The Public Needs a Better Answer than AI Data Centers
The way the industry has decided to run AI is to build enormous buildings full of chips, and then to build a great many more of them. Those buildings need staggering amounts of electricity and water, and the bill for both is increasingly landing on people who never opened a chatbot. It is worth asking whether that is the only way to do this, or just the first way we tried.
The scale is not small, and it is not slowing
The International Energy Agency projects that electricity used by data centers will roughly double, from about 485 terawatt hours in 2025 to around 950 by 2030, with demand from AI focused data centers more than quadrupling over that stretch. In the United States the picture is sharper still. The 2024 U.S. Data Center Energy Usage Report from Lawrence Berkeley National Laboratory found that data centers already drew about 4.4 percent of the country's electricity in 2023, and projected that share could reach somewhere between 6.7 and 12 percent by 2028.
The public is already paying for it
This is not an abstract line on a corporate budget. Residential electricity prices rose around 11.5 percent in 2025, and in regions thick with data centers they are forecast to climb much further by the end of the decade, according to reporting from Consumer Reports. One grid operator alone, PJM, has been tied to roughly 23 billion dollars in customer price increases driven largely by data center demand, a cost expected to sit on bills for years, as Fortune reported. Utility shut offs have climbed past three million households, and higher bills fall hardest on the people least able to absorb them.
And then there is the water
The chips run hot, so the buildings drink. Research from the University of California, Riverside, led by Shaolei Ren, estimated that training a single large model could directly evaporate on the order of 185,000 gallons of freshwater, and that a model in use spends roughly a bottle of water for every ten to fifty answers it gives. The team's paper, Making AI Less Thirsty, argues that this footprint has stayed largely out of sight while it keeps growing. At national scale it adds up to new water demand measured against the daily supply of a major city.
The shape is the problem
Notice the pattern. Capability that millions of people want is being poured into a small number of resource hungry facilities, owned by a small number of companies, while the electricity, the water, and the strain on the grid are absorbed by the public around them. The benefit concentrates. The cost distributes. That is the wrong shape for a tool this important, and it is the same shape we objected to when we built Osiris Compute.
There is another answer
The compute does not have to live in one giant building. Most of the machines that could run a great deal of everyday AI already exist, sitting mostly idle in homes and offices. Peer to peer browser compute pools those machines into a circle that can carry a real workload, with no new building, no new substation, and no reservoir behind it.
We are honest about the limits. A circle of ordinary devices will not train the next frontier model, and it trades some raw speed for the ability to run without a data center at all, which we measured plainly in the numbers behind Osiris Compute. But for a large and growing share of the AI that people actually use day to day, the honest question is why it should require a power plant and a reservoir in the first place.
The question is not whether AI is worth having. It is whether the only way to have it is to make everyone pay for someone else's data center. We do not think it is, and we are building the alternative.