Satellites Reveal AI Data Centers Devouring US Farmland
File photo: An aerial view of a data center roof in Arizona, lined with rows of cooling towers and backup generators — the kind of sprawling physical footprint the Washington Post's satellite analysis mapped across rural America. (Photo: Rsparks3, CC0 1.0 Universal, via Wikimedia Commons)
Update, Sept. 30: the satellite story becomes the debate
The Post’s satellite investigation is now its own media story. Over the weekend, a trending-topic debate on Bluesky erupted over whether the headline framing — data centers “reshape” the American landscape — sanitizes the environmental damage: one viral reply argued the summary should say “destroy,” while others pushed back that such projects are being opposed nearly everywhere they are proposed. Both sides have a point. The backlash is real — farm bureaus calling for moratoriums, the Lancaster County easement fight, even death threats against project proponents in Michigan reported by CNBC — and the rollout is being contested project by project, in county hearing rooms and online alike. This post already carries the strongest corrective in the genre — the “Not all of it will be built” analysis below — and it is this post’s biggest credibility edge over the hot takes: satellite counts of planned sites are not the same as operating campuses.
The numbers behind the sprawl
The Post’s satellite analysis, built on the Cleanview tracker, puts hard numbers on what has until now been mostly anecdote: 1,310 operating data centers across the United States and 2,231 planned. Roughly 100 hyperscale campuses — the 100-megawatt-and-up giants that anchor the AI boom — are running today, with about 120 under construction and some 460 more on the drawing board. A 93-site analysis found that about a third of the large facilities sat on farmland, while 40 percent were built on previously developed land.
Why developers want farmland
The appeal is brutally practical. Data center developers need three things in bulk: large, flat parcels; water rights for cooling; and access to transmission lines. Farmland counties happen to bundle all three, plus cheaper rural land near existing fiber routes. The Post’s reporting found developers quietly assembling parcels for years before announcing projects, sometimes through shell LLCs to avoid tipping off land markets — a practice that has infuriated neighbors who learn about a 500-megawatt neighbor from a satellite photo rather than a public notice.
When the grid can't keep up, bring your own power plant
The most striking finding in the Post’s reporting may be what is happening with electricity. Because data centers are straining the capacity of the electric grid, the paper reports, some are being built with their own power generation — in part to blunt local criticism that these operations raise electricity costs for everyone else. That makes the sites even more sprawling: a campus that generates its own power needs room for the turbines, fuel storage and batteries too.
The workarounds have entered genuinely strange territory. In Texas, data-center operators are deploying aeroderivative gas turbines — effectively retired commercial aircraft engines, including General Electric’s CF6-80C2 and LM6000 turbine cores once found on 767s and Airbus A310s, converted into modular generators by vendors like ProEnergy and Mitsubishi Power — that can spin up tens of megawatts in months rather than the years a grid connection takes, according to IEEE Spectrum’s reporting. FTAI Aviation is converting retired CFM56 jet engines into 25-megawatt natural gas turbines aimed at hyperscale sites, with production starting this year. GE Vernova is supplying similar aeroderivative units to large AI projects.
The industry is trying to turn this improvisation into strategy: Google, NVIDIA and Emerald AI’s new grid-flexibility alliance aims to make data centers flexible grid assets through demand response, colocated generation and performance-based interconnection standards.
The boldest example belongs to xAI. The company’s Colossus complex in the Memphis area sits next to an existing gas power plant, and rows of giant batteries help stabilize the power supply during peak demand, the company has said. By one industry estimate cited in the reporting, the site could eventually draw 1.5 gigawatts — more than the entire city of Memphis at peak. The Barn project in Michigan’s Saline Township (a Related Digital build for Oracle and OpenAI under the Stargate umbrella, reported at $16 billion) and Caterpillar’s $725 million generator plant in Indiana complete the picture of an industry building its own power sector in parallel with the one it keeps breaking.
The backlash in farm country
The political backlash has been swift and bipartisan. The Pennsylvania Farm Bureau has called for a moratorium on large data-center projects; in Michigan, opposition to The Barn project turned openly hostile, with reports of threats against project figures. In Lancaster County, Pennsylvania, Representatives Thompson and Kennedy have been drawn into an easement fight over farmland preservation. The Post’s own comment sections and the Bluesky debate capture the mood: a technology sold as weightless keeps landing, very heavily, on someone else’s soil.
The tension has a number that explains it. The grid operator’s monitor for the PJM region — covering 13 states plus D.C. — has identified data-center load as “the primary reason” for high capacity-market prices, which cleared $23.1 billion through 2028. When a farmer’s electricity bill rises because a hyperscaler three counties over needs more electrons, the abstract AI boom becomes a kitchen-table issue.
Not all of it will be built
The essential corrective, and this post’s strongest card: satellite counts of planned sites are not building permits. Analysts note that land-clearing for new U.S. data centers has plateaued, and Columbia Business School’s Stijn Van Nieuwerburgh has argued that a large share of announced projects — his estimate runs to 45 percent — will never be built, stranding some $9.3 trillion in implied investment. The Jefferies/SynMax analysis that data-center chip supply (HBM and advanced packaging) will cap deployment in the low-20s of gigawatts by 2027 says the same thing from the supply side: the physical build-out has a ceiling, and the satellite maps show the ambition, not the outcome.
AI data centers and farmland: your questions, answered
How many US data centers are operating, and how many are planned? Per the Cleanview tracker cited by the Post, 1,310 are operating and 2,231 are planned. Roughly 100 hyperscale campuses (100 MW or more) exist today, with about 120 under construction and 460 more planned — which is exactly why analysts keep warning that a large share of the “planned” column will never break ground.
How much farmland are data centers actually taking? In the Post’s 93-site analysis, about a third of the large facilities sat on farmland, while 40 percent were built on previously developed land. The pattern varies wildly by county: farmland is attractive where parcels are large, flat and cheap — not because developers are targeting crops.
Why are data centers being built on farmland at all? Three things developers need in bulk: large flat parcels, water rights for cooling, and access to transmission lines — plus cheaper rural land near existing fiber routes. Farmland counties happen to bundle all four.
Are data centers raising Americans’ electricity bills? The grid operator’s own monitor has called data-center load “the primary reason” for high capacity-market prices in the PJM region, where auctions cleared $23.1 billion through 2028. That is the figure fueling the ratepayer anger from Virginia to Ohio — and the reason some developers now build their own power plants to blunt the criticism.
What is actually at stake
The Post’s satellite work reframes the AI boom as a land-use story, and that is the right frame. The fight over where the models live — on whose farmland, on whose grid, at whose water table — will shape the technology as much as any benchmark. Trump’s Sept. 29 accord included energy pledges for the AI build-out; we covered the accord and its data-center energy pledges in detail here. Whether the pledges survive contact with county zoning boards is the story to watch next.
Sources: Washington Post satellite analysis (Sept. 28, 2026) via the Cleanview tracker; CBS News; CNBC; IEEE Spectrum; Wall Street Journal; Barron’s.
Comments
Post a Comment