The AI Power Crisis: America's Grid Cannot Keep Up With Its Ambitions
By Sanna the Weaver • Thu Feb 26 2026 • Technology
The artificial intelligence revolution has an energy problem. The hyperscale data centers being built to train and run the world's most powerful AI models — facilities housing hundreds of thousands of NVIDIA GPUs drawing power continuously — are creating electricity demand that is growing faster than the US power grid can supply. Morgan Stanley's "Intelligence Factory" model now projects a net US power shortfall of 9 to 18 gigawatts through 2028. To put that in context, a single gigawatt is the output of approximately one large nuclear power plant or three to four utility-scale gas plants. The shortfall is equivalent to building 9 to 18 nuclear plants from scratch — which takes a decade. The Scale of Demand A single large AI training cluster — the kind used to train a frontier model like GPT-5.4 or Claude Mythos 5 — can consume 50 to 150 megawatts continuously for months at a time. The largest planned AI data center campus currently under development, in Texas, is designed to consume approximately 1 gigawatt at full capacity — more electricity than the city of Atlanta uses. There are more than a dozen comparable facilities in various stages of planning or construction across the United States, not counting the significant buildout underway in Europe and Asia. Who Is Competing for Power Power purchase agreements have become the most critical competitive asset in AI infrastructure. Amazon, Microsoft, and Google are competing aggressively for long-term power contracts, including deals that involve direct investment in new generation capacity — nuclear, natural gas, and increasingly large-scale solar paired with battery storage. Microsoft's 20-year agreement with Constellation Energy to restart the Three Mile Island nuclear plant generated significant publicity. Amazon has signed deals with small modular reactor startups. The irony is not lost on climate advocates: AI, which its proponents promise will help solve climate change, is driving a surge in new fossil fuel generation in markets where renewables cannot scale fast enough. "We are asking the grid to do something it was not designed to do: absorb massive, continuous, geographically concentrated demand that grows faster than we can build supply." — Department of Energy analyst, January 2026 Grid Vulnerability Beyond the capacity issue is a reliability issue. AI data centers require extremely stable, uninterrupted power — not the variable supply that increasingly characterizes grids with high renewable penetration. Multiple grid operators have raised concerns that concentrating enormous power demand in specific geographic areas — northern Virginia has roughly 35% of the world's data center capacity — creates single points of failure that could have cascading consequences for both AI infrastructure and broader grid stability. The energy crisis created by AI is, for now, a feature of the industry's success. If left unaddressed, it could become a constraint on that success.