Energy constraints cap AI data center growth at 10 gigawatts yearly

Naveen Rao, founder of Unconventional AI, contends that energy supply is the critical constraint on artificial intelligence infrastructure expansion. Rao estimates the cost of building one gigawatt of capacity at $50 billion, a figure that exceeds the $20 billion to $30 billion cost of constructing a semiconductor fabrication plant. He notes that while the US power grid averages 900 gigawatts of output with peak capacities between 1.2 and 1.3 terawatts, China surpassed total US capacity in 2023 after expanding its grid six times faster over the previous two decades.

Rao predicts that the US will exhaust its excess capacity within a few years. This limitation will restrict new data center construction to approximately 10 gigawatts per year, a volume worth half a trillion dollars, rather than the 40 gigawatts some industry figures propose. He observes that a two-tiered pricing market is emerging, where data centers pay significantly more per kilowatt-hour than general users. Rao cited a recent deal involving Bloom Energy for several gigawatts in Arizona as evidence of this trend. Regarding financial stability in the sector, Rao views current debt concerns at companies like CoreWeave as a temporary blip rather than a bubble burst. He attributes the resilience of the market to infinite demand for intelligence, arguing that energy scarcity, not a lack of demand, defines the current landscape.