AI Data Centers Face Power Gaps Through 2031 as Grid Limits Force Delays

Multiple major institutions have published forecasts showing AI data center power demand will far outpace supply well into the 2030s. JPMorgan projects global AI compute demand will reach roughly 530 GW by 2031 while supply reaches only 250 GW, leaving a roughly 280 GW shortfall even as supply quadruples. Morgan Stanley estimates US data centers could require roughly 160 GW of power by 2029, with a potential 76.5 GW shortfall even after accounting for all planned capacity. SemiAnalysis forecasts 51.2 GW of new data center capacity will be added in 2027 alone, more than double the 25 GW added in 2026, with satellite imagery confirming more than 24 GW already in advanced construction and approximately $1 trillion in non-cancellable leases signed in 2026.

The physical bottlenecks are already producing concrete project failures. Microsoft walked away from a planned data center in Loughton, Essex after grid officials indicated sufficient electricity would not be available until the 2030s. Texas paused new data center approvals after ERCOT's large load queue grew from 63 GW to 474 GW in 18 months, with roughly 90% tied to data centers but only 4.3 GW operational. High-voltage transformers carry lead times of five or more years. Building a gigawatt-scale data center in Texas now requires paying $584,000 per megawatt for grid-connected land, while roughly $750 billion in hyperscaler capital expenditure is competing for the same turbines, steel, copper, and fiber, pushing power-gear backlogs past 2030. Amazon CEO Andy Jassy told an investor that AI compute capacity is sold out beyond 2028, even for a hyperscaler with enormous capital access.

On policy, President Trump posted on Truth Social that large new data centers must generate their own electricity and pay host communities a cash dividend. Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI had previously signed the Ratepayer Protection Pledge requiring signers to build or buy new generation and cover infrastructure upgrades, though experts consider the pledge likely unenforceable at the federal level.

SemiAnalysis analyst Dylan Patel argues that while power is a real bottleneck, supply-chain limits in substrates, memory, lasers, and optics are more binding than power, and that on-site solutions such as turbines, reciprocating engines, and fuel cells are helping ease the power constraint.