Electricity demand is expected to increase significantly, propelled by ongoing electrification and the rapid growth of artificial intelligence (AI) workloads. However, this demand surge may outstrip the capacity of grid infrastructure. In the United States, constraints include interconnection queue backlogs, regional transmission bottlenecks, and supply-chain delays for critical components. At the same time, there is a counter-narrative: AI efficiency improvements, through model compression, data-centric design, and inference optimization, could meaningfully temper future demand. This article examines these intersecting tensions, drawing on recent data from the International Energy Agency (IEA) and the U.S. Department of Energy’s Grid Deployment Office (DOE GDO), among other sources.

The IEA’s Building the Future Transmission Grid analysis reports that supply-chain bottlenecks are putting increasing pressure on global transmission deployment. Lead times for essential components have lengthened: for example, procuring cables now takes 2–3 years, while large power transformers can require up to 4 years, and high-voltage DC cables may face wait times exceeding 5 years [1]. Meanwhile, prices have surged: cable costs have nearly doubled since 2019, and transformer prices have risen approximately 75% [1]. These trends contribute to project delays and escalating costs, undermining the pace of grid expansion.

The IEA emphasizes that meeting future transmission needs will require not only increased capital investment but also more strategic procurement, transparent planning, and streamlined permitting [1].

A major bottleneck in U.S. grid expansion lies in interconnection. As of the end of 2023, nearly 2,600 GW of generating and storage capacity was actively seeking transmission connection, more than twice the size of the current U.S. generating fleet [2].

Over 95% of this queued capacity comes from clean-energy sources: solar, wind, and battery storage [2]. Delays are significant. Data shows that the median time from interconnection request to commercial operation for recent projects has stretched to five years, compared with under two years for earlier vintages [3]. Furthermore, project attrition is high: around 80% of interconnection requests are eventually withdrawn before completion [4].

These dynamics reflect systemic inefficiencies in the interconnection regime: speculative applications, protracted impact studies, and ambiguous cost-allocation. Such challenges constrain the ability of clean-energy projects to proceed reliably.

The Midcontinent Independent System Operator (MISO) continues to face structural transmission constraints, particularly along its seam with the Southwest Power Pool (SPP), leading both organizations to develop the Joint Targeted Interconnection Queue (JTIQ) portfolio as a coordinated solution. According to the JTIQ fact sheet, the study identified five transmission projects costing approximately $1.6 billion, designed to unlock about 28.6 GW of new generation capacity across the seam by replacing piecemeal upgrades with a coordinated backbone approach [5].

The JTIQ portfolio was approved by the SPP Board on 9 December 2024 and by the MISO Board on 12 December 2024 as part of the broader MISO Transmission Expansion Plan for 2024 (MTEP24), which represents one of the most extensive grid-investment packages in MISO’s history.

MISO reports that MTEP24 consists of 488 approved transmission projects across its fifteen-state footprint, covering more than 5,000 miles and totaling roughly $30 billion in anticipated investment when accounting for local, regional, and joint upgrades [6]. Within this portfolio, 459 local reliability projects extend across 932 miles at a cost of $6.7 billion, while the Long-Range Transmission Planning (LRTP) Tranche 2.1 contains 24 major regional projects spanning 3,631 miles, including new 765 kV backbone lines with a combined capital cost of $21.8 billion and an expected long-term benefit exceeding $72 billion [6].

MISO emphasizes that this integrated suite of projects, developed through more than 300 stakeholder meetings, reflects an effort to strengthen reliability, enhance transfer capability, and accommodate a rapidly evolving generation mix. Although the JTIQ projects reduce interconnection barriers and improve coordination with SPP, both sources note that sustained transmission expansion remains essential given continued queue growth, rising load, and the operational demands of resource diversification.

The Electric Reliability Council of Texas (ERCOT) operates as a largely isolated grid, with only three direct current (DC) ties connecting it to neighboring interconnections. This limited connectivity constrains ERCOT’s ability to import or export power during peak periods or emergency conditions, creating inherent reliability vulnerabilities [7]. The 2023 DOE Grid Deployment Office Needs Study emphasizes that substantial increases in interregional transfer capacity will be necessary to support moderate-to-high clean energy growth scenarios. Without such enhancements, the integration of large-scale wind and solar generation may be impeded, reducing overall system efficiency [7]. ERCOT’s internal transmission network also experiences localized congestion, particularly in regions with high wind generation in West Texas and the Panhandle. These bottlenecks restrict the optimal utilization of renewable energy, often necessitating curtailment during periods of excess generation. The limited interconnections further reduce operational flexibility, requiring ERCOT to maintain higher levels of internal spinning reserves or invest in energy storage to balance variable renewable resources [7].

The DOE report underscores the consequences of ERCOT’s isolation during extreme events. For instance, simultaneous generation and load stress can significantly increase the risk of reliability failures, as observed during the 2021 Texas Winter Storm. The study suggests that enhancing interregional transfer capacity, alongside targeted internal transmission upgrades and energy storage deployment, is critical for mitigating such risks [7]. To support future clean energy deployment, the DOE recommends that ERCOT pursue additional interconnections with neighboring regions, implement strategic transmission expansions to relieve internal congestion, and adopt advanced grid planning and forecasting tools. These measures would not only increase operational flexibility but also enable higher utilization of renewable resources while maintaining system reliability. Quantitative projections within the study indicate that interregional capacity increases on the order of multiple gigawatts may be necessary under high renewable adoption scenarios to prevent curtailments and ensure adequate reserve sharing [7].

Overall, ERCOT’s current configuration presents both challenges and opportunities: while its isolation limits flexibility and the ability to import power during stress events, strategic interconnection and transmission enhancements can significantly strengthen the grid, facilitating the integration of clean energy resources and supporting long-term reliability objectives [7].

The protracted interconnection process creates considerable risk for Power Purchase Agreements (PPAs). Developers often apply speculatively, but many withdraw when studies reveal high costs or long delivery times [4]. This pattern undermines contract certainty: only a subset of queued capacity ever materializes. A study by Berkeley Lab found that from 2000–2018, only 19% of projects, and just 14% of queued capacity, ultimately reached commercial operation [3]. This high “drop-out” rate complicates long-term planning for off­-takers, financiers, and utilities.

To address speculative queue behavior, FERC Order 2023 introduced a cluster-study process alongside stricter readiness requirements. These reforms aim to reduce immature applications and streamline studies. At the same time, the DOE GDO identified key transmission corridors for potential designation as National Interest Electric Transmission Corridors (NIETCs) to relieve interregional congestion [7]. The 2023 guidance specifies expected large capacity increases across several seams by 2035, including 414% median growth in transfer capacity between the Delta and Plains regions [7]. However, implementing these reforms is challenged by long lead times for equipment and unclear funding and cost-allocation mechanisms.

While many grid-planning models assume steep demand growth from AI and data centers, emerging research suggests this growth may be less aggressive than commonly projected.

Research in Green AI demonstrates how optimizing models can sharply reduce energy consumption. In one study, modifying the dataset during training (a data-centric approach) reduced training energy by up to 92% with little to no performance loss [8]. These results indicate that smarter training strategies, rather than brute-force scaling, can dramatically lower the carbon footprint of AI. Another preprint advocates selecting smaller, task-appropriate models. The authors estimate that global AI energy consumption could be reduced by 27.8%, potentially saving tens of terawatt-hours annually [9]. Though preliminary, such work underscores how model choice alone may decelerate energy demand growth.

At inference (when models respond to live queries), bottom-up analyses estimate the median energy per query on high-end hardware (e.g., H100 GPU) at 0.34 Wh, far lower than some prior public estimates [10]. The same study projects that combining optimizations at the model, platform, and hardware level could deliver 8–20× reductions in per-query energy. These efficiency improvements suggest that AI workload growth may not require commensurate scaling of electricity demand, challenging assumptions of unchecked load growth.

If these efficiency trends scale broadly, grid planners may be overestimating demand in their long-term scenarios. Rather than a simple “exponential demand → build more grid” narrative, a more nuanced trajectory might emerge:

• Infrastructure investments based on aggressive demand may overbuild.

• PPAs negotiated under optimistic assumptions risk over-commitment.

• Planning frameworks may over prioritize supply-side expansion at the expense of demand-side flexibility.

Several strategic risks arise if planners base infrastructure decisions on overly aggressive demand forecasts:

1. Stranded Transmission Assets: New lines, substations, or transformers built under assumptions of runaway demand could be underutilized, weakening their economic justification.

2. Financial Exposure: Developers, off-takers, and utilities could lock in PPAs that do not align with actual realized demand, exposing them to risk.

3. Procurement Mismatch: Manufacturers may over invest in production capacity for cables and transformers only to face lower orders if demand growth slows.

4. Regulatory Misalignment: Policies and incentives designed for high-growth scenarios may misallocate resources if demand moderates due to efficiency.

5. Operational Inefficiencies: Excess capacity may lead to low utilization, depressing returns and discouraging future build-out.

In light of these risks, a balanced, forward-looking strategy is needed:

1. Scenario Planning: Incorporate demand-moderated scenarios that reflect possible AI efficiency gains. Use sensitivity analyses that account for lower-than-expected load growth.

2. Proactive Procurement: Negotiate long-term supply contracts for critical components (cables, transformers) linked to realistic demand paths.

3. Demand-Side Flexibility: Encourage shifting of AI workloads (e.g., non-urgent inference) to times or regions with abundant renewable generation.

4. Regulatory Acceleration: Implement reforms like FERC Order 2023 and NIETC corridor designations; align cost-recovery mechanisms with updated demand scenarios.

5. Incentivize Efficiency: Support “green AI” research, model compression practices, and hardware-performance matching via grants, procurement standards, or recognition programs.

6. Monitoring & Feedback: Establish regular reviews of demand, actual grid usage, and interconnection pipeline data to adjust infrastructure plans in real time.

The energy grid faces a profound paradox. On the one hand, transmission and interconnection bottlenecks threaten to constrain growth just as demand from AI and electrification accelerates. On the other hand, rapid advances in AI efficiency challenge the assumption that demand will continue to grow unchecked. If efficiency gains are realized at scale, the projected surge in electricity consumption may not materialize as anticipated. This duality underscores the importance of flexible, data-driven planning. Rather than betting solely on supply expansion, stakeholders must also invest in demand-side innovation and robust scenario analysis. By aligning infrastructure build-out with realistic demand forecasts, and incentivizing AI energy efficiency, policy makers, utilities, and developers can better navigate the tradeoffs of a rapidly evolving energy future.

1. International Energy Agency (IEA) (2025) | Building the Future Transmission Grid: Strategies to navigate supply-chain challenges (Executive Summary)
   https://www.iea.org/reports/building-the-future-transmission-grid/executive-summary

1. Lawrence Berkeley National Laboratory (LBNL) (2024) | Grid connection backlog grows by 30% in 2023, dominated by requests for solar, wind, and energy storage
   https://emp.lbl.gov/news/grid-connection-backlog-grows-30-2023-dominated-requests-solar-wind-and-energy-storage

1. Lawrence Berkeley National Laboratory (LBNL) (2024) | Queued Up: Characteristics of Power Plants Seeking Transmission Interconnection as of the End of 2023
   https://emp.lbl.gov/queued-characteristics-power-plants-seeking-transmission-interconnection-end-2023

1. Lawrence Berkeley National Laboratory (LBNL) (2025) | Grid Connection Barriers to New-Build Power Plants in the United States
   https://emp.lbl.gov/news/grid-connection-barriers-new-build-power-plants-united-states

1. MISO & Southwest Power Pool (2024) | Joint Targeted Interconnection Queue (JTIQ) Fact Sheet
   https://cdn.misoenergy.org/JTIQ%20Fact%20Sheet%20Website666572.pdf

1. MISO (2024) | MISO Board Approves Historic Transmission Plan to Strengthen Grid Reliability (MTEP24)
   https://www.misoenergy.org/meet-miso/media-center/2024/miso-board-approves-historic-transmission-plan-to-strengthen-grid-reliability/

1. U.S. Department of Energy, Grid Deployment Office (DOE GDO) (2023) | 2023 Needs Study: NIETC Final Guidance Document
   https://www.energy.gov/sites/default/files/2023-12/2023-12-15%20GDO%20NIETC%20Final%20Guidance%20Document.pdf

1. Verdecchia, R., Cruz, L., Sallou, J., Lin, M., Wickenden, J., & Hotellier, E. (2022) | Data-Centric Green AI: An Exploratory Empirical Study
   https://arxiv.org/abs/2204.02766

1. Barros, T. da S., Giroire, F., Aparicio-Pardo, R., & Moulierac, J. (2025) | Small is Sufficient: Reducing the World AI Energy Consumption Through Model Selection
   https://arxiv.org/abs/2510.01889

1. Oviedo, F., Kazhamiaka, F., Choukse, E., Kim, A., Luers, A., Nakagawa, M., Bianchini, R., & Lavista Ferres, J. M. (2025) | Energy Use of AI Inference: Efficiency Pathways and Test-Time Compute
   https://arxiv.org/abs/2509.20241

Electricity demand is expected to increase significantly, propelled by ongoing electrification and the rapid growth of artificial intelligence (AI) workloads. However, this demand surge may outstrip the capacity of grid infrastructure. In the United States, constraints include interconnection queue backlogs, regional transmission bottlenecks, and supply-chain delays for critical components. At the same time, there is a counter-narrative: AI efficiency improvements, through model compression, data-centric design, and inference optimization, could meaningfully temper future demand. This article examines these intersecting tensions, drawing on recent data from the International Energy Agency (IEA) and the U.S. Department of Energy’s Grid Deployment Office (DOE GDO), among other sources.

The IEA’s Building the Future Transmission Grid analysis reports that supply-chain bottlenecks are putting increasing pressure on global transmission deployment. Lead times for essential components have lengthened: for example, procuring cables now takes 2–3 years, while large power transformers can require up to 4 years, and high-voltage DC cables may face wait times exceeding 5 years [1]. Meanwhile, prices have surged: cable costs have nearly doubled since 2019, and transformer prices have risen approximately 75% [1]. These trends contribute to project delays and escalating costs, undermining the pace of grid expansion.

The IEA emphasizes that meeting future transmission needs will require not only increased capital investment but also more strategic procurement, transparent planning, and streamlined permitting [1].

A major bottleneck in U.S. grid expansion lies in interconnection. As of the end of 2023, nearly 2,600 GW of generating and storage capacity was actively seeking transmission connection, more than twice the size of the current U.S. generating fleet [2].

Over 95% of this queued capacity comes from clean-energy sources: solar, wind, and battery storage [2]. Delays are significant. Data shows that the median time from interconnection request to commercial operation for recent projects has stretched to five years, compared with under two years for earlier vintages [3]. Furthermore, project attrition is high: around 80% of interconnection requests are eventually withdrawn before completion [4].

These dynamics reflect systemic inefficiencies in the interconnection regime: speculative applications, protracted impact studies, and ambiguous cost-allocation. Such challenges constrain the ability of clean-energy projects to proceed reliably.

The Midcontinent Independent System Operator (MISO) continues to face structural transmission constraints, particularly along its seam with the Southwest Power Pool (SPP), leading both organizations to develop the Joint Targeted Interconnection Queue (JTIQ) portfolio as a coordinated solution. According to the JTIQ fact sheet, the study identified five transmission projects costing approximately $1.6 billion, designed to unlock about 28.6 GW of new generation capacity across the seam by replacing piecemeal upgrades with a coordinated backbone approach [5].

The JTIQ portfolio was approved by the SPP Board on 9 December 2024 and by the MISO Board on 12 December 2024 as part of the broader MISO Transmission Expansion Plan for 2024 (MTEP24), which represents one of the most extensive grid-investment packages in MISO’s history.

MISO reports that MTEP24 consists of 488 approved transmission projects across its fifteen-state footprint, covering more than 5,000 miles and totaling roughly $30 billion in anticipated investment when accounting for local, regional, and joint upgrades [6]. Within this portfolio, 459 local reliability projects extend across 932 miles at a cost of $6.7 billion, while the Long-Range Transmission Planning (LRTP) Tranche 2.1 contains 24 major regional projects spanning 3,631 miles, including new 765 kV backbone lines with a combined capital cost of $21.8 billion and an expected long-term benefit exceeding $72 billion [6].

MISO emphasizes that this integrated suite of projects, developed through more than 300 stakeholder meetings, reflects an effort to strengthen reliability, enhance transfer capability, and accommodate a rapidly evolving generation mix. Although the JTIQ projects reduce interconnection barriers and improve coordination with SPP, both sources note that sustained transmission expansion remains essential given continued queue growth, rising load, and the operational demands of resource diversification.

The Electric Reliability Council of Texas (ERCOT) operates as a largely isolated grid, with only three direct current (DC) ties connecting it to neighboring interconnections. This limited connectivity constrains ERCOT’s ability to import or export power during peak periods or emergency conditions, creating inherent reliability vulnerabilities [7]. The 2023 DOE Grid Deployment Office Needs Study emphasizes that substantial increases in interregional transfer capacity will be necessary to support moderate-to-high clean energy growth scenarios. Without such enhancements, the integration of large-scale wind and solar generation may be impeded, reducing overall system efficiency [7]. ERCOT’s internal transmission network also experiences localized congestion, particularly in regions with high wind generation in West Texas and the Panhandle. These bottlenecks restrict the optimal utilization of renewable energy, often necessitating curtailment during periods of excess generation. The limited interconnections further reduce operational flexibility, requiring ERCOT to maintain higher levels of internal spinning reserves or invest in energy storage to balance variable renewable resources [7].

The DOE report underscores the consequences of ERCOT’s isolation during extreme events. For instance, simultaneous generation and load stress can significantly increase the risk of reliability failures, as observed during the 2021 Texas Winter Storm. The study suggests that enhancing interregional transfer capacity, alongside targeted internal transmission upgrades and energy storage deployment, is critical for mitigating such risks [7]. To support future clean energy deployment, the DOE recommends that ERCOT pursue additional interconnections with neighboring regions, implement strategic transmission expansions to relieve internal congestion, and adopt advanced grid planning and forecasting tools. These measures would not only increase operational flexibility but also enable higher utilization of renewable resources while maintaining system reliability. Quantitative projections within the study indicate that interregional capacity increases on the order of multiple gigawatts may be necessary under high renewable adoption scenarios to prevent curtailments and ensure adequate reserve sharing [7].

Overall, ERCOT’s current configuration presents both challenges and opportunities: while its isolation limits flexibility and the ability to import power during stress events, strategic interconnection and transmission enhancements can significantly strengthen the grid, facilitating the integration of clean energy resources and supporting long-term reliability objectives [7].

The protracted interconnection process creates considerable risk for Power Purchase Agreements (PPAs). Developers often apply speculatively, but many withdraw when studies reveal high costs or long delivery times [4]. This pattern undermines contract certainty: only a subset of queued capacity ever materializes. A study by Berkeley Lab found that from 2000–2018, only 19% of projects, and just 14% of queued capacity, ultimately reached commercial operation [3]. This high “drop-out” rate complicates long-term planning for off­-takers, financiers, and utilities.

To address speculative queue behavior, FERC Order 2023 introduced a cluster-study process alongside stricter readiness requirements. These reforms aim to reduce immature applications and streamline studies. At the same time, the DOE GDO identified key transmission corridors for potential designation as National Interest Electric Transmission Corridors (NIETCs) to relieve interregional congestion [7]. The 2023 guidance specifies expected large capacity increases across several seams by 2035, including 414% median growth in transfer capacity between the Delta and Plains regions [7]. However, implementing these reforms is challenged by long lead times for equipment and unclear funding and cost-allocation mechanisms.

While many grid-planning models assume steep demand growth from AI and data centers, emerging research suggests this growth may be less aggressive than commonly projected.

Research in Green AI demonstrates how optimizing models can sharply reduce energy consumption. In one study, modifying the dataset during training (a data-centric approach) reduced training energy by up to 92% with little to no performance loss [8]. These results indicate that smarter training strategies, rather than brute-force scaling, can dramatically lower the carbon footprint of AI. Another preprint advocates selecting smaller, task-appropriate models. The authors estimate that global AI energy consumption could be reduced by 27.8%, potentially saving tens of terawatt-hours annually [9]. Though preliminary, such work underscores how model choice alone may decelerate energy demand growth.

At inference (when models respond to live queries), bottom-up analyses estimate the median energy per query on high-end hardware (e.g., H100 GPU) at 0.34 Wh, far lower than some prior public estimates [10]. The same study projects that combining optimizations at the model, platform, and hardware level could deliver 8–20× reductions in per-query energy. These efficiency improvements suggest that AI workload growth may not require commensurate scaling of electricity demand, challenging assumptions of unchecked load growth.

If these efficiency trends scale broadly, grid planners may be overestimating demand in their long-term scenarios. Rather than a simple “exponential demand → build more grid” narrative, a more nuanced trajectory might emerge:

• Infrastructure investments based on aggressive demand may overbuild.

• PPAs negotiated under optimistic assumptions risk over-commitment.

• Planning frameworks may over prioritize supply-side expansion at the expense of demand-side flexibility.

Several strategic risks arise if planners base infrastructure decisions on overly aggressive demand forecasts:

1. Stranded Transmission Assets: New lines, substations, or transformers built under assumptions of runaway demand could be underutilized, weakening their economic justification.

2. Financial Exposure: Developers, off-takers, and utilities could lock in PPAs that do not align with actual realized demand, exposing them to risk.

3. Procurement Mismatch: Manufacturers may over invest in production capacity for cables and transformers only to face lower orders if demand growth slows.

4. Regulatory Misalignment: Policies and incentives designed for high-growth scenarios may misallocate resources if demand moderates due to efficiency.

5. Operational Inefficiencies: Excess capacity may lead to low utilization, depressing returns and discouraging future build-out.

In light of these risks, a balanced, forward-looking strategy is needed:

1. Scenario Planning: Incorporate demand-moderated scenarios that reflect possible AI efficiency gains. Use sensitivity analyses that account for lower-than-expected load growth.

2. Proactive Procurement: Negotiate long-term supply contracts for critical components (cables, transformers) linked to realistic demand paths.

3. Demand-Side Flexibility: Encourage shifting of AI workloads (e.g., non-urgent inference) to times or regions with abundant renewable generation.

4. Regulatory Acceleration: Implement reforms like FERC Order 2023 and NIETC corridor designations; align cost-recovery mechanisms with updated demand scenarios.

5. Incentivize Efficiency: Support “green AI” research, model compression practices, and hardware-performance matching via grants, procurement standards, or recognition programs.

6. Monitoring & Feedback: Establish regular reviews of demand, actual grid usage, and interconnection pipeline data to adjust infrastructure plans in real time.

The energy grid faces a profound paradox. On the one hand, transmission and interconnection bottlenecks threaten to constrain growth just as demand from AI and electrification accelerates. On the other hand, rapid advances in AI efficiency challenge the assumption that demand will continue to grow unchecked. If efficiency gains are realized at scale, the projected surge in electricity consumption may not materialize as anticipated. This duality underscores the importance of flexible, data-driven planning. Rather than betting solely on supply expansion, stakeholders must also invest in demand-side innovation and robust scenario analysis. By aligning infrastructure build-out with realistic demand forecasts, and incentivizing AI energy efficiency, policy makers, utilities, and developers can better navigate the tradeoffs of a rapidly evolving energy future.

1. International Energy Agency (IEA) (2025) | Building the Future Transmission Grid: Strategies to navigate supply-chain challenges (Executive Summary)
   https://www.iea.org/reports/building-the-future-transmission-grid/executive-summary

1. Lawrence Berkeley National Laboratory (LBNL) (2024) | Grid connection backlog grows by 30% in 2023, dominated by requests for solar, wind, and energy storage
   https://emp.lbl.gov/news/grid-connection-backlog-grows-30-2023-dominated-requests-solar-wind-and-energy-storage

1. Lawrence Berkeley National Laboratory (LBNL) (2024) | Queued Up: Characteristics of Power Plants Seeking Transmission Interconnection as of the End of 2023
   https://emp.lbl.gov/queued-characteristics-power-plants-seeking-transmission-interconnection-end-2023

1. Lawrence Berkeley National Laboratory (LBNL) (2025) | Grid Connection Barriers to New-Build Power Plants in the United States
   https://emp.lbl.gov/news/grid-connection-barriers-new-build-power-plants-united-states

1. MISO & Southwest Power Pool (2024) | Joint Targeted Interconnection Queue (JTIQ) Fact Sheet
   https://cdn.misoenergy.org/JTIQ%20Fact%20Sheet%20Website666572.pdf

1. MISO (2024) | MISO Board Approves Historic Transmission Plan to Strengthen Grid Reliability (MTEP24)
   https://www.misoenergy.org/meet-miso/media-center/2024/miso-board-approves-historic-transmission-plan-to-strengthen-grid-reliability/

1. U.S. Department of Energy, Grid Deployment Office (DOE GDO) (2023) | 2023 Needs Study: NIETC Final Guidance Document
   https://www.energy.gov/sites/default/files/2023-12/2023-12-15%20GDO%20NIETC%20Final%20Guidance%20Document.pdf

1. Verdecchia, R., Cruz, L., Sallou, J., Lin, M., Wickenden, J., & Hotellier, E. (2022) | Data-Centric Green AI: An Exploratory Empirical Study
   https://arxiv.org/abs/2204.02766

1. Barros, T. da S., Giroire, F., Aparicio-Pardo, R., & Moulierac, J. (2025) | Small is Sufficient: Reducing the World AI Energy Consumption Through Model Selection
   https://arxiv.org/abs/2510.01889

1. Oviedo, F., Kazhamiaka, F., Choukse, E., Kim, A., Luers, A., Nakagawa, M., Bianchini, R., & Lavista Ferres, J. M. (2025) | Energy Use of AI Inference: Efficiency Pathways and Test-Time Compute
   https://arxiv.org/abs/2509.20241