4

Impact of Data Centers on the Grid

As grid operators work to anticipate and meet growing demands from artificial intelligence (AI) data centers, they face a variety of challenges in forecasting needs, expanding capacity, and maintaining reliability. Participants examined these issues in a panel discussion of the increasing impact of large data centers on the grid and in a fireside chat highlighting the on-the-ground experiences of utilities.

PANELIST REMARKS

Thomas Wilson, Electric Power Research Institute (EPRI), moderated a panel on the impact of data centers on the grid and introduced the speakers: Costa Samaras, Carnegie Mellon University; Line Roald, University of Wisconsin–Madison; T. Bruce Tsuchida, Brattle Group; and Ravi Jain, Tapestry.

Advancing Climate and Energy Goals with Ethical Artificial Intelligence

Despite the many unknowns surrounding the impacts of AI data centers on the electric power grid, Samaras posited that it is still possible to achieve net-zero emissions as well as energy affordability, reliability, and resilience. “Even if we don’t know everything, we can still take actions that are consistent with the outcomes that we want,” he said. Analyses suggest that AI workloads in practice are overestimated by looking at the capacity

of existing data centers, which is encouraging because it means there is perhaps more time to address energy issues than would be assumed based on the maximum thermal design power for the number of chips that are in use.¹ However, this only buys a little bit of time, and there remain many important issues to address.

One challenge is that AI workloads are far more variable and have much wider swings in energy use than traditional computing, which makes them computationally efficient but disruptive to the electric grid. There are various possible solutions that could help to smooth these volatile loads—by adding harmonics, addressing quality issues, building virtual power plants (VPPs),^2,3 or designing new algorithms. However, to spur progress in this space, Samaras said that it will be important to have more transparency into data centers’ energy use and assess how they affect resilience and sustainability.

Samaras outlined a set of principles for ethical AI that attend to climate and energy goals, proposing that AI data centers should be net-zero, avoid adding local pollution, and add more clean power to the grid than they use; share open data and rising efficiency targets; invest in new infrastructure that increases reliability, flexibility, and resilience; avoid impacts on consumer rates; and develop community benefits agreements.⁴ “I’m confident that if we center principles around people, equity, and emissions, we will be in a better place 10 years later after the broader adoption of artificial intelligence,” Samaras concluded.

The Role of Data Centers in Creating a Greener Grid

Roald discussed opportunities for AI data centers to help the grid operate more economically, efficiently, sustainably, and reliably through demand response and market participation. Roald presented an analysis of how data centers can adapt their energy needs to consume power when electricity is generated with low-carbon sources. To provide the necessary feedback, this strategy would require a new carbon-intensity signal indicating when the grid is “green.” There are different approaches to measure the “greenness” of electricity, including average carbon emissions,

___________________

¹ A.C. Newkirk, J. Fernandez, J. Koomey, et al., 2025, “Empirically-Calibrated H100 Node Power Models for Reducing Uncertainty in AI Training Energy Estimation,” Carnegie Mellon University, https://doi.org/10.48550/arXiv.2506.14551.

² Virtual power plants are systems that aggregate and coordinate energy production and distribution from large collections of small energy producers.

³ A.C. Newkirk, J. Fernandez, J. Koomey, et al., 2025, “Empirically-Calibrated H100 Node Power Models for Reducing Uncertainty in AI Training Energy Estimation,” Carnegie Mellon University, https://doi.org/10.48550/arXiv.2506.14551.

⁴ The White House, n.d., “Blueprint for an AI Bill of Rights,” https://bidenwhitehouse.archives.gov/ostp/ai-bill-of-rights, accessed February 15, 2025.

computed by dividing total emissions by total consumption at different timescales, or marginal carbon emissions, computed by dividing changes in emissions by changes in consumption. Both methods have benefits and drawbacks and, importantly, adapting power consumption according to either of these metrics could increase stress on the grid or emissions, underscoring the need for more research in this area.⁵

Perhaps a better opportunity for data centers to support a more sustainable and reliable grid, Roald suggested, is by implementing incentivized, improved computing flexibilities that enable them to participate in the electricity market. In this scenario, data centers would make bids based on real-time usage, allowing them access to cheaper (and typically greener) electricity. This flexibility saves money for the data centers, improves grid stability, and enables data centers to actively shape carbon and price signals themselves, Roald said.

Forecasting and Meeting Electric Demand

Tsuchida discussed opportunities and challenges in forecasting future demand for electric power and preparing the grid to meet future needs. Data centers are just one of many drivers of increasing electricity demand; other major drivers include onshoring and industrial electrification, cryptocurrency mining, and increasing electrification of buildings and vehicles. While there are different projections of future load, all predict significant growth.⁶

A key challenge is to determine how accurate these forecasts are and then price energy and build services accordingly. Overestimating future loads can lead to higher utility rates, while underestimating them has multiple negative consequences, including reduced reliability. Demand response is one strategy to keep costs lower, but it is challenging for utilities to price and serve new large loads when there is so much uncertainty. Tsuchida underscored the importance of explicitly considering options for expansion in all generation and transmission plans, including through optimizing the usage of existing infrastructure, such as through the adoption of grid-enhancing technologies,⁷ and incorporating the full effects of demand-side resources and non-wire alternatives.

___________________

⁵ J. Gorka, N. Rhodes, and L. Roald, 2025, “ElectricityEmissions.jl: A Framework for the Comparison of Carbon Intensity Signals,” arXiv:2411.06560, https://doi.org/10.48550/arXiv.2411.06560.

⁶ T.B. Tsuchida, L. Lam, P. Fox-Penner, et al., 2024, Electricity Demand Growth and Forecasting in a Time of Change, The Brattle Group, https://www.brattle.com/wp-content/uploads/2024/05/Electricity-Demand-Growth-and-Forecasting-in-a-Time-of-Change-1.pdf.

⁷ Grid-enhancing technologies include various hardware and software tools that help to alleviate congestion and control the flow of power through the electric grid.

A Holistic Approach to Grid Management, Planning, and Operations

In an electric power environment that is highly distributed, decentralized, variable, and rapidly growing, Jain said there are two main challenges to improving the grid: handling load growth and integrating clean energy. Given the scale of these challenges and their impact on the wider ecosystem, he highlighted the need for a holistic approach to grid management, planning, and operations. He posited that siloing among these elements and across energy markets has impeded the ability to effectively address these issues and described how AI, a key driver of current load growth, can also help abate it through new tools that leverage AI, data integration, and software innovation.

To this end, Google X created Tapestry, a new, secure grid platform that enables cross-silo collaboration. Jain said that this end-to-end product approaches grid planning, management, and operations holistically to better leverage physical infrastructure and enable more effective planning. It uses modernized, AI-powered grid operations tools to integrate data on load and supply variability at the needed scale, speed, and granularity.

As an example of how collaborative AI–energy partnerships can implement coordinated solutions at scale, Jain described how a Tapestry pilot is helping Chile meet its ambitious decarbonization goals. Preliminary results show dramatically reduced computing time, more accurate weather predictions (critical for wind and solar generation), and an automated yet collaborative workflow, reducing the planning and operations burdens of load growth and clean energy integration. In response to a question, Jain noted that while the team has not explicitly modeled flexibility in terms of data center loads, they plan to extend the modeling to capture collaboration between data center owners and utilities. He added that there is a great deal of flexibility in terms of the parameters that are set up to capture demand, generation, planning, and optimization for different scenarios at different time horizons and granularities.

PANEL DISCUSSION

In an open discussion, panelists considered how the growth of AI data centers may intersect with other pressures that grid operators face, discussed the economics of energy and data center operations, and identified possible policy solutions that could help to advance progress and forestall some of the anticipated challenges.

Anticipating Growth

Panelists discussed different ways in which the expansion of AI data centers could impact the electric grid at national and local scales. Wilson

asked panelists to comment on opportunities to address the impact of AI data centers on energy reliability, in light of the wide spikes and dips the grid can experience as AI processes cycle on and off. Samaras replied that the integration of data centers into the grid will be easier if AI algorithms can run more efficiently, smooth out the dips and spikes, and be responsive to price and valuation changes. Right now, data center owners are not sensitive to price, which suggests that a tariff that enables data centers to participate in demand response could help, but he emphasized the need to take a community-first approach in order to ensure that critical infrastructure, such as hospitals, can continue to access electricity that is affordable, reliable, clean, and resilient.

Wilson asked how to balance short-term operational opportunities with long-term investment goals. Roald replied that both are important, but that better integration and coordination between data centers and the grid in the short term is an important foundation for long-term investments. Roald also pointed out that business interests, more so than utilities, have driven power purchase agreements that have accelerated the clean energy transition. Unfortunately, grid constraints have created long interconnection queues and bottlenecks to integrating more renewable energy sources, which are growing in tandem with data center needs. To address this, she said that data centers need to create flexibilities that enable them to match their loads, regionally and hourly, with renewable power generation to further accelerate and leverage clean energy and avoid wasting it, which can happen if the grid is not able to absorb or distribute it.

When asked what makes today’s large load increases different from those that utilities have faced in the past, Tsuchida replied that data centers are coming online that need power more quickly than utilities’ traditional planning cycles allow for. He also noted that data on the total expected load amount is not enough—what is needed is a better understanding of the more granular and temporal energy needs so that utilities can direct resources more efficiently and cost effectively.

Wilson asked Jain if the tools he described could help to address some of these challenges. Jain replied that long interconnection queues exist because clean energy generation is a highly decentralized and variable two-way flow that is counter to how the grid was designed, creating a bottleneck impeding affordable, equitable, reliable energy. He expressed agreement with Tsuchida that traditional grid planning mechanisms and power sources are inadequate given the uncertainties, and he posited that AI-driven methods are necessary to help utilities address emerging challenges and plan with confidence at a higher level of granularity and accuracy.

Tsuchida added that incorporating flexible demand response into planning may depend on a utility’s location and the willingness of system

operators to pursue this approach. Wilson pointed out that some utilities are starting to incentivize flexibility by moving flexible sources to the front of their load interconnection queues. Roald expressed optimism about data center flexibilities and improved tools, models, and predictions to support grid operations, and said she hoped they would be widely adopted and fully leveraged to optimize the grid.

Julie Bolthouse, Piedmont Environmental Council, remarked that enormous data center growth in Northern Virginia has impacted state and local climate goals by increasing reliance on fossil fuel generation. She asked whether panelists envision a trajectory for AI that will ultimately plateau with some end amount of data generated (and energy needed) or whether the future will be more like a treadmill of unending growth. Tsuchida replied that as the cost of data storage goes down, it creates economies of scale and scope for data generation and use, but how much energy or data is needed ultimately depends on the consumers. Roald noted that efficiency improvements are helpful in keeping energy demands in check, but actual energy constraints would better incentivize data centers to minimize demand. Jain agreed that energy consumption is growing fast but speculated that it will plateau again—after an initial lag—when companies are incentivized to find efficiencies. As for data, he said that data generation will likely continue to accelerate, and since much of the data is of low quality, new software and AI techniques are needed to make better use of the data, especially for data relevant to improving grid planning and operation to meet energy and climate goals.

Benjamin Lee, University of Pennsylvania, asked if data centers were optimally sited. Tsuchida replied that from the perspective of data center owners, they are, because developers consider local factors—like energy rates, workforce potential, building costs, and weather—when siting their facilities. The siting of data centers creates more challenging issues for utilities, however, because they may need to build new generation or transmission capabilities and attempt to pass those costs onto data centers or all consumers, leading to complicated negotiations with multiple competing interests.

Cost, Market, and Supply Chain Issues

Ayse Coskun, Boston University, asked which energy market programs data centers could most easily join. Roald replied that there are a range of options for integrating with the grid, and the answer will depend on a data center’s flexibility across workloads and timescales. Real-time ancillary markets, which pay for providing capacity, require minute-by-minute adjustments, while power envelopes enable participation in day-ahead markets. To facilitate participation in markets and make programs

more accommodating to data centers, she noted that there is a need for a greater exchange of information between data centers and grid operators. Wilson added that European data centers are more involved in frequency regulation and participating in ancillary markets—markets that are currently less well defined in the United States.

Another participant pointed out that AI data centers have a very different balance of capital and operational expenditures than other types of data centers, such as those that support high-energy physics computation. Their capital expenditures are so high that operational costs like electricity essentially amounts to noise compared to hardware costs, which lessens the impact of ancillary service charges in influencing their decision-making. The participant asked if tariffs targeting AI data centers would be a more effective mechanism. Tsuchida replied that there are two tariff ideas—a new, all-inclusive tariff on AI data centers, or applying existing tariffs to AI data centers. He expressed his belief that the latter is more effective and avoids discrimination among customers, noting that some jurisdictions testing new tariffs are coming up against unresolved questions, such as who is charged, how long the tariff lasts, and how a data center can become part of the native load.

Samaras reiterated that the switch from AI training to inference has positive and negative grid impacts. As companies race to get ahead, an urgent request to bring a data center online would be less sensitive to price, so a tariff levied early on can help local utilities build more infrastructure. He also suggested that data centers could become more responsible grid participants by helping to accelerate the transition to cleaner power as overall electrification creates even more load demands. Systematic, deliberate, and ambitious use of AI can maximize benefits, minimize risks, and reduce emissions, but this vision will be off to a poor start if data centers are powered by diesel, natural gas, or other legacy fuels, he said.

Another participant noted that data center growth will impact not just the grid, but also its equipment suppliers, which are experiencing bottlenecks as demand skyrockets. AI data centers are being built by the world’s largest companies, which are far larger financially than utilities and have very few price constraints; the participant pointed out that these companies could potentially vertically integrate suppliers, leading to unfair resource allocation. Wilson added that there is already a backlog of equipment supplies for the electric power grid, and regulated entities often cannot pay above a certain price without special approval. The issue of uncertainty also affects this dynamic, making investment decisions just as challenging for suppliers as for utilities and data centers. Tsuchida expressed his belief that the open market does eventually level out bottlenecks, equipment shortages, and high prices—sometimes with secondary markets and sometimes through new competitors—usually within a few years.

Policy Suggestions

Wilson asked panelists to highlight policy or research ideas that they view as important for making progress toward addressing challenges around the impacts of data centers on the grid. Samaras offered several suggestions to work toward a clean, affordable, reliable, and resilient grid. First, he expressed his support for continuing policies that ramp up nuclear power generation, along with advancing new policies to investigate geothermal power, leverage existing nuclear and hydropower capabilities, and advance long-duration energy storage and upgrades to reconductoring, transmission, and distribution infrastructure. He added that supply-chain challenges could be addressed with a policy compelling advanced market commitments or transmission expansion investments from developers, alongside community benefits agreements to ensure that critical infrastructure does not lose energy access. “Our existing system is built for both the temperature and the loads of the 20th century,” he stated. “We have to rebuild that from scratch in a way that is digitally responsive to all these new loads.”

Samaras also suggested that additional policy mechanisms could help to support the use of AI in enabling 100 GW of VPPs in the next decade to ensure that data center growth does not outpace grid capabilities while simultaneously demonstrating how AI can propel decarbonization goals. Finally, he underscored the need for better carbon accounting methods that enable industry to consider uncertainties, disclose their carbon footprints, and deliver more benefits to society.

Roald suggested advancing policies that incentivize data centers to participate in electricity markets and improve load flexibility. She noted that data centers are not merely reacting to price and carbon signals from the grid but contributing to them both in real time and over the long term, and incentives are needed to encourage them to be more active in grid orchestration. She also suggested that policies are needed to improve carbon accounting for real-time electricity, as current methods are ineffective and may actually increase emissions. If energy customers receive signals that reflect the physical reality of the grid’s complex operations and suggest concrete actions, they can lower their own emissions and that of the grid as a whole.

Jain agreed with the other panelists’ suggestions and recommended, as a short-term solution, a grid planning and operations policy to alleviate the clean energy interconnection bottleneck and enable more AI-driven efficiencies. To address inadequacies in the grid’s digital architecture in the longer term, he said that a new digital architecture is needed to enable secure data sharing within and between companies and utilities and support more precise and auditable carbon accounting measures. He

suggested that this new architecture should be decentralized, data-driven, and malleable—similar to the Internet.

Tsuchida suggested that the AI industry could lobby for a fair overall pricing structure that incentivizes data centers and utilities to make improvements in ways that are mutually beneficial. For example, he said that AI data center operators could do more to try to shrink their energy usage peaks, and creating a standard transmission charge could encourage grid operators and data centers to increase the adoption of renewables and grid-enhancing technologies.

SYSTEM OPERATOR PERSPECTIVES ON CONNECTING GRID-FRIENDLY DATA CENTERS

K. John Holmes, National Academies of Sciences, Engineering, and Medicine, moderated a discussion with Agee Springer of the Electric Reliability Council of Texas (ERCOT) and Robert Wright of Dominion Energy of Virginia. ERCOT is an independent system operator that maintains electric power reliability, ensures equitable interconnection access, and runs wholesale markets across most of Texas. Dominion is a traditional utility that serves residential and commercial customers, generates and transmits its own power, and participates in the wholesale electricity market. The discussion focused on current impacts of AI data centers, future opportunities, and the unknowns surrounding the interactions between AI data centers and the electric power grid.

Holmes noted that AI is at the leading edge of electricity growth as decarbonization strategies take hold. Wright said that AI has generated a lot of excitement in the electric power industry, which has historically been a relatively mundane industry. While AI is proving to be more complex and demanding than other challenges utilities have faced, it also presents unique opportunities to drive innovation and deliver safer, more reliable energy, such as from small nuclear reactors or long-duration energy storage. In addition, the incredibly fast growth of AI is driving innovation in other industries. Springer agreed that AI creates opportunities to make a positive change in the energy sector and the world. Based on current trends, he added that the supply, demand, and reliability challenges in 10 years will likely be very different from those seen today.

Wright stated that AI data center energy demands have the potential to both help and harm the grid. These data centers typically have very high loads that operate at peak capacity with few fluctuations, which is good for grid stability. However, data center load is highly sensitive to grid disturbances and can shut down, requiring major recovery operations for the customer. Data centers often have multiple redundancies to protect against these shutdowns, but as loads become more dense, they

can also upset the balance between service and generation, another large challenge that needs to be addressed.

Springer echoed Wright’s concerns. ERCOT is not synchronized with the rest of the U.S. grid and is much smaller. Data center demands represent only a small percentage of ERCOT’s peak summer load, but they are very sensitive to grid disturbances, especially when they are clustered and still growing, like in Texas. These new loads, totaling an additional 5.5 GW of demand, are seeking immediate interconnection, and utilities are struggling to plan for and model them while maintaining day-today reliability and stability. Storms or load disruptions impact ERCOT’s service-generation balance and could cause cascading outages. Springer emphasized the need to address these issues now through data center designs that prioritize grid stability and reliability and minimize grid overwhelm.

Building on Springer’s remarks, Wright said that Virginia is also experiencing unprecedented demand and new interconnection requests. This has forced Dominion to innovate to maximize existing facilities’ capabilities and improve outage scheduling processes to allow building new projects to maintain power reliability. In addition to making physical grid upgrades, Dominion is removing supply-chain constraints, employing new methods such as energized reconductoring, and improving customer transparency through service request platforms that share planning and process information, he said. Springer stated that ERCOT utilities are dealing with similar supply-chain, transmission, and workforce challenges. He said that the process that governs building new transmission lines is inadequate for today’s fast-paced growth demands, because data center developers want to build where there is already adequate power. Holmes noted that rate payers also have a large role to play that is beyond the scope of this discussion but is fundamental to the electric power system.

Holmes asked if ERCOT or Dominion had experienced demand fluctuations during AI training. Springer replied that ERCOT suspects but has not yet confirmed this and sees these fluctuations as a growing concern. He posited that it may be possible to coordinate load fluctuations with grid operations to avoid disrupting stability or reliability, benefiting everyone.

Wright noted that unknowns around AI’s training and inference process variabilities make predicting grid impacts challenging. To enable a better understanding of these fluctuations, he suggested that utilities and data centers should share data to improve coordination, predict impacts, and understand needed response. He also suggested that training centers may need to be located separately from inference centers, given that their energy needs are different. Springer agreed that grid impacts from AI

fluctuations are challenging to predict and more data is needed. “Every layer we peel back, at least in the ERCOT region, we keep finding five more questions underneath every one that we answer,” he stated. “So, I think understanding is really going to be important for those of us charged with reliability.”