Ask what decides the AI race and the usual answers come quickly: the best chips, the smartest models, the most agents. A run of reporting this week from Seoul, Mumbai, Singapore and Seattle points at something far more mundane and far harder to fake. The binding constraint on artificial intelligence is turning out to be electricity, and the firms that lead will be the ones that can get power onto a site fastest. As the Seoul Economic Daily put it bluntly, what keeps AI executives awake is not the precision of their algorithms but the question of how to secure electricity.
The numbers behind that worry are becoming hard to wave away. The International Energy Agency reckons global data-centre power demand will reach around 1,000 terawatt-hours in 2026, more than double the level of 2022, and keep climbing past 1,500 TWh in the following decade. For scale, the higher figure is more than two and a half times South Korea's entire annual electricity consumption. Crucially, almost all of that growth is AI: electricity use by the accelerated servers that run AI workloads is projected to rise about 30% a year, against single digits for conventional computing. The physical layer, in other words, is now the bottleneck, and it is a slow one to build.
That is reshaping energy policy in real time. Countries are reaching for an "all of the above" strategy, mixing nuclear, renewables and gas, but the tech giants in a hurry are mostly reaching for gas, because a gas plant can be built quickly next to a data centre and feed it directly without waiting years for grid connections. Investors are piling into small modular reactors and grid upgrades on the theory that constant, carbon-free baseload power has gone from a nice-to-have to a precondition for expansion. The scramble is global: India's operational data-centre capacity is forecast to leap from 2.2 gigawatts to 12 by 2030, a build-out that one official estimate says will add more than 26 gigawatts of load to a grid already under strain.
All that demand has a public-relations problem, and this week produced a vivid attempt to manage it. Matt Garman, the chief executive of Amazon Web Services, published a lengthy defence arguing that data centres use "very little water" compared with other industries. US golf courses, he claimed, consume roughly 200 times more water than all of Amazon's data centres combined, and a typical Amazon facility uses less than 13,000 gallons a day, the equivalent of 42 households. He compared the build-out to the interstate highway system and warned that the more than 100 proposed data-centre moratoriums across the United States could cost the country its lead. When a hyperscaler's boss is personally rebutting water soundbites, the politics of power has clearly arrived.
The ripples reach places most AI coverage never looks. At a trade-finance roundtable in Singapore, senior bankers described how energy security and the AI build-out are together redrawing Asia's trade map: longer supply chains as buyers shift from Middle Eastern to American and Latin American energy, vast prepayment requirements for power projects, and a revival of old-fashioned financing instruments to fund it all. The through-line from Seoul to Singapore is the same. The romance of AI is in the models; the reality is in megawatts, cooling water and letters of credit. The open question is whether grids and politics can be built out as fast as the server racks waiting to draw on them, and so far the honest answer is no.