Three reports landed within a few hours of each other this week, on three different continents, and they were really all the same report. Greenpeace Australia commissioned an analysis arguing that the AI-driven data centre buildout could demand more new electricity over the next fifteen years than the country's cars or homes. The American Chamber of Commerce in India and Applied Materials warned that "the electricity grid is the single largest constraint on India's AI ambitions." On Wall Street, analysts at Goldman Sachs and the Department of Energy circulated forecasts projecting that U.S. data centre power demand could nearly triple by 2030. Different framings, different politics, identical underlying point.
The eye-watering numbers are starting to add up across jurisdictions. The Australian Energy Market Operator now projects data centres could account for thirteen per cent of national electricity demand by 2040, up from two per cent today, with the forecast having risen more than fivefold between AEMO's 2024 and 2025 reports. New South Wales grid operator Transgrid told a parliamentary inquiry it has fielded interest representing more than ten gigawatts of new load, of which six gigawatts have progressed to formal applications. The agency said its existing regulatory arrangements "were not designed for this level of large, clustered load growth." That is bureaucratic language for: the system is being asked something it was never built to answer.
The financial markets have already noticed. Bloom Energy, a solid-oxide fuel cell maker that sells on-site power generation, signed a ten-year, $2.6 billion master agreement with European AI infrastructure firm Nebius on May 20 to put fuel cells across Nebius's global data centre portfolio. Two days later, Daiwa Securities raised its price target on the stock from $98 to $324, a 230 per cent upward revision. The shares, up roughly 1,400 per cent over the past year, briefly touched $322.83. Constellation Energy, the largest U.S. nuclear operator and the company restarting a Three Mile Island reactor for Microsoft, is being valued as an AI infrastructure play rather than a utility.
The interesting part of this story is not the numbers, which will be revised again next quarter, but what people are now willing to do because of them. Bloom's pitch to AI customers is that they no longer have to wait for the grid: a fuel cell deployment can be running on natural gas in months rather than the half decade some new substations are taking. That is also, of course, an admission. The cleanest version of the AI build-out was always going to involve plugging new compute into new renewables. The actually-happening version, in many places, is going to involve burning more gas to bridge the gap, and running coal plants for longer than their retirement schedules suggested. Australian researcher Dylan McConnell from the University of New South Wales put it plainly: AI demand is "new load," not a replacement for fossil consumption. "As a consequence, we are going to be burning fossil fuels for longer than we would otherwise be."
Politics is starting to follow the meters. Australian state and federal energy ministers, with the exception of Queensland, agreed this month to a proposal that would require data centres to "fully offset" their power consumption with new renewable energy. New South Wales is fast-tracking fifteen projects but the grid operator is asking that the operators, not existing customers, pay for the network upgrades they trigger. The Indian report calls for treating power, semiconductors and compute as a single integrated national system. None of these are yet binding, but they are unusually unified for a debate that did not exist two years ago.
There is one tension worth sitting with. The same companies driving the demand are also among the biggest direct buyers of new clean generation. Microsoft, Amazon and Google have signed nuclear, geothermal and renewable PPAs at unprecedented scale. The argument that hyperscaler load is purely a climate problem is weaker than it was eighteen months ago. But the argument that it is a grid problem, a permitting problem, a transmission problem, and increasingly a community-consent problem, is stronger. The chips are the easy part now. The hard part is the wire to the building, the substation, and the conversation with the neighbours.
If the AI race is going to be won by whoever has the most compute, then the next chapter of that race will be won by whoever can plug it in. That is a very different competition, and most of the incumbents who have been pretending to lead it have not started running yet.