Generated by Codex with GPT 5.6 Sol High

AI may feel weightless when it arrives through a chat window, but its infrastructure is anything but. Training and running frontier models requires warehouses packed with power-hungry chips, cooling equipment, transmission lines and, increasingly, private power plants. Austyn Gaffney’s investigation follows that physical chain to the Mississippi–Tennessee border, where SpaceXAI assembled scores of gas turbines to bring its Colossus data centers online faster than the electric grid could accommodate them.

The result is a revealing test case for the AI boom. It shows how the race for computing capacity can shift costs onto nearby communities, exploit regulatory categories designed for smaller temporary equipment and overwhelm local governments with projects too lucrative to refuse.

A power plant built at data-center speed

Colossus 2 sits in Memphis, Tennessee, but much of its electricity comes from a turbine complex just across the state line in Southaven, Mississippi. The grid could not supply enough power on SpaceXAI’s schedule, so the company installed portable natural-gas turbines instead of waiting for a conventional plant or a utility upgrade. By July 2026 Mississippi regulators counted 69 temporary turbines at the site—more than twice the number the company had acknowledged at the start of the year. Their combined capacity was about 1,775 megawatts, comparable to a major utility plant and theoretically enough to power more than 1.4 million homes.

That scale reflects an industry-wide constraint. U.S. data centers already consume roughly 4 to 5 percent of the nation’s electricity, and one projection cited in the article puts their 2030 share at 9 to 17 percent. The incentive to move quickly is immense: energy-data firm Cleanview estimates that an AI data center can generate \$10 million to \$12 million in annual revenue per megawatt. A facility that opens a few years early can therefore unlock tens of billions of dollars.

Colossus illustrates the economics. Its first Memphis facility uses more than 220,000 Nvidia graphics processors, and the second is estimated to use more than twice as many. SpaceX disclosed an agreement under which Anthropic would pay \$1.25 billion a month for computing capacity across the two sites, potentially exceeding \$40 billion by May 2029. Against those sums, waiting for ordinary permitting and grid construction can look like an unacceptable delay.

Temporary machines, permanent consequences

Mississippi allowed each portable turbine to operate for up to a year without an air permit. The central legal dispute is whether dozens of nominally temporary machines working together should instead count as one stationary power plant under the Clean Air Act. The NAACP sued, arguing that the fleet was an unpermitted plant. State regulators later permitted 41 permanent turbines with about 1,200 megawatts of capacity, while an agreement gave parts of the temporary fleet a rolling retirement schedule extending into 2027.

The distinction matters because portable turbines are fast but inefficient. Unlike combined-cycle plants, they do not reuse waste heat to make additional electricity; the article reports that a combined-cycle plant can be about 1.5 times as efficient. An NAACP filing estimated that 33 temporary turbines could emit 2,508 tons of nitrogen oxides a year—about 26 times the emissions of the nearby TVA Allen gas plant. Nitrogen oxides irritate lungs and help form ground-level ozone, and both the Memphis and Southaven counties already received failing grades for high-ozone days.

Residents describe harms that are immediate even when their exact source is difficult to quantify. An acoustics report estimated that low-frequency turbine noise could be four to 16 times as loud as the background level in a quiet suburb. Taylor Logsdon says the unending drone keeps her children awake. Air monitors around her home found daily fine-particle concentrations above the Environmental Protection Agency’s health standard about a third of the time, although the monitors could not determine how much pollution came from the turbine site. That uncertainty is important: the evidence supports serious investigation and precaution, not a claim that every measured pollutant originated at Colossus.

The Southaven strategy followed an earlier version in Memphis. There, xAI sought permits for 15 permanent turbines, but environmental-group flyovers found 35 machines operating. The temporary units later disappeared around the time similar equipment began appearing across the border. Community advocates see a repeatable playbook: install modular generation faster than regulators can respond, then convert the project into a permanent facility once its economic importance makes reversal politically difficult.

Who controls the power—and the risk

The project’s leverage extends beyond its size. Mississippi officials point to more than \$20 billion in promised investment, jobs and tax revenue. The U.S. Department of Justice moved to support xAI in the NAACP case, arguing that interrupting Colossus 2 could threaten AI innovation and national security. The Pentagon had already awarded xAI a contract worth up to \$200 million and identified Grok as one of four proprietary frontier models capable of supporting national-security work.

Those benefits are real, but the article asks who gets to negotiate their terms. Residents learned that voluntary promises can be fragile. In Memphis, xAI paused construction of a recycled-water facility before it was finished; officials had not secured a binding completion agreement and were left relying on corporate goodwill. Southaven residents likewise say the company did not consult them before building beside their homes. At the same time, the White House’s voluntary “build, bring, or buy” pledge encourages AI companies to supply their own power without creating an enforcement mechanism for the local consequences.

Communities are beginning to push back. At least 15 states have considered restrictions or pauses on new data centers, and some local governments have imposed temporary moratoriums. Memphis adopted an ordinance directing part of SpaceXAI’s property-tax revenue to nearby public-benefit projects, including environmental work. Yet money does not answer every concern. Advocates want enforceable limits on pollution, water use and siting—not compensation after infrastructure has already locked those burdens in place.

Gaffney closes with Krystal Polk, whose family has owned land beside the Southaven site for generations. She now carries tax records because company representatives have questioned why family members and visitors are near the fence. Her experience gives the article its sharpest lesson: the AI race is not only about chips, models or national competitiveness. It is also a contest over land, law and whose health and peace can be traded for speed. The crucial policy question is not simply how to power artificial intelligence, but who has the authority to decide what that power costs.