NEW YORK — What happened: Major artificial intelligence laboratories are increasingly finding that securing compute hardware is easier than securing the electrical grid capacity required to power multi-hundred-megawatt and gigawatt clusters.
Why it matters: Training frontier models demands continuous power draws comparable to small cities. Without immediate electrical infrastructure upgrades, planned exascale training runs risk indefinite postponement, shifting competitive advantages toward firms with pre-existing energy access.
Deep dive: Modern AI training facilities require massive substation transformers, redundant cooling systems, and steady baseload power to prevent thermal runaway and micro-outages that can corrupt long-running gradient descent calculations. Grid operators are struggling to process interconnection queues in time.
Report check (claims vs what is verified vs still rumor): Tech firms claim they are partnering directly with nuclear and geothermal providers to bypass public grid delays. It is verified that average grid interconnection wait times in key US technology hubs have surpassed five years. Rumors suggest several secret desert sites are relying on temporary diesel generation.
Open questions: How will tech companies balance their corporate net-zero commitments while aggressively deploying fossil-backed backup generation to keep training runs from pausing?