According to an industry survey, around a third of data centers could be fully self-sufficient from the electricity grid by 2030, and around two-thirds of players are already planning or deploying on-site generation. Demand from US data centers is reported to reach 82.3 gigawatts in 2026, up 28% year-on-year, with projections of around 167 gigawatts by 2030. A one-gigawatt campus consumes roughly as much electricity as the city of San Francisco.
All year we have documented the physical constraints of AI: water in India, local approval in Pennsylvania, the component shortage. Here is the industry's answer, and it is more radical than expected.
Why they are leaving
Three reasons, and they do not carry equal weight.
Grid connection delays. This is the main reason and the most prosaic. Getting a grid connection for several hundred megawatts takes years: studies, permits, line construction, transformers with delivery lead times measured in quarters. In an industry that builds in eighteen months, waiting five years is disqualifying.
Local opposition. We have documented this on several occasions. Around twenty projects are reported to have been delayed or cancelled within a few months in the United States. Generating your own electricity sidesteps a large share of objections about the strain placed on the grid and on local residents' bills.
Predictability. A negotiated electricity tariff for twenty years beats exposure to market prices, especially when your competitor has signed a twenty-year lease on 191 megawatts.
The accompanying geographic shift
This movement comes with a dramatic redistribution. Texas is expected to capture nearly 30% of the US market by 2028, Georgia's share is set to grow by 75%, while California, Oregon, Iowa and Nebraska would lose more than half of their relative share.
The criterion is no longer proximity to users or land costs. It is electricity availability and how quickly it can be secured.
This is exactly the mechanism we anticipated regarding local blockages: a shift towards territories least equipped to say no, with the equity question that raises.
Here is the point the public debate has not yet absorbed. A data center connected to the grid consumes an electricity mix that is gradually decarbonising, driven by public policy. A data center that generates its own electricity chooses itself its source, and the overriding criterion is rapid availability. Yet what deploys quickly is not what emits the least. An on-site generation facility also escapes part of the territorial energy planning mechanisms. In other words: the industry solves its connection problem by removing itself from the collective decarbonisation trajectory.
The democratic consequence
There is a second effect, less visible and perhaps more important.
We described the decree requiring local authority approval as a constraint that neither engineering nor capital can bypass, because it is settled through negotiation.
Self-generation partially bypasses that constraint. A large part of residents' leverage rested on the strain imposed on the grid and its effect on tariffs. A site that does not touch it deprives the opposition of its strongest argument.
This does not remove everything: land, cooling water, and nuisances remain. But the balance of power shifts clearly in favour of the operator.
The favourable reading, which exists
It would be dishonest to present only one side.
A data center that generates its own electricity does not worsen strain on the grid, and it does not push up household bills. That is precisely what many opponents were demanding.
Some projects come with investments in capacity that subsequently benefit the territory, or with arrangements where the site can feed electricity back into the grid during peak periods.
And a self-sufficient site can, in theory, be powered by low-carbon sources if the operator decides so. Nothing forces the worst outcome; nothing prevents it either.
What would need to be measured
We wrote in our article on AI and climate that the net balance was undecidable for lack of data. This shift makes the problem more acute, because it moves consumption outside public statistics.
A grid-connected site appears in the grid operator's data. A self-sufficient site does not. The ability to measure the sector's real footprint, already weak, degrades at the very moment it becomes most necessary.
What would be needed to settle the matter is fairly simple to state: a reporting obligation for consumption and electricity mix, regardless of source. That is what exists for heavy industry, and AI has become one.
What to take away
This movement is the logical response of an industry in a hurry facing constraints it deems insurmountable. It solves one real problem, and it creates two.
It removes massive and growing consumption from collective planning and measurement mechanisms, at a time when the climate trajectory depends on them. And it strips territories of part of their negotiating power.
The question is therefore no longer whether data centers should be powered, but who decides how they are. As long as they depended on the grid, the answer was collective. It is becoming private, without anyone having debated that transfer.