For two years the story of artificial intelligence was told in chips. Whoever had the most Nvidia silicon would win. This week, a run of reports from four continents made clear that the industry has quietly moved past that framing and hit a harder limit: the electricity and water needed to run the machines. The constraint is now the grid, and every country is reaching it at roughly the same moment. What is striking is how far apart their responses are already drifting.
Australia offered the sharpest picture of the shock. The Australian Energy Market Operator now expects data centres to draw 34 terawatt-hours a year by 2036, roughly as much as every home in New South Wales and Victoria uses today. A year ago the same body did not expect that level until well after 2050. The number of centres in its development pipeline has more than doubled in twelve months, from 97 to 225. Prime Minister Anthony Albanese responded by moving to require new centres to bring their own renewable supply, striking power purchase agreements that also help finance the wind and solar projects the grid needs anyway.
The United States is racing in the opposite direction. According to Global Energy Monitor, developers there have about 189 gigawatts of gas-fired capacity in development aimed at data centres, nearly double last year's total, and the country is now building twice as much gas power as China. The Trump administration frames this as the winning move, urging AI firms to supply their own power so bills do not rise. But the report warns that two-thirds of proposed gas capacity worldwide has no turbine or engine maker named yet, locking in emissions for plants that may never clear the multi-year queue for parts.
India, meanwhile, is treating the same demand as an opening. A Wood Mackenzie analysis projects data-centre electricity use rising nearly twentyfold, from 10 terawatt-hours in 2025 to 191 by 2040, and casts the whole thing as a renewable-energy investment story: solar, wind, battery storage and transmission suppliers all stand to gain, with hyperscalers such as Meta already pairing new campuses with gigawatt-scale clean procurement.
Underneath the power question sits a quieter one about water. A new Ceres report estimates that generating the electricity for data centres in just seven US states relies on about 3.4 trillion gallons of freshwater withdrawals a year, much of it in already stressed regions. In Kenya, a planned Microsoft and G42 campus near Lake Naivasha has stalled, and residents who already ration water fear being outbid by, as one pastoralist put it, "the multibillion-dollar companies." The centres promise jobs; the neighbours count what leaves the tap.
The common thread is that speed to power, not speed to chips, now decides where AI gets built. The divergence is in what "power" is allowed to cost. Australia is tying growth to new clean supply and community consent, the US is betting on gas and self-generation, India sees a green windfall, and places like Naivasha are left asking who pays when a data centre and a village want the same river. Those are not technical choices. They are political ones, and they are being made right now, mostly before the public has been asked.