The gap between how fast AI labs can build more capable models and how fast the power grid can supply the electricity to train them has become one of the tightest bottlenecks in the entire industry.
Training runs for the largest frontier AI models now require power capacity comparable to a small city, a scale of demand that has turned electricity access, not chip availability or algorithmic breakthroughs, into one of the tightest bottlenecks facing AI labs racing to build the next generation of models.
Several major AI companies have responded by signing direct power purchase agreements with utilities and, in some cases, investing directly in new generation capacity, including nuclear and gas projects, rather than relying solely on the existing grid to supply the enormous, steady load their data centers require.
Grid interconnection queues have become a genuine competitive factor
Getting a new large-scale data center connected to the grid can now take years in many regions due to lengthy interconnection queues, a timeline that has become a genuine competitive factor in the AI race, since a lab that secures power capacity faster can simply train larger models sooner than a rival still waiting on its grid connection.
“We used to think chips were the constraint on how fast AI could scale. Increasingly, the actual constraint is how fast you can get electrons to the building.”
With power demand from AI expected to keep climbing well beyond current grid capacity in many regions, utilities and regulators are increasingly treating data center power requests as a distinct planning category, one that’s already reshaping regional energy infrastructure investment decisions years into the future.
