AI's Next Bottleneck Is Not a Chip

The next binding constraint in the AI race is not silicon. It is the wall outlet.

For a decade, the U.S.–China contest was framed as a chip race. A growing chorus of infrastructure executives, most recently Hannan Happi, co-founder and CEO of Exowatt, is reframing the contest as a power race. Whichever country can plug new data centers into the grid fastest wins, regardless of who trains the better model.

Happi anchors the case with a number: China adds 540 gigawatts of generating capacity a year, which he puts at more than ten times U.S. capability. Independent data backs the direction, if not the exact multiple. China has been adding the equivalent of several large U.S. grids each year, while U.S. interconnection queues routinely stretch four to seven years. The lead, on this measure, looks structural.

The reader should also know the source. Happi runs Exowatt, a power-infrastructure startup backed by Sam Altman and Andreessen Horowitz, and he sells into the bottleneck he is diagnosing. The product (modular solar-thermal storage sited on so-called frontier land) and the thesis are the same product.

That does not make the claim wrong. It does make it specific and falsifiable. If China's permitting pace slows, or U.S. interconnects clear, the bottleneck migrates back to chips and capital. Until then, the race is being run on a substation, not a fab.