Artificial intelligence is not just a new tool for writing emails or sifting massive data sets – it’s rapidly becoming one of the biggest new consumers of electricity on the planet.

The foundation models that drive systems like ChatGPT and other generative AI tools require staggering amounts of energy. Training one of these models – sometimes known as large language models, or LLMs – can draw more power than some towns use in a year. Running them daily across vast data centers adds substantial, around-the-clock demand.

This sudden and uncertain growth presents an unprecedented challenge to electricity grids.

Power systems, designed to serve predictable patterns of use from homes and businesses, now face irregular usage spikes. To meet the energy needs of AI, utilities must decide where to place new transmission lines, build backup generation, and manage reserves – all while addressing climate challenges. In places like Georgia and North Carolina, regulators facing forecasts of explosive data center growth are reconsidering long-term energy plans.

The paradox: AI as both stress and solution

Yet the very systems creating these pressures also offer a way forward. The foundation models behind everyday AI platforms are capable of interpreting complex data, advanced reasoning, and assisting humans with decision-making. When applied to the power sector, they are tools to help solve the challenges they create.

For example, utilities often treat data centers as just another customer, even though their consumption rivals that of entire cities. Deciding how and where to connect them to the grid requires highly technical studies of stability, reliability, and transmission capacity. In Texas, the grid operator ERCOT has seen a surge of requests to plug in from AI and cryptocurrency operators. Each request demands months of analysis with specialized software, strict regulatory filings, and discussions with multiple agencies. Foundation models – working independently as “AI agents” – could automate large parts of this process: running simulations in parallel, summarizing results, and proposing solutions like adding batteries or adjusting operations. Engineers would still oversee the results, but the process could move from months to weeks.

PowerAgent: Building an open community

This is not just theory. Researchers and practitioners have begun adapting AI for energy management.

One example is the PowerAgent Community, which brings together tools, workflows, and domain-specific models designed to study the grid. The idea is simple: Instead of each utility or data center building proprietary solutions, PowerAgent provides shared infrastructure for experimenting with AI agents that can interface with trusted engineering software, evaluate trade-offs, and support decision-making in planning and operations.

Imagine a utility planner or grid operator facing a flood of new data center requests. Using PowerAgent workflows, AI agents could automatically launch standard grid impact studies, summarize which requests meet technical requirements, and highlight where system upgrades are needed. Rather than drowning in paperwork and simulations, human engineers would focus on oversight and high-stakes judgment calls.

Energy and cost savings

This kind of collaborative, open ecosystem is essential to making AI a tool for sustainability rather than just another driver of demand.

Companies like IBM and NVIDIA are already deploying AI-powered monitoring systems to spot inefficiencies in cooling and electricity use. These tools help reschedule energy-intensive training jobs to off-peak hours or shift workloads across regions, reducing strains on the grid.

A single data center adjusting its load is like tens of thousands of homes turning down their air conditioning. By making both the grid and data centers more efficient, agentic AI help contain infrastructure costs and stabilize electricity prices. Those savings flow through to ratepayers.

Still, technology alone won’t solve the problem. The rules governing how data centers buy electricity vary dramatically across states. Some are allowed to contract directly with renewable projects, while others must pay fixed retail rates regardless of grid conditions, sucking up supply and increasing costs for others.

That is why researchers are calling for new planning tools that integrate AI, grid models, and policy levers. A shared simulator that captures feedback between computing demand, grid operations, and regulatory decisions would allow stakeholders to explore futures together – testing scenarios like, “What if AI demand doubles in five years?” or “What if data centers are built near renewables instead of coal-fired power plants?”

The choice

The rise of AI represents a crossroads for the energy transition. On one path, unchecked demand locks us into higher emissions, strained grids, and expensive new infrastructure. On the other, the intelligence of foundation models is harnessed to build a smarter, cleaner, more resilient power system.

The PowerAgent community and other open approaches, like the Power and AI Initiative (PAI) at Harvard, show that the second path is possible. But it will require collaboration between researchers, utilities, policymakers, and the tech sector. The same ingenuity that gave us AI can also keep the lights on – sustainably.

Further reading

Qian Zhang and Le Xie, “PowerAgent: A Road Map Toward Agentic Intelligence in Power Systems: Foundation Model, Model Context Protocol, and Workflow,” in IEEE Power and Energy Magazine, vol. 23, no. 5, pp. 93-101, Sept.-Oct. 2025.

Le Xie, Qian Zhang, Minlan Yu, Paul L. Joskow, and Chanan Singh, “Crucial Role of Foundation Models in Enhancing the Interaction of AI and Power Systems”, IEEE Energy Sustainability Magazine, 2025, invited paper.

All perspectives expressed in the Harvard Climate Blog are those of the authors and not of Harvard University or the Salata Institute for Climate and Sustainability. Any errors are the authors’ own. The Harvard Climate Blog is edited by an interdisciplinary team of Harvard faculty.