The Pentagon wants AI data centers on military bases; a single one in Texas would rival El Paso's electricity consumption

By Zendoric · July 25, 2026.

The U.S. Army has already closed two specific conditional deals: Carlyle Group will build an artificial intelligence data center on 1,384 acres at Fort Bliss (El Paso, Texas) by 2027, and CyrusOne will do the same on 1,201 acres at the Dugway Proving Ground (Utah) by 2029, according to a press release and fact sheet from the Army itself cited by Politico. The Air Force, for its part, has already offered unused land at five bases for private companies to build similar facilities. The Pentagon's stated goal is to use these servers to power the drone swarms and other AI-based warfare systems that Washington considers key to not falling behind China.

The push began last year with an executive order from Donald Trump that directed the Department of Defense to identify military land suitable for AI infrastructure, and that also required the EPA and the White House Council on Environmental Quality to facilitate construction. The Army maintains that these projects will come "at no upfront cost to the taxpayer" and with "comprehensive energy and water solutions off the public grid" (behind-the-meter, in industry jargon: self-owned generation and supply, not connected to the general grid).

That last point is exactly what a bipartisan group in Congress does not believe. The figure that has set off the alarms, according to a report by El Paso Matters cited by Politico: the Fort Bliss data center alone could require more electricity than all 460,000 electricity customers in the entire city of El Paso combined. It is the kind of number that turns an abstract discussion about military infrastructure into a very concrete problem of power grid and water supply for the base's neighbors.

The legislative response, for now, does not seek to halt the projects but to shed light on them. Most of the amendments under way —in the House and the Senate, in the versions of the Defense Authorization Act (NDAA) and the fiscal year 2027 defense appropriations— call for studies, assessments or prior reports, not bans. Republican Ken Calvert, chairman of the Defense Appropriations Subcommittee, managed to include a provision requiring the Pentagon to brief Congress on the impact on communities before construction begins. Democrat John Garamendi went further: he proposed that the Department of Defense formally assess energy and water consumption, light and noise pollution, and security and supply chain risks before closing each deal; Republicans on the House Armed Services Committee rejected the amendment during its markup.

Other proposals do aim to cut things off at the root. Rashida Tlaib proposed directly banning the construction of AI data centers on federal land, including bases; after it failed as an amendment, she reintroduced it this week as a standalone bill. Hillary Scholten wants to impose minimum water efficiency standards and is preparing her own legislative version. Republican Cory Mills got the NDAA passed by the House this week to include a clause preventing land from being handed over to data centers with "significant" components manufactured by U.S. adversaries. In the Senate, the Armed Services Committee —with a Republican majority— rejected a Democratic amendment that would have blocked land leasing until the Pentagon finalized its data center strategy, already required by law.

The other side of the debate —the Pentagon itself and part of the Republican caucus— warns that all this oversight, however well-intentioned, could turn into friction precisely when moving forward is most urgent. An Army official, quoted by Federal News Network, called Mills's clause an "unworkable standard" and warned that it could scare off commercial partners and hold the country back against its adversaries in the race for advanced AI capabilities. Republican Pat Harrigan was more graphic: he said that requiring "precertification" does not protect bases or communities, it only "buries them in paperwork"; he prefers annual reports. His colleague Derrick Van Orden argues outright for putting data centers on military installations rather than on farmland.

Our read: this fight in Congress is the military version, in miniature, of the problem we already identified as the real bottleneck of this AI decade —not the silicon, but the electrons. As sector context, projections point to strong growth in data center electricity consumption globally over the coming years, heavily concentrated in the U.S. and China; the source does not provide figures on this. What is happening at Fort Bliss or Dugway is that same clash —compute against power grid and aquifer— moved to terrain where national security is also invoked to skip much of the environmental and planning scrutiny that a civilian data center does face.

That is the most interesting part of the episode: the dominant pattern is not "ban military AI" but "require it to be measured and reported", the same mold of evidence-based governance —briefings, reports, impact studies— that we have already seen take hold in other high-risk AI areas, from ultimate human control over weapons systems to algorithmic surveillance. It is a healthy sign: neither the doomsaying of banning outright (the Tlaib route, a minority position) nor the Pentagon's "trust us" (no cost, no friction, everything solved "behind the meter") is prevailing. For now, the middle line wins: yes to the infrastructure, with mandatory transparency.

In the short term, this is a real problem with no shortcuts: the residents of El Paso or the area around Dugway did not choose to compete for megawatts with a defense data center, and the tension between electricity prices, grid capacity and AI projects —civilian or military— will keep escalating as long as new generation is not built at that pace. In the long term, however, every dispute like this one pushes toward solving sooner the energy problem that underlies the entire AI revolution; and that same technology, applied to reactor design, smart grids or battery materials, is the best positioned to solve the bottleneck it created itself. The Pentagon is not going to stop building these centers —none of the serious amendments ask for that—; what is really at stake in Congress is who pays the energy bill for the AI arms race, and who finds out before the crane arrives at the site.