Revenue Worldwide in 2022, 2023, and 2028. The Y axis represents revenue in billion U S dollars from 0 to 50. The X axis lists three specific years. In 2022, revenue is 7.6 billion. In 2023, revenue increases slightly to 9.8 billion. For the year 2028, the chart shows a projected surge to 38.5 billion dollars.

However, this progress comes with significant consequences: a rapidly increasing demand for computational resources required to support the tasks in the life cycle of AI models, which involves both training (a one-time or occasional effort) and inference (repeatedly executed when the model is deployed for consumption). Although the work done (measured in FLOPs) per dollar and per watt has increased, meaning the power consumption per FLOP has decreased (see Figure 2 [left]), the number of FLOPS required to train AI models has continued to rise (see Figure 2 [right]). Consequently, the total power consumption for training AI models keeps increasing. This translates directly into escalating energy consumption (AI to Drive 165% Increase in Data Center Power Demand by 2030, 2025). For instance, it was reported in Patterson et al. (Reference Patterson, Gonzalez, Le, Liang, Munguia, Rothchild, So, Texier and Dean2021) that training the LLM powering ChatGPT-3 consumed almost 1300 megawatt hours of energy, equivalent to the annual energy usage of 130 American homes. Furthermore, an analysis by OpenAI suggests that the amount of power needed to train AI models has grown exponentially since 2012, doubling roughly every 3.4 months as the models become more sophisticated (AI and Compute, 2018).

Figure 2.

The figure illustrates the decrease in energy consumption per GFLOPS over the years (left) and the number of FLOPS required to train different models (right).

Figure 2. Long description

The first panel is a line graph titled Energy Consumption per G F L O P S Over the Years. The x-axis represents Year from 2000 to 2025. The y-axis represents Energy Consumption in joules per G F L O P S from 0.0 to 1.0. A blue line shows a steep exponential-like decrease from 1.0 in the year 2000 to 0.1 in 2010, continuing to decline toward nearly 0.0 by 2025.

The second panel is a bar chart titled Number of F L O P S by Model. The x-axis lists four models: Transformer, G A N, G P T-3, and G P T-4. The y-axis is a log scale representing the Number of F L O P S from 10 super 11 to 10 super 25.

* Transformer and G A N bars are the lowest, positioned near 10 super 11.

* G P T-3 shows a massive increase, reaching approximately 10 super 23.

* G P T-4 is the highest, exceeding 10 super 25.

3.1. Computational demands of training and inference

AI models, especially deep learning models and LLMs, require immense computational power for both training and inference (de Vries, Reference de Vries2023).

• • Training involves feeding the model vast amounts of data and adjusting its parameters to achieve the desired level of accuracy. This process can take weeks or even months and accounts for a significant portion (30%–40%) of an AI model’s total electricity consumption (see Figure 3 for the electricity consumption of training several prominent LLMs).

• • Inference is the deployment and utilization of the trained AI models in real-world applications, where the model interprets new data and generates outcomes or predictions. Despite innovations in model efficiency, the inference phase accounts for the majority (60%) of the energy footprint (Power Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption, 2024) (see Supplementary Table 1 for electricity consumption of various tasks that AI chatbots are typically asked to execute).

Figure 3.

The figure illustrates the electricity consumption of several prominent large language models published in recent years. Although the authors usually do not provide these statistics directly, some do. Many factors influence electricity consumption, but typically, if they disclose the hardware details—such as the number of GPU hours used and the types of GPUs—it is possible to calculate the overall power consumption using the GPUs’ wattage. However, this calculation does not account for other components like CPUs, data storage, and cooling technology. Nonetheless, it indicates the resources required to train these models.

Figure 3. Long description

The chart is titled Energy Consumption of Different Models in M W h. The y-axis is labeled Model and the x-axis is labeled M W h with a scale from 0 to 1750 in increments of 250. The orange bars represent the following data from top to bottom.

* G P T 4: approximately 1750 M W h.

* G P T 3: approximately 1280 M W h.

* Gopher: approximately 1060 M W h.

* Llama 2: approximately 750 M W h.

* Llama 3.1: approximately 500 M W h.

* BLOOM: approximately 430 M W h.

* O P T: approximately 320 M W h.

* T 5: approximately 10 M W h.

While this section focuses heavily on LLMs due to their high energy intensity per query, it is important to note that LLMs represent only one manifestation of AI. Consumer-facing generative AI accounts for a massive surge in inference compute; however, the industrial applications of AI discussed later in Section 7—such as grid optimization and CCS surrogate modeling—often utilize smaller, purpose-built models (e.g., Fourier neural operators) or inference-light architectures. Therefore, the energy footprint calculated here represents a high-intensity scenario dominated by generative AI, rather than a universal baseline for all machine learning applications.

While the trends clearly indicate a rise in the computational resources required for AI models, understanding the global consumption trends for all AI-related workloads necessitates knowing where these models are being trained and hosted during the post-deployment phase. These details will be discussed in the next section.

4. Data centers: The epicenter of AI’s energy demand

Data centers are the backbone of the AI ecosystem, serving as the primary energy consumers that power the computational workloads required for training and deploying AI models. These facilities house vast arrays of servers, storage devices, and networking equipment, all of which demand significant amounts of electricity to operate. The rapid expansion of AI, particularly deep learning and LLMs, has led to an unprecedented surge in data center growth, intensifying concerns about their environmental impact (Power Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption, 2024).

The data center industry is experiencing rapid growth worldwide, driven by the increasing demand for cloud computing, digital services, cryptocurrencies, and AI applications. The number of data centers is increasing, and their sizes are also expanding. As of March 2024, there were approximately 9,580 data centers globally, with more than half of them located in the United States (Data Center Map, 2024). Between 2015 and 2024, the electricity consumption at data centers increased from under 200 TWh to over 400 TWh (Energy and AI, 2025). Data centers are projected to consume between 4.6% and 9.1% of US electricity generation annually by 2030, up from an estimated 4% today. Some forecasts even predicted that data centers could draw up to 21% of the world’s electricity supply by 2030 (Jones, Reference Jones2018). In the United States, 15 states account for 80% of the national data center load, with data centers comprising a quarter of Virginia’s electric load in 2023. New data centers are being built with capacities ranging from 100 to 1000 megawatts, equivalent to the load from 80,000 to 800,000 homes.

It is critical to distinguish between the baseline energy consumption of data centers and the marginal growth driven by AI. Historically, data center energy use has not grown exponentially despite rising demand, due to significant efficiency gains (Masanet et al., Reference Masanet, Shehabi, Lei, Smith and Koomey2020). Current estimates suggest that AI workloads account for approximately 10%–20% of total data center electricity consumption (Energy and AI, 2025). However, AI represents the primary driver of new capacity demand. In the “Power Intelligence” report by EPRI Power Intelligence: Analyzing AI and Data Center Energy Consumption (2024), the authors note that in high-growth scenarios, the majority of the additional load is attributable to AI-specific infrastructure, distinct from traditional streaming or cloud storage services.

In Power Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption (2024), the authors have developed four growth scenarios for US data center electricity load from 2023 to 2030: low growth (3.7% annual), moderate growth (5% annual), high growth (10% annual), and higher growth (15% annual). In the higher growth scenario, data center loads are projected to increase by a staggering 166% from 152 TWh/year in 2023 to 403 TWh/year in 2030, with the majority of this demand attributed to AI. Such projections highlight the significant energy demands of AI and data centers and the need for innovative energy management and infrastructure planning to accommodate this growth.

Compared to traditional data center infrastructure, AI-focused data centers have significantly different operational requirements and substantially higher energy demands. Hyperscale data centers, often required for AI, contain a minimum of 5,000 servers and at least 10,000 sq ft of space, enabling extreme scalability. These AI-ready data centers rely on accelerators such as GPUs (Graphical Processing Unit), NPUs (Neural Processing Unit), and tensor processing units (TPUs) to support machine learning and deep learning workloads, which generate far greater heat loads than conventional CPUs (Central Processing Unit). As a result, they require advanced cooling solutions to maintain reliability and prevent outages.

To manage these high thermal densities, many operators are shifting from air-based cooling to liquid-based technologies. In addition to direct-to-chip liquid cooling, immersion cooling—where entire servers are submerged in thermally conductive dielectric fluids—is emerging as one of the most energy-efficient approaches. Immersion cooling can dramatically improve heat-transfer efficiency, lower fan and pump power consumption, and deliver superior power usage effectiveness (PUE). Because these systems operate with fully contained coolant loops, they can also reduce or eliminate the need for evaporative cooling, thereby decreasing overall water consumption and improving water usage effectiveness (WUE). By enabling higher rack densities while lowering operational overhead, immersion cooling helps AI data centers scale more efficiently with a smaller environmental footprint.

Beyond cooling, AI data centers require high-speed networking and ultra-low-latency interconnects to move data rapidly between distributed accelerators. They also depend on specialized storage and data management systems for handling large volumes of unstructured data, such as NVMe (Non-Volatile Memory Express) SSDs (Solid-state drives) and parallel file systems like Lustre. While these capabilities support rapid scaling and high computational throughput, they also drive higher capital costs due to the increased expense of accelerators, immersion-ready infrastructure, and next-generation networking components.

Despite increasing investment in renewable energy, many regions continue to rely heavily on fossil fuels—particularly natural gas—to meet growing electricity demands (Rotman, Reference Rotman2025). Consequently, a significant share of data center energy consumption still contributes directly to GHG emissions. In regions with limited access to clean power, the surge in electricity demand from AI workloads can deepen reliance on fossil generation, complicating efforts to decarbonize the energy sector. As noted in Rotman (Reference Rotman2025), natural gas is often viewed as the only reliable and affordable option for meeting the constant 24/7 electricity load associated with large AI-focused data centers.

5. The energy consumption scenarios and near-term outlook for data centers and AI

5.1. Shell’s energy security scenarios and the role of AI

Shell’s 2025 Energy Security Scenarios explore AI’s transformative role across different pathways, each highlighting a distinct impact on energy systems and societal change (Shell plc, 2025). In the Surge scenario, AI is a key driver of economic growth and societal transformation. AI optimizes manufacturing processes, enhances productivity across various sectors, and leads to increased economic activity. This surge in economic activity results in higher energy demand, which is managed efficiently by AI-powered smart grids. Additionally, AI accelerates the transition to low-carbon energy sources by optimizing the integration of renewable energy and enhancing energy storage solutions. The Archipelagos scenario presents a different picture, where AI development is hindered by global concerns about resource security, border control, and trade restrictions. Geopolitical tensions limit the sharing of AI technologies and data, slowing progress. As a result, technological advancements are slower, impacting the transformation of energy systems. In the Horizon scenario, AI plays a crucial role in achieving net-zero emissions by 2050. AI-driven innovations in energy efficiency and carbon capture technologies are pivotal in reaching this goal. AI also helps limit global warming to 1.5°C by optimizing energy consumption and reducing GHG emissions (Figure 4).

Figure 4.

Electricity generation and associated CO₂ emissions between 2020 and 2060 under the three Shell Energy Security Scenarios. Of those, only Horizon is reaching Net Zero by the 2050s (Courtesy of Shell Energy Security Scenarios, 2025).

Figure 4. Long description

A two-panel figure showing energy projections.

Left Panel: Electricity Generation 2020 to 2060. The y-axis measures T W h from 0 to 90000. The x-axis shows years from 2020 to 2060. All three scenarios start at approximately 23000 T W h in 2020.

- Surge (orange line) shows a steep linear increase, reaching the highest point at approximately 85000 T W h by 2060.

- Archipelagos (green line) follows a steady upward trend, reaching approximately 66000 T W h by 2060.

- Horizon (blue line) tracks closely with Archipelagos until 2040, then curves slightly higher to end near 66000 T W h.

Right Panel: C O sub 2 from Electricity Generation 2020 to 2060. The y-axis measures M t C O sub 2 forward slash year from negative 2000 to 9000. The x-axis shows years from 2020 to 2060. All scenarios start near 4800 M t in 2020.

- Surge (orange line) peaks at approximately 8200 M t around 2042 before declining to 5300 M t by 2060.

- Archipelagos (green line) rises slowly to a plateau of approximately 7400 M t between 2050 and 2054, ending at 6900 M t in 2060.

- Horizon (blue line) peaks earliest at 6400 M t in 2034, then enters a steep decline, crossing the zero line in the late 2050s and reaching approximately negative 800 M t by 2060.

Each scenario underscores the diverse roles AI can play in shaping future energy landscapes. The Surge scenario emphasizes AI’s potential to drive rapid economic growth and energy demand, while the Archipelagos scenario highlights the constraints and risks associated with geopolitical tensions. The Horizon scenario focuses on AI’s role in achieving sustainability goals and mitigating climate change.

5.3. AI scenarios

In a recent study (Paccou and Wijnhoven, Reference Paccou and Wijnhoven2024), the authors presented a comprehensive outlook using a system dynamics approach to model four scenarios for AI electricity consumption: Sustainable AI, Limits To Growth, Abundance Without Boundaries, and Energy Crisis. These scenarios provide insights into potential future trajectories of AI electricity use, influenced by current AI infrastructure decisions. While all scenarios show initial growth, their paths diverge after 2030 due to various internal and external factors (Figure 5b). The Sustainable AI scenario prioritizes efficiency and sustainability, with energy consumption increasing from 100 TWh in 2025 to 785 TWh in 2035. This scenario balances technological advancement with environmental stewardship. In contrast, the Limits To Growth scenario shows minimal increase in consumption, from 510 TWh in 2030 to 570 TWh in 2035, indicating stunted economic expansion due to constraints like power availability and data scarcity. The Abundance Without Boundaries scenario projects unchecked growth, with consumption reaching 1370 TWh by 2035. This scenario highlights risks such as power centralization and inequitable AI access. Meanwhile, the Energy Crisis scenario underscores the risks of mismatched energy demand and infrastructure, with consumption peaking at 670 TWh before dropping to 190 TWh by 2035, indicating a potential energy crisis.

Figure 5.

(a) Four scenarios for electricity demand by data centers globally between 2020 and 2035 (reproduced from Energy and AI, 2025); (b) Plausible scenarios for electricity usage by AI applications between 2020 and 2035 (reproduced from Paccou and Wijnhoven, Reference Paccou and Wijnhoven2024).

Figure 5. Long description

Panel a is a line graph titled Global electricity demand from data centers 2020-2035. The X axis shows years from 2020 to 2035 and the Y axis shows T W h from 0 to 2000. All four scenarios start at approximately 270 T W h in 2020 and diverge after 2024. The Lift-Off Case shows the steepest exponential growth reaching about 1700 T W h by 2035. The Base Case follows a steady upward curve to 1200 T W h. The High Efficiency Case reaches nearly 1000 T W h. The Headwinds Case shows the slowest growth leveling off around 700 T W h.

Panel b is a line graph titled A I Electricity Usage Scenarios Between 2020 and 2035. The X axis shows years from 2020 to 2035 and the Y axis shows T W h from 0 to 1600. All scenarios show near-zero usage until 2025. The Abundance Without Boundaries A I scenario rises sharply to 1400 T W h by 2035. The Sustainable A I scenario reaches approximately 800 T W h. The Limits To Growth A I scenario peaks around 600 T W h in 2032 and plateaus. The Energy Crisis A I scenario peaks at 700 T W h in 2030 before dropping sharply to 200 T W h by 2035.

Specific insights from these scenarios include the dominance of generative AI in the Sustainable AI scenario, focusing on less energy-intensive models. The Limits To Growth scenario highlights constraints in AI development due to power, chip manufacturing, and data scarcity. The Abundance Without Boundaries scenario warns of risks associated with unchecked AI growth, including unsustainable infrastructure costs, global inequality, and increased e-waste. The Energy Crisis scenario emphasizes the likelihood of underestimated future electricity needs due to insufficient grid planning and inaccurate demand projections. The authors note that projections for electricity consumption in TWh face challenges due to the nascent nature of the discipline, making direct comparisons difficult. However, some convergence is observed: the AI Abundance Without Boundaries scenario projects 880 TWh usage by 2030, aligning with estimates by Patel et al. (Reference Patel, Nishball and Ontiveros2024). Another projection by Kindi (Reference Kindig2024) of 652 TWh by 2030 aligns with the Sustainable AI scenario, suggesting a balanced growth trajectory.

To provide a comprehensive understanding of the potential futures for AI development and energy consumption, we have matched the Shell scenarios with the IEA sensitivity cases and the AI scenarios from Paccou and Wijnhoven (Reference Paccou and Wijnhoven2024). This alignment is based on the shared themes and underlying assumptions of each scenario. The Shell scenarios offer a broad perspective on global economic and technological trends, while the IEA sensitivity cases focus on the specific impacts of AI on the energy sector. The AI scenarios from Paccou and Wijnhoven (Reference Paccou and Wijnhoven2024) provide detailed insights into the potential trajectories of AI electricity consumption. By matching these scenarios, we can explore how different factors such as geopolitical tensions, technological advancements, and sustainability efforts might influence the future of AI and energy. This integrated approach allows us to identify common challenges and opportunities, offering a more nuanced and holistic view of the interplay between AI development and energy consumption:

1. 1. Surge—Lift-Off Case—Abundance Without Boundaries: The Surge scenario envisions rapid AI adoption and technological advancements, driving significant economic growth. Similarly, the Lift-Off Case from the IEA assumes high AI adoption rates and proactive measures to reduce energy bottlenecks, highlighting AI’s transformative potential. The Abundance Without Boundaries scenario by Paccou and Wijnhoven (Reference Paccou and Wijnhoven2024) projects unrestricted AI growth, emphasizing its power to drive economic and technological progress without major constraints. These scenarios share a vision of rapid and expansive AI growth fueling economic and technological advancements. Under these scenarios, the energy demand by data centers continues to grow past 2035, exceeding 1700 TWh. AI accounts for more than three-quarters of the electricity usage (Figure 6a).

2. 2. Archipelagos—Headwinds Case—Limits To Growth: The Archipelagos scenario depicts technology development hindered by global concerns about resources, borders, and trade security, leading to fragmented efforts and slower progress. Similarly, the Headwinds Case from the IEA includes various bottlenecks, such as macroeconomic challenges and infrastructure delays, representing cautious growth with significant obstacles. The Limits To Growth scenario by Paccou and Wijnhoven (Reference Paccou and Wijnhoven2024) highlights constraints on AI development due to resource limitations, emphasizing the need for careful management of AI’s energy demands. These scenarios share a focus on limitations and challenges that hinder AI and energy sector progress. Under these scenarios, energy demand by data centers plateaus by 2030 at around 700 TWh. AI accounts for approximately 80% of the electricity usage (Figure 6b).

3. 3. Horizon—High Efficiency Case—Sustainable AI: The Horizon scenario focuses on achieving net-zero emissions by 2050 and limiting global warming to 1.5 °C, emphasizing sustainability and efficiency in energy use. Similarly, the High Efficiency Case from the IEA assumes significant improvements in energy efficiency, both in AI technologies and in the broader energy sector, envisioning reduced energy consumption per unit of AI output. The Sustainable AI scenario by Paccou and Wijnhoven (Reference Paccou and Wijnhoven2024) balances AI development with sustainability, highlighting the importance of generative AI inferencing and traditional AI in decarbonization efforts. These scenarios share an emphasis on achieving sustainability and efficiency in AI and energy consumption. Under these scenarios, energy demand by data centers is approaching a plateau by 2035 at nearly 1,000 TWh. AI accounts for approximately 80% of the electricity usage (Figure 6c).

4. 4. Baseline Case—Energy Crisis: The Baseline Case from the IEA assumes current trends and policies continue without significant changes, serving as a benchmark for comparing other scenarios and representing a future where AI development follows existing trajectories. The Energy Crisis scenario by Paccou and Wijnhoven (Reference Paccou and Wijnhoven2024) considers the possibility of an energy crisis caused by AI, highlighting the risks of unmanaged AI development leading to significant energy consumption and potential shortages. It emphasizes the need for strategic planning and management to avoid such crises. These scenarios share a focus on the potential risks and challenges associated with AI development and energy consumption. Under these scenarios, energy demand by data centers is approaching a plateau in 2035 at around 1,200 TWh. AI electricity usage reaches its maximum in 2030, when it accounts for 70% of the data center’s electricity usage. However, due to the energy crisis, it drops to around 16% by 2035 (Figure 6d).

Figure 6.

Historical trends and forecasts for AI electricity demand through 2035. Depending on the scenario, it is estimated that electricity demand from data centers could range between 500 and 1700 TWh, with AI workload expected to pick between 600 and 1400 TWh at some point between 2030 and 2035.

Figure 6. Long description

A four-panel set of stacked bar charts. Each chart shares a common Y-axis representing electricity demand in T W h from 0 to 1800 and an X-axis representing years from 2020 to 2035.

* Top-Left Panel: Lift-Off and A w B scenarios. The blue base representing Lift-Off Case no A I remains stable around 400 T W h. The orange section representing Abundance Without Boundaries A I grows exponentially, reaching a total of approximately 1700 T W h by 2035.

* Top-Right Panel: Headwinds and L t G scenarios. The blue base for Headwinds Case no A I declines significantly after 2027. The orange Limits To Growth A I section grows moderately, with the total demand plateauing around 700 T W h by 2035.

* Bottom-Left Panel: High Eff. and Sustainable A I scenarios. The blue High Efficiency Case no A I base drops to roughly 200 T W h by 2035. The orange Sustainable A I section grows steadily, bringing the total demand to just under 1000 T W h.

* Bottom-Right Panel: Base and Energy Crisis scenarios. The blue Base Case no A I grows steadily to 1000 T W h. The orange Energy Crisis A I section adds a smaller consistent layer, resulting in a total demand of approximately 1200 T W h by 2035.

6. The near-term carbon footprint: AI as a net contributor

Although energy consumption scenarios for data centers are based on global estimates, they can greatly vary by country. This leads to a diverse global landscape influenced by global trends, local conditions, policies, and decisions. In fact, geopolitical factors play a crucial role in shaping AI’s impact on the energy transition and carbon footprint in the short term. Regulatory frameworks across different countries can either accelerate or hinder AI advancements, influencing how quickly AI technologies are adopted in energy sectors. National security concerns drive investments in AI for defense and surveillance, which can divert resources from sustainable energy initiatives. International trade policies and technological alliances also affect the global distribution of AI technologies, impacting their deployment in renewable energy projects (World Economic Forum, 2024). Control over semiconductor supply chains, essential for AI hardware, can create geopolitical tensions, affecting the availability and cost of AI technologies. Additionally, countries leading in AI development, such as the United States and China, have significant influence over global AI policies and standards, shaping the direction of AI applications in energy efficiency and carbon reduction. These geopolitical dynamics, combined with societal factors like public perception and workforce adaptation, will determine AI’s role in achieving a sustainable energy transition and reducing carbon emissions (Iqbal et al., Reference Iqbal, Zhang and Jahangir2025).

Despite the growing emphasis on renewable energy, many regions still rely heavily on fossil fuels to power their electricity grids. This means that a significant portion of data centers’ energy consumption is indirectly contributing to GHG emissions. In regions with limited access to clean energy sources, the increasing electricity demands of data centers can exacerbate the reliance on fossil fuels, hindering efforts to decarbonize the energy sector (Maher et al., Reference Maher, Meinecke, Gromier, Garcia-Novelli and Fortmann2022).

As we probe the various scenarios in data center growth and the rise of AI workloads at these facilities, it will be important to deliberate on their near-term carbon footprint and its implications for achieving net-zero emissions. Depending on the energy mix used to generate electricity, the CO₂ emissions largely vary from region to region. In some locations, the data centers claim to be 100% renewable, but in many different places, it varies. Due to the lack of detailed CO₂ data for electricity generation in the three Shell scenarios, we based our calculations on the CO₂ emissions from the primary energy production (not only electricity). As the primary energy mix changes over time from fossil fuels to renewable energy, the associated CO₂ emissions are declining accordingly. As we see in Figure 7a, the rate of decline varies between scenarios. For the CO₂ estimates by data centers and AI workloads, the corresponding Shell scenarios were used as described in the previous section. For the Base Case scenario, the average energy mix between the Surge and Archipelagos scenarios was used, as they represent rather opposing cases. More precisely, the CO₂ emissions by data centers in the Headwinds Case and the High Efficiency Case reached a plateau by 2030 at around 130–150 Mt CO₂ per year. In the High Efficiency Case, the trend started pointing downwards by 2035. Finally, in the Base and Lift-Off Cases, the CO₂ emissions continued to increase past 2035, albeit at different rates, reaching 210 and 300 Mt per year, respectively (Figure 7a). In the case of AI, the CO₂ emissions from electricity usage in Sustainable AI and Limits to Growth plateaued around 2030 at around 115 Mt per year. On the contrary, the Abundance without Boundaries scenario grew at a much faster pace, reaching 240 Mt CO₂ by 2035 with no clear signs of slowing down. Finally, the Energy Crisis scenario peaked at 130 Mt CO₂ in 2030, followed by a rapid decline to 35 Mt within 5 years (Figure 7b). For context, 100 Mt CO₂ is equivalent to the emissions from nearly 13.5 million American homes’ energy usage for 1 year (EPA Calculator, n.d.).

Figure 7.

Historical trends and forecasts for CO₂ emissions from (a) data centers and (b) AI workloads through 2035. These forecasts are based on the data provided in the Shell plc (2025), Energy and AI (2025), and Paccou and Wijnhoven (Reference Paccou and Wijnhoven2024).

Figure 7. Long description

Two side-by-side line graphs.

Left graph: C O sub 2 emissions by data centers 2020-2035.

* X-axis: Years from 2020 to 2035.

* Y-axis: M t C O sub 2 forward slash year, ranging from 0 to 350.

* Data: All scenarios start at approximately 55 M t in 2020. The Lift-Off Case (orange) shows a steep sigmoid curve peaking near 300 M t. The Base Case (blue) rises to about 210 M t. The High Efficiency Case (dark green) peaks at 150 M t in 2032 before a slight decline. The Headwinds Case (purple) plateaus at 130 M t by 2030.

Right graph: C O sub 2 emissions from electricity usage by A I 2020-2035.

* X-axis: Years from 2020 to 2035.

* Y-axis: M t C O sub 2 forward slash year, ranging from 0 to 250.

* Data: Emissions are near zero until 2024. The Abundance Without Boundaries A I (green) shows an exponential-like increase reaching 240 M t. The Sustainable A I (purple) and Limits To Growth A I (orange) follow similar paths, peaking around 110-120 M t. The Energy Crisis A I (light blue) peaks at 130 M t in 2030 before sharply declining to 35 M t by 2035.

As illustrated in Figure 7, these scenarios present contrasting perspectives on the penetration of AI and its impact on CO₂ emissions. For example, the Lift-Off Case from IEA is more optimistic, projecting the highest CO₂ emissions, whereas the Energy Crisis scenario from Paccou and Wijnhoven (Reference Paccou and Wijnhoven2024) is pessimistic, predicting the least emissions. Although these reports present extreme scenarios, we believe the reality lies somewhere in between. It is often preferable to provide an interval within which the forecasted CO₂ emissions are expected to fall. To determine this interval, we consider the mean scenario presented in Figure 7 and calculate the P5, P50, and P95 probability cases around that, which will be described next.

6.1. Mean scenario and probabilistic assessment

To generate the potential range of CO₂ emissions under various assumptions and scenarios and to understand the variability and uncertainty in future emissions projections, a comprehensive methodology has been developed. This methodology includes a list of factors that influence emissions in data centers and the relationship between these factors and emissions. Supplementary Table 2 describes the factors that we considered to influence the emissions at the data centers. It is important to note that the mean scenario calculated from Figure 7 is captured in the last column, and the values for other factors were taken from various reports and studies carried out that represent the actual numbers available until 2024 and the forecasts until 2035.

The energy mix factor and AI workload at the data centers used for Figure 7 have also been reported here. The mean and standard deviation for each variable are used to define the 99% confidence interval, which provides the probabilistic range that the MC simulations can use to generate many such possibilities and then calculate the emissions for each possibility. Our main hypothesis is that the extreme scenarios depicted in Figure 7 would be highly unlikely to occur. An important point to note here is that these years include two pandemic years, where there is no significant deviation in the variation for these factors. In order to make predictions for each possibility, a mathematical model was fitted to the data presented in the Supplementary Table 2 with Columns 2–5 as inputs and Column 6 as output. Historical data is integrated into the model to validate its accuracy and ensure that the projections are grounded in real-world observations. For example, the mean scenario curve perfectly honors the historical and forecast data presented in Supplementary Table 2. Different scenarios (P5, P50, and P95) represent varying levels of probability, with P5 being the pessimistic one, with 5% of the time the outcome will fall below this value, P50 the median, and P95 the optimistic one, with 95% of the time that the outcome will fall below this value. Key assumptions made in the methodology include the stability of certain factors over time and the exclusion of extreme events. These assumptions, along with potential limitations and uncertainties, are acknowledged to provide a balanced view of the projections. The impact of AI workloads on CO₂ emissions is specifically accounted for in the model, recognizing that AI-related activities can significantly influence energy consumption. The energy mix factor, which defines the ratio of fossil versus renewable energy used, is a critical component of the model, as it directly affects the emissions calculations. Below is a summary of the methodology and steps involved to compute the P5, P50, and P95 probability scenarios:

1. 1. Calculate the mean CO₂ emission scenario from all the scenarios presented in Figure 7.

2. 2. Identify factors potentially impacting CO₂ emissions at data centers and gather past and future forecasts from published reports and literature from 2020 to 2035.

3. 3. Develop and validate a model honoring the data from Step 2.

4. 4. Determine the mean and standard deviation of each variable and the corresponding 99% confidence interval for data variation.

5. 5. Perform MC simulations using combinations of parameters sampled from the confidence intervals and apply the approximated curve to find the corresponding emissions for each realization.

6. 6. Calculate the P5, P50, and P95 probability cases based on the predictions.

Figure 8 illustrates the P5, P50, and P95 probability scenarios calculated using the methodology described above. As it can be clearly depicted, the P5, P50, and P95 cases showcase a healthy deviation from the mean case; however, the deviation is not as extreme as predicted in either the IEA or Paccou et al. Our modeling does not prioritize an energy crisis scenario as the most probable outcome, though such risks cannot be ruled out. We expect a deviation from the mean scenario, but not an extreme one. The reason is that most of the emissions are expected to come from data centers, and the data center growth does not support such an extreme scenario. On the other hand, we do not expect an energy crisis scenario, as we have not seen such a one in the last 20 years (including the COVID-19 pandemic). In summary, while AI applications have clear benefits for environmental initiatives, it is challenging to translate these insights into actual carbon reduction efforts due to the time it takes for AI to have an impact on carbon reduction. Despite technical advances such as model optimizations (e.g., low-rank approximations, smaller and more efficient models), green AI hardware, and the use of renewable energy sources to power data centers, we anticipate that AI will have a net positive impact on carbon emissions in the near future. It will take several years, possibly beyond 2035, for AI to start having a net negative impact on carbon emissions due to technical advancements and penetration, which will be discussed in the next section.

Figure 8.

The P5, P50, and P95 probability scenarios were generated based on the following assumptions: (a) The input variables listed in Supplementary Table 2 were assumed to follow a normal distribution. The ranges for these variables were selected within a 99% confidence interval around the mean scenario. (b) The scenarios presented in Paccou and Wijnhoven (Reference Paccou and Wijnhoven2024) were reproduced without considering the energy crisis scenario.

Figure 8. Long description

Two panels, (a) and (b), show Predicted C O sub 2 Emissions in M t per year on the Y-axis against Time in years on the X-axis from 2018 to 2036.

Both graphs share a common legend including:

- Low-Case Scenario P 5

- Mid-Case Scenario P 50

- High-Case Scenario P 95

- Observed C O sub 2 Emissions (black triangles)

- Abundance Without Boundaries A I (light blue dashed line)

- Limits To Growth A I (green dashed line)

- Sustainable A I (dark blue dashed line)

- Energy Crisis A I (red dashed line)

- Mean Scenario Data (black squares)

- Mean Scenario Model (yellow line)

In panel (a), all scenarios show a slow increase until 2024, followed by a steep rise. By 2030, the scenarios diverge. The High-Case P 95 and Mean Scenario plateau around 125 to 150 M t per year. The Energy Crisis A I scenario peaks at 2030 and then shows a sharp decline toward 40 M t per year by 2035. The Abundance Without Boundaries A I continues a steep linear increase, reaching nearly 250 M t per year by 2035.

In panel (b), the trends are similar but show higher variance in the upper-bound scenarios. The High-Case P 95 and Abundance Without Boundaries A I lines are closely grouped, both trending toward 240 M t per year by 2035. The Mid-Case P 50 scenario settles higher than in panel (a), reaching approximately 160 M t per year. The Energy Crisis A I scenario again shows a significant drop-off after 2030.

7. The long-term promise: AI as a net-zero enabler

8. Mitigating AI’s energy impact: Strategies and solutions

Research in green AI hardware can transform the sustainability of AI by prioritizing energy efficiency and reducing environmental impact. Recent advancements include energy-efficient GPUs, TPUs, and neuromorphic chips, which consume significantly less power compared to traditional hardware (Clemm et al., Reference Clemm, Stobbe and WimalawarneKand Druschke2024; Tabbakh et al., Reference Tabbakh, Amin, Islam, Mahmud, Chowdhury and Mukta2024). Advances in hardware design are now targeting the creation of AI accelerators that combine high performance with environmental sustainability. This effort includes developing processors that can run AI algorithms with very low energy consumption (Rahmani et al., Reference Rahmani, Shetty, Wagih, Ghasempour, Palazzi, Carvalho, Correia, Costanzo, Vital, Alimenti, Kettle, Masotti, Mezzanotte, Roselli and Grosinger2023). Ongoing research aims to create AI devices capable of harvesting energy from their surroundings to power AI systems (Bolón-Canedo et al., Reference Bolón-Canedo, Morán-Fernández, Cancela and Alonso-Betanzos2024). Examples include utilizing ambient light, vibrations, and heat. By incorporating energy-harvesting features, these AI devices could become more self-sufficient, reducing reliance on external power sources and promoting a more sustainable deployment of AI (Divya et al., Reference Divya, Panda, Hajra, Jeyaraj, Paul, Park, Kim and Oh2023; Ali et al., Reference Ali, Shaukat, Bibi, Altabey, Noori and Kouritem2023). Neuromorphic computing represents a significant shift in computing technology, inspired by the architecture of the human brain to create more efficient and adaptive systems. The human brain’s remarkable energy efficiency allows it to perform complex tasks with far less energy than conventional computers. By emulating this efficiency, neuromorphic computing aims to reduce the energy consumption of computational processes to levels similar to those of biological brains. This is achieved through the use of spiking neural networks (SNNs), which mimic the behavior of biological neurons, leading to substantial energy savings and performance enhancements, especially in AI applications that require real-time processing and decision-making. Additionally, a sustainability proposal by Oh et al. (Reference Oh, Kim, Kim, Kim, Park and Lee2023) suggests creating biodegradable neural network hardware to mitigate the impact of technological waste. Powering data centers with renewable energy sources is a crucial step toward reducing the carbon footprint of the rapidly growing digital infrastructure. Integrating renewable energy, such as solar and wind power, into data center operations can help mitigate the environmental impact of these energy-intensive facilities (Brown and Gorham, Reference Brown and Gorham2014). However, the variability of renewable energy sources poses challenges, which can be addressed through innovative solutions like virtual batteries that shift computational demand to match power availability (Agarwal et al., Reference Agarwal, Sun, Noghabi, Iyengar, Badam, Chandra, Seshan and Kalyanaraman2021). Additionally, advancements in energy storage technologies and smart grid integration are essential for ensuring a stable and reliable power supply (Hussain, Reference Hussain2024). These efforts not only support sustainability goals but also align with the increasing corporate commitments to carbon neutrality (Agarwal et al., Reference Agarwal, Sun, Noghabi, Iyengar, Badam, Chandra, Seshan and Kalyanaraman2021). CCS is a versatile technology and can be an enabler for data center decarbonization, with the potential to capture 90–95+ % of a power generation facility’s CO₂ emissions (Barlow and Kearns, Reference Barlow and Kearns2025). A business model is emerging that applies CCS to purpose-built power generation facilities “behind the meter” (not connected to municipal electric grids) to minimize data center carbon emissions while meeting their firm, high-reliability power demand. In this model, CO₂ emissions from natural gas turbines are captured, transported, and permanently stored before they reach the atmosphere—significantly decarbonizing the data center power supply. Industry is responding to the data center decarbonization opportunity in the United States and other countries. Further details can be found in the Role of CCS in US Data Centre Decarbonisation (2025).

Recent advancements in deep learning have focused on reducing the size of neural networks to enhance computational efficiency and enable deployment in resource-constrained environments. One approach involves a pruning algorithm that retrains the smaller, pruned model at its faster, initial learning rate, resulting in models that are simpler and more accurate (Renda et al., Reference Renda, Frankle and Carbin2020). Another comprehensive review highlights techniques such as pruning, quantization, low-rank factorization, knowledge distillation, and transfer learning, all of which effectively reduce model size while retaining performance. A study on global weight compression demonstrates that deploying deep neural networks to devices requiring real-time processing can save memory, reduce storage size, and lower computational requirements (Dantas et al., Reference Dantas, Sabino da Silva, Cordeiro and Carvalho2024). An example advancement is the PocketNet paradigm, which aims to reduce the size of deep learning models used in medical image analysis. The authors propose a modification to existing convolutional neural network architectures that limits the growth of the number of channels, thereby significantly reducing the number of parameters. This approach maintains comparable performance to conventional neural networks while using up to 90% less GPU memory and speeding up training times by up to 40%. The paper demonstrates the effectiveness of PocketNet in various segmentation and classification tasks, highlighting its potential for deployment in resource-constrained settings (Celaya et al., Reference Celaya, Actor, Muthusivarajan, Gates, Chung, Schellingerhout, Riviere and Fuentes2023). DeepSeek, a Chinese AI model, has garnered significant attention for its energy-efficient approach to AI. Unlike traditional models, DeepSeek employs a “mixture of experts” technique during training, where only a portion of the model’s parameters are activated at any given time. This method, combined with improved reinforcement learning techniques, has led to more efficient training processes. However, while DeepSeek’s training phase is notably energy-efficient, its inference phase tends to be more energy-intensive. The model generates longer responses compared to similar models, which offsets some of the energy savings achieved during training (O’Donnell, Reference O’Donnell2025). Despite its energy-efficient training, DeepSeek’s overall impact on energy consumption remains complex. The model’s longer inference responses require substantial computational power, which can lead to higher energy usage in data centers. This phenomenon is consistent with the broader trend in AI development, where efficiency gains in training often result in increased overall energy consumption due to the expanded capabilities and applications of the models (O’Donnell, Reference O’Donnell2025). As companies strive to develop more intelligent systems, they tend to invest more in training, thereby using more energy. This dynamic underscores the importance of considering both training and inference phases when evaluating the energy efficiency of AI models. The introduction of DeepSeek has significant implications for the tech industry and energy sector. While its energy-efficient training methods offer a promising solution to the growing energy demands of AI, the model’s inference phase highlights the need for balanced approaches to AI development. The rapid adoption of DeepSeek-like models could strain power grids and challenge energy security. To ensure AI remains an environmentally friendly tool, several strategies are crucial: These include focusing on green AI hardware and powering data centers with renewable energy sources to reduce the environmental impact of AI operations. Research into model optimization techniques can also lead to more energy-efficient models, as discussed above. Furthermore, standardized environmental impact assessments and collaboration between AI developers and sustainability experts can promote transparency and responsible development throughout the AI lifecycle.

9. Scope and limitations

While this paper brings together diverse and credible data sources to project AI’s energy trajectory, several conceptual and methodological limitations must be clarified to appropriately interpret our conclusions regarding AI’s long-term emissions footprint and its capability as a decarbonization enabler:

• • Analytical Separation of Footprint and Enablement: This study addresses two mathematically distinct phenomena: (1) the quantitative environmental footprint of AI infrastructure, including data center electricity demand, and (2) the potential emissions reductions enabled by AI applications across other economic sectors. Because our models are not dynamically coupled into a unified macroeconomic framework, we caution against synthesizing the footprint and enablement narratives into a singular “net impact.” Our quantitative risk evaluation is fundamentally isolated from our discussion on AI’s qualitative enabling potential. To state definitively whether the benefits mathematically offset the costs remains beyond our current methodology.

• • Qualitative Assessment of Reductions: The potential benefits of AI across sectors (e.g., energy systems optimization, transport, and Carbon Capture, utilization and Storage (CCUS)) detailed in this paper are largely drawn from qualitative surveys of existing literature. We do not attempt to deterministically quantify the magnitude or probability of these reductions within our MC simulations. For example, applying a simple breakeven calculation to our P50 projections implies AI would need to catalyze systemic global emissions reductions on a massive scale merely to neutralize its own operational footprint.

• • Reliance on Scenario Synthesis: Our MC probabilistic simulations layer mathematical uncertainty analysis over base scenarios derived from previously published studies, particularly the IEA Energy and AI report (Paccou and Wijnhoven, Reference Paccou and Wijnhoven2024). Consequently, our primary analytical contribution entails synthesizing, reframing, and probabilistically modeling the risk profiles of these external reference scenarios, rather than generating fundamentally new foundational empirical benchmarks.

• • The Rebound Effect Remains Unmodeled: While we address Jevons’ paradox conceptually, dynamic rebound effects have not been mathematically integrated into our modeling architecture. Because improved computational efficiency conventionally induces disproportionately higher AI consumption, ignoring downstream rebound dynamics implies our infrastructure emission trajectories likely function as conservative, lower-bound estimates. Effectively analyzing Jevons’ paradox would necessitate computable general equilibrium (CGE) modeling, constituting an important frontier for future research.