AI's surging energy demands reshape grid

AI energy consumption is forcing a fundamental rethink of grid planning. Current AI infrastructure can require approximately 230 kilowatts per rack, a massive leap from traditional data center demands. Future AI systems could approach one megawatt per rack, creating new pressures as large loads may seek service faster than infrastructure can be planned, permitted and built.

Deterministic forecasts alone become less effective when planners face many possible futures. An AI facility may be delayed, resized, relocated, or phased, and its load profile may change as technology evolves. Renewable output, electrification, distributed resources and transmission constraints add further variables. Utilities should complement traditional studies with scenario-based and probabilistic planning to handle this uncertainty. Key considerations now include project probability and timing, load ramp rates, hourly and seasonal demand, generation and transmission availability, storage performance, flexible-load participation, behind-the-meter resources and permitting timelines.

Certain AI training processes could be shifted, throttled, or scheduled for different times or locations. However, inference and other time-sensitive applications may have strict availability requirements. Behind-the-meter generation may accelerate projects but raises complex questions about fuel, permitting, reliability, market impacts and economics. Planning teams are evaluating more interconnection requests, operating conditions and investment scenarios than ever, often without equivalent growth in engineering resources.

Utilities face complex coordination demands

Multiple stakeholders must align early on timelines, locations and flexibility to manage AI's uncertain growth. Utilities, regulators, system operators, developers, technology providers and customers must coordinate earlier on project certainty, ramp schedules, location, reliability and potential load flexibility. Dominion Energy’s Northern Virginia experience showed integrating large loads is not solely a utility challenge. Partial flexibility could reduce stress during constrained periods, improve use of existing assets and potentially defer selected investments. Options include workload scheduling, demand response, battery storage and phased energization.

Automation and AI-enabled workflows can expand analytical capacity by handling repeatable tasks, supporting data analysis and helping planners focus on higher-value engineering decisions. Faster simulation, improved data integration and interoperable platforms can shorten study cycles and create a consistent view of assumptions and results. Combined with probabilistic methods, AI can expand scenario analysis without compromising engineering rigor. The objective is not to replace engineers, but to give them the speed, scale and information to plan across more possible futures. Utilities must preserve reliability, safety and affordability while becoming more agile and growth-oriented. Physical reinforcement will remain necessary, but building alone is not enough. Utilities must also use existing capacity more effectively, improve data quality, automate repeatable work and evaluate uncertainty systematically.

Real-world examples show strategic power solutions

Amazon's projects demonstrate how existing infrastructure, new gas plants and strategic partnerships enable AI's power needs. Amazon began construction on the world's largest data center in New Carlisle, Indiana, in spring 2024. That facility reached 1 GW of compute capacity in late 2024, 18 months after construction began, by using existing transmission lines and grid capacity. Indiana Michigan Power cut residential rates by $100 per year due to Amazon's demand on that existing grid infrastructure.

For new capacity, Amazon plans to build a 7.65 GW natural gas power plant in West Texas to power a data center campus. Seven additional gas plants are linked to Amazon's data centers, with utilities and independent power producers building 17.5 GW of new gas generation in total. That 17.5 GW equals roughly 1.5 times the peak annual electricity consumption of New York City. Amazon also filed plans for a 36-building campus in Pennsylvania powered by the 4.5 GW Homer City natural gas plant. Satellite images confirm construction has begun at Homer City, with first turbines expected to operate in 2027.

Legislative action is also enabling development. In January 2024, Mississippi passed legislation enabling data center investment and utility infrastructure development. Amazon announced a $10 billion investment in the state, targeting at least $10 billion in data center capital spending. The legislation exempted Entergy's Amazon supply contract from regulatory review and allowed utility infrastructure build-out without a certificate of public convenience and necessity. Entergy is investing $3.4 billion to build 2.3 GW of new gas generation capacity for Amazon in Mississippi, plus 270 MW of new solar generation. By the end of 2027, Amazon is expected to have 1.5 GW of data center capacity in Mississippi, supported by three new gas plants under development.

Amazon has built the largest data center in the world without constructing new power plants or transmission lines. The company operates 825 data centers across 149 campuses in the U.S. And has contracted power for 8 gas plants to support its future data centers. As NVIDIA observed during the webinar, "energy has become part of the AI technology stack." Power availability is now a strategic factor in where AI infrastructure is developed. The goal for planners is to identify decisions that remain sound across many futures, as some AI workloads may be more flexible than commonly assumed. Utilities must implement scenario-based planning to maintain reliability while accommodating AI's evolving power demands.