AI Power Demand Outpacing US Grid Infrastructure

US AI power demand has grown nearly tenfold—from roughly 3 GW in 2023 to a projected 28 GW by end of 2026—outstripping the grid's ability to absorb it. [1] PJM, the largest US grid operator, and MISO face interconnect queues that were never designed for this scale of step-function demand growth. [1][3] AI labs have responded by quietly adopting onsite natural gas generation as 'the default planning assumption for the next wave of US training clusters,' effectively bypassing the grid. [4] Meanwhile PJM is accelerating its backstop processes and partnering with Google and startup Tapestry to deploy AI tools to clear the backlog. [8][9]

The grid mismatch is not a short-term planning gap but a structural realignment: private AI infrastructure is outrunning public electricity infrastructure, pushing the industry toward a parallel power system built on onsite gas. This has compounding implications—for carbon commitments, for who pays the utility bill, for grid reliability for existing customers, and for whether regulatory timelines can ever catch up to the pace of AI deployment.

The US electrical grid is experiencing a demand shock with few historical precedents. AI data centers have driven power consumption from roughly 3 GW in 2023 to a projected 28 GW by end of 2026—a near-tenfold increase in three years. [1] PJM, the largest US grid operator (serving some 65 million people across 13 states and parts of Canada), and MISO face interconnect queues that were designed for incremental demand growth, not the step-function jumps driven by GPU clusters. Grid engineers, utility executives, and regulators have described a system under mounting strain. [2][3]

Facing multi-year interconnection delays, AI labs have made a pragmatic pivot: onsite natural gas generation has shifted from a niche fallback to the dominant planning assumption for new US AI training clusters, according to a SemiAnalysis investigation that tracked the trend across major hyperscalers. [4] Rather than wait for grid access, labs are building their own generation capacity at data center sites—sidestepping interconnect queues but also, critics note, sidestepping clean energy commitments. [5][6] The shift has been documented across industry sources including Latitude Media and Marketplace, and is reflected in a broader wave of data centers planning their own gas plants. [7][6]

PJM has begun mounting a response. In mid-May 2026, the grid operator accelerated its backstop interconnection process for data centers. [8] It also announced a partnership with Google and startup Tapestry to deploy AI software to help clear the interconnection backlog—an arrangement with a certain irony, given that AI demand helped create the crisis. [9][10][11] FERC separately directed PJM to develop new rules embracing innovation in grid interconnection. [12] Merchant power generators well-positioned to supply data centers directly—Constellation Energy and Vistra—saw their stocks rally in response to PJM's accelerated processes. [13] Utility consolidation is also in play: observers note that NextEra's reported interest in Dominion's PJM interconnect positions reflects AI compute demand meeting utility-scale M&A logic. [14]

The problem extends well beyond the United States. Denmark paused all new data center grid connections after total queued requests reportedly reached 60 GW—a figure that dwarfs the country's current generation capacity. [15] Microsoft's planned billion-dollar data center in Kenya has sparked fears of blackouts among local residents whose grid cannot absorb the load. [16][17] India faces its own tension between AI infrastructure ambitions and electricity supply constraints. [18] Hardware bottlenecks compound the picture domestically: large power transformers now carry lead times of roughly four years. [19] Battery storage firms are positioning their technology as essential for managing the sharp, fast demand swings characteristic of AI workloads, with some observers comparing batteries' coming scarcity to that of semiconductors. [20][21]