What the data centres cost
The cloud is somebody's neighbourhood. Here is the bill.
The privacy page is about what happens to your data once it leaves your house. This page is about what happens to a place when the building that holds it arrives. Every large AI service you can type into runs in a warehouse of machines somewhere, and those warehouses have to be put down in a real town, on real land, drawing real water off a real aquifer, connected to a grid that other people are also paying for.
None of this is a secret and none of it is contested. It is in national statistics offices, federal court filings, peer-reviewed journals and the companies' own environmental disclosures. Every claim below links to its source.
1. The scale is not "a few sheds"
The US Department of Energy asked Lawrence Berkeley National Laboratory to measure it. Their 2024 report found data centres consumed about 4.4% of all US electricity in 2023, up from 58 terawatt-hours in 2014 to 176 in 2023, and projected 6.7% to 12% of the entire national supply by 2028. That is a range wide enough to be honest about the uncertainty, and even the bottom of it is a doubling in five years.
Globally the International Energy Agency puts data centres at around 415 TWh in 2024, roughly 1.5% of world electricity, and projects about 945 TWh by 2030. The IEA names AI as the main driver: consumption by AI-accelerated servers is projected to grow 30% a year, against 9% for ordinary ones.
In Ireland it has already arrived at the end state. The Central Statistics Office measured data centres at 22% of all metered electricity in 2024, up from 5% in 2015. Urban homes were 18%, rural homes 10%. The grid operator stopped accepting new Dublin connections in 2021 and the freeze held for four years; the rules that replaced it now require a new data centre to bring its own generation or storage to cover its full demand. A country ran out of headroom.
The National Wildlife Federation counts roughly 4,000 data centres operating in the US with about 3,000 more built or planned, and points out the timing problem underneath all of this: a data centre takes 18 to 36 months to build, and the transmission line to feed it takes seven to ten years. The gap gets filled with whatever can be switched on fastest, which is section 3.
2. The power bill lands on the neighbours
Researchers at NC State, Carnegie Mellon, Pittsburgh and Toronto modelled what the announced data centre and crypto load does to American power prices by 2030. Published in Environmental Research Letters in May 2026, the finding is a 6% to 29% rise in demand-weighted wholesale electricity prices nationally, up to 57% in the worst-hit regions, which they name: Virginia, eastern North Carolina, Pennsylvania, Maryland, Delaware, New Jersey, west Texas, Ohio, West Virginia and New York. The same modelling has CO2 emissions up to 28% higher than a future without the buildout.
Nobody in those states voted for that. The load arrives, the grid is upgraded to carry it, and the upgrade is recovered from everyone's bill, because that is how regulated utilities work. Whether the data centre or the household pays is decided by tariff design in front of a public utilities commission, in hearings almost nobody attends. EESI has a readable summary of how that cost allocation actually works. It is worth knowing that this is a decision someone makes, not weather.
3. Water, in places that did not have spare water
Cooling a building full of GPUs mostly means evaporating water, and the water is usually the same treated drinking water everyone else is on.
In Newton County, Georgia, Beverly and Jeff Morris live about 1,000 feet from a Meta data centre. The New York Times reported in July 2025 that months after construction started their dishwasher, washing machine and toilet stopped working, and the pressure in the taps fell to nothing. They are on a well. Meta disputes that its construction caused it. What is not in dispute is the county's own trajectory: local water costs have risen sharply and Newton County is projected to be in water deficit by 2030.
In Uruguay, during the country's worst drought in 74 years, Google proposed a data centre that would have drawn 7.6 million litres of drinking water a day to cool its servers, equivalent to the domestic use of 55,000 people, on the environment ministry's own figures. Montevideo was under a state of emergency and the authorities were mixing salty water into the public drinking supply to keep taps running. People went to the streets, and the project only proceeded after Google dropped water cooling for air cooling.
In Chile, a court partially revoked the environmental permit for a planned Google facility in Cerrillos over its groundwater use, and Google paused the project and redesigned the cooling. In both cases the fix was available all along. It was cheaper not to use it until somebody made them.
4. Air, and the sound that never stops
When the grid cannot deliver power fast enough, the answer is often to burn gas on site. In April 2026 the NAACP, with the Southern Environmental Law Center and Earthjustice, sued Elon Musk's xAI in federal court over 27 methane gas turbines running without a Clean Air Act permit at Southaven, Mississippi, to power the Colossus 2 data centre behind the Grok chatbot. The filing puts the turbines' potential emissions at over 1,700 tons of nitrogen oxides, 180 tons of fine particulate matter, 500 tons of carbon monoxide and 19 tons of formaldehyde a year. The company's other site sits beside majority-Black neighbourhoods in south Memphis that already carry a refinery, a steel mill and a gas plant. The case is live and xAI contests it. Read the complaint rather than anybody's summary of it, including this one.
Then there is the noise, which sounds trivial until it is your bedroom window. Data centre cooling runs 24 hours a day and produces a low hum that carries. Residents in Prince William County, Virginia describe it as catastrophic, Chandler, Arizona rewrote its zoning code in 2022 after nearly a decade of complaints, and there is now an established wave of nuisance and negligence litigation over it. One Michigan operator's response to the lawsuits was to offer to buy the neighbours' houses.
5. Land, and the things living on it
These buildings are enormous and they are increasingly going up in rural places, because that is where the cheap land and the substation capacity are. In Virginia, conservation groups have raised objections to a Botetourt County proposal that would clear about 120 acres inside the range of the endangered Indiana bat, and to a project of some 2,000 acres next to Manassas National Battlefield Park.
The habitat cost is not only the footprint. The transmission lines that feed these sites cause direct habitat loss, fragmentation and disruption of migratory routes, and greater sage grouse have been shown to avoid habitat near them outright. A Wisconsin transmission line built to serve an AI data centre is being fought by farmers, conservation groups and local government over farmland, wetlands and reforested land. Rural land converted to industrial use does not usually convert back.
The freshwater picture is only now being studied properly. Jager and Yoon published the first risk-assessment framework for data centre effects on US freshwater biodiversity in Water Biology and Security in 2026, covering both the direct effects of water withdrawal and warm discharge and the indirect ones through thermoelectric and hydroelectric generation. Their honest conclusion is that the question had not been examined before. We are building at speed in front of the research.
6. What the town gets in return
The pitch to a local council is jobs and tax base. The jobs number is the part worth checking. Good Jobs First found an average of about 2 million dollars in subsidy per job across major data centre deals, and recommends public officials cap incentives at 50,000 per job. In Illinois, subsidised operators invested over 5.4 billion dollars and created 339 jobs. A JPMorganChase expansion in New York drew a 77 million dollar break for one job. Virginia's data centre sales tax exemption cost the state 1.6 billion dollars in 2025.
A finished data centre is a very large building with a small number of technicians in it. That is not a scandal in itself, it is simply what the building is. It becomes a problem when it is sold to a community as an employer and paid for as if it were one.
7. This is not fringe. It is 833 groups and 49 states
Data Center Watch, which tracks the industry's own losses rather than campaigning against it, recorded at least 75 projects worth around 130 billion dollars blocked or delayed in the first quarter of 2026 alone, matching the whole of 2025 in three months. Active opposition groups more than doubled in a quarter, from 396 to 833. Over 300 state bills were filed in the first six weeks of the year and moratorium proposals appeared in 14 states, from both parties. Their earlier two-year report found the politicians opposing large projects split 55% Republican and 45% Democrat.
Nobody is against computers. People are objecting to a specific bargain being offered to them: your water, your grid, your quiet and your fields, in exchange for a service you will then be charged a monthly fee to use.
What Fred does instead, and what it does not fix
Fred runs the model on a computer you already own, in your house. When you ask Fred something, nothing travels to a data centre, nothing is queued behind someone else's inference, and no capacity has to be built anywhere for you to be able to ask it. Your AI use stops being a claim on a town you have never visited and becomes a line on your own electricity bill, where you can see it.
Being straight about the edges, because an environmental page that overclaims is just marketing:
- Fred does not make data centres unnecessary. The open-weight models Fred runs were trained in one. Training is a real cost that has already been paid, and running an existing model on your own machine does not undo it. What it avoids is adding to the demand that justifies the next building.
- We are not claiming a better watt-hour per question. A data centre GPU is genuinely more efficient per answer than a machine in a spare room, and honest measurement of a single cloud query, Google puts a median Gemini text prompt at 0.24 watt-hours, makes any one query look small. The argument on this page is not about one query. It is about what a continent-scale build of them costs the places it is built in, and who is asked before it happens.
- Inference is the part that keeps growing. Training is a one-off per model; answering is forever, and the IEA expects AI-server electricity to grow 30% a year on the back of it. That is the share a local machine actually removes from the grid's future.
- Fred still uses power. It runs on hardware you own, it idles when you are not using it, and you pay for what it draws. That is the point: the cost is visible and it is yours.
- The relay is ours and it is small. When your phone reaches Fred from outside your home, it goes through a relay in Chur, Switzerland, running on our own hardware. It handles encrypted messages only and stores none of your content. The privacy policy covers it line by line.
A computer in your house is not a small idea
For about forty years, personal computing meant the machine was in the room with you and the electricity it used was on your own meter. The last fifteen years quietly moved that machine into a warehouse in someone else's county, and the bill for it into their water table, their grid and their sleep. Fred is an attempt to move it back. It is free, it runs on hardware you already own, and the only place it costs anything is your house.