Georgetown Law
Technology Review

The data centers that support artificial intelligence consume electricity at a scale rivaling cities. Training OpenAI’s ChatGPT-4 required roughly 50,000 megawatt-hours of electricity. That is enough to power San Francisco for three days. As models grow more complex, training requirements and operational demands soar. ChatGPT-5 is estimated to use around 45,000 megawatt-hours of electricity each day, consuming the same amount of energy it takes to fully charge an iPhone battery for each of its 2.5 billion daily queries.

The development of data centers in the U.S. is projected to increase electricity demand by hundreds of millions of megawatt-hours over the next decade. Many recognize that the sheer volume of this demand threatens the reliability of the bulk power system. Clearly, if we do not have enough electricity due to AI-driven load growth, there will be a problem. Yet, few appreciate that how AI uses electricity, consuming huge and rapidly varying amounts of electricity, is challenging for an electrical system engineered to supply a steady frequency of alternating current. While this issue is secondary to meeting the volume of AI’s energy needs, it could become increasingly important as AI-driven load takes up a larger share of demand. To understand this threat, this Tech Explainer will analyze the nature of electrical currents, examine the origins of our centralized bulk power system, and finally consider why the modern computing equipment at the heart of AI challenges reliability.