Ten big ideas
AI needs electricity the way a factory needs workers.
Chips do nothing without power. An idle chip costs its owner far more than the electricity it would have used, so AI firms will pay almost any price for power that arrives soon. Electricity is now one of the limits on how fast AI can grow.
Power has its own needs
Power is harder than it looked.
Demand arrived faster than supply can be built. A data centre goes up much faster than the power it needs: fully connecting a large one to the grid can take five to ten years. Electricity also needs transformers, pipelines, permits and crews, and large transformers now take more than two years to arrive. Many in AI expected to buy power. They found a system they had to fit into.
Electricity needs China the way a factory needs workers.
The same pattern, one step down the chain. China holds 60–85% of the world’s production capacity in the key supply chains for clean-energy technology, such as solar panels, batteries and wind turbines, and over 95% for some steps. It also refines most of the rare earths that go into magnets, from wind turbines to data centres. The country that wants the most AI depends for its power on the country it is racing.
Living with the grid
When asking is cheap, people ask for too much. Then the rules change.
In Texas, reserving grid capacity for a data centre costs little, and walking away costs little too. Developers apply at several sites at once, and planners cannot tell real projects from hopeful ones. By June 2026, requests for new connections came to more than five times the state’s record demand. The rules followed: requests studied in batches, then a freeze on new permits, with new fees on the way. New York’s governor ordered a pause of up to a year on new large data centres.
You can build your own power plant. Few want to stay on their own.
A data centre can make its own power to start sooner. But if its one plant fails, nobody covers for it; on the grid, hundreds of other plants take up the slack. Almost every campus that builds its own power plans to join the grid when it can.
The grid is insurance.
It pools everyone’s risk, so a failure anywhere is covered from everywhere else. It is built for the few hottest and coldest hours of the year, and paid for all year. That is why a large user should pay for the size of its connection, not only for the power it draws.
Cutting back on the worst days is cheaper cover.
The grid strains for only a few hours a year. A data centre that agrees to use less in those hours, by running on batteries or pausing work, can connect sooner and cost others less. One Duke study estimates the US grid could take about 100 GW of new demand this way, if that demand gave up half a percent of its yearly power.
Who pays, who builds
Who pays decides the politics. In India, so does who votes.
In an August poll, 56% of Texas voters said more data centres would hurt their energy bills. In India, factories and offices already pay more so that farms and homes pay less, and cheap farm power is an election promise. The high payers now have reasons to leave the grid, which pushes the cost onto those who stay. A data centre in India walks into that bargain.
In India, the gap is between announcing and building.
Texas has too many requests for power. India’s problem is finishing. Of all the power capacity announced in India, only about 15% of conventional and 9% of renewable capacity has been completed. For AI and for the power it needs, money is not the main constraint; the ability to execute is. The data-centre grid filings tracked so far come to about 5.9 GW, of which 1.5 GW has been granted (September 2026). Those data centres will run on what gets built, not on what gets announced.
What changes
AI makes us rethink the grid.
How it pools risk, how it prices cover, and who depends on whom. When you read a headline about AI and power, ask who is asking for capacity, who covers the risk, and who pays if the demand never comes.
PoolingPricingDependence