Energy for AI
Every gigawatt of AI infrastructure is a negotiation between a hyperscaler, a utility, a grid operator, a regulator, and a community. This piece describes where that negotiation sits right now.
What xAI did in Memphis
The piece opens with the facility. By early 2026, xAI had installed 35 natural-gas turbines on the Colossus site, more than twice what its permit allowed. The Southern Environmental Law Center filed suit on behalf of the NAACP. Mississippi then approved 41 more methane turbines at Colossus 2 across the state line. xAI bought a former power plant site for $10, installed 168 Tesla Megapacks, announced an $80 million wastewater treatment plant, and announced a solar farm next door. The Memphis Chamber of Commerce endorsed the build while community residents filed air-quality complaints. This is the concrete case that frames the rest of the piece.
Section 2What a gigawatt costs to connect
Anthropic said in July 2025 that US AI needs 50 gigawatts by 2028 and that frontier model data centers will each need 5 gigawatts. This section describes what it actually takes to connect that much load: interconnection queues (PJM has a six-year backlog), transmission buildout (AEP took a $1.6 billion DOE loan for 5,000 miles of new line), substation upgrades, and generation procurement. Grid planning operates on 10-year cycles. Data center deployment operates on 18-month cycles. The two time horizons do not fit.
Section 3How these facilities get financed
These builds are not funded only from tech balance sheets. Blackstone, Apollo, and Brookfield are underwriting the pipeline through infrastructure credit. The $7 billion Saline Township project, the $165 billion OpenAI Santa Teresa deal, and xAI's $16 billion Colossus 2 expansion all required external capital. This section describes the financing structure and uses Brian Janous — the former VP of Energy at Microsoft, now at Cloverleaf Infrastructure — as the named source on how capital is actually flowing into this space.
Section 4The FERC decision on Amazon and Talen
In April 2025, the Federal Energy Regulatory Commission denied a rehearing on the co-located Amazon-Talen nuclear pact. That pact would have let Amazon plug a data center directly into a Talen-owned nuclear plant, bypassing transmission and some grid costs. This section describes what colocation means, why FERC's decision mattered, and what happened next: in December 2025, FERC directed PJM to write new colocation rules for large loads. That rulemaking is still running.
Section 5Moratoriums and local opposition
Maine enacted the first statewide moratorium on large data centers in April 2026. New York Democrats proposed a three-year ban. Saline Township, Michigan blocked a $7 billion facility. Bernalillo County, New Mexico adopted strict requirements. New Brunswick, New Jersey cancelled after public protest. Bradley County, Tennessee passed a zoning ban. Doña Ana County, New Mexico sued to block a $165 billion OpenAI facility. Polling shows 40 percent support and 65 percent opposition depending on whose survey. The political reaction function is activating at the local and state level faster than at the federal level.
Section 6What's actually getting built — and why gas
Nuclear gets the headlines. Gas gets built. OpenAI's Stargate facility in Abilene, Texas: 29 gas turbines. xAI's Colossus 2 in Mississippi: 41 gas turbines approved. Crusoe in Wyoming: 1.8 gigawatts of stranded methane. NRG: 4 new gas plants for PJM and Texas. Exxon in talks on gas with carbon capture for AI data centers. The reason is the equipment bottleneck — Bloomberg reported that more than half of US data centers face delays due to equipment shortages, with Vertiv, Cummins, GE Vernova, Caterpillar, and Siemens Energy all supply-constrained. Hyperscalers cannot get clean permanent grid gear fast enough, so they install gas turbines on-site.
Section 7Who pays, and the counter-case
In November 2025, Michigan's Public Service Commission ruled that data centers must cover up-front energy costs instead of spreading them across ratepayers. Xcel Energy filed a new data center tariff in Colorado in April 2026. PJM's Independent Market Monitor documented a $7.3 billion cost shift toward ratepayers. This section describes the regulatory pattern of pushing cost back onto the facilities generating the load. Then it presents the steelmanned counter-case: if tariffs are designed well, data center growth is actually good for utilities — fixed costs spread over more megawatt-hours, and hyperscalers have the balance sheets to sign long-term power purchase agreements that fund advanced nuclear and geothermal. The section closes on the flexible-load question — can training workloads genuinely ramp with grid stress (as Lancium claims in Texas), or is that a greenwashing pitch?
This is a working outline, not the final piece. Research is done and the frame has been stress-tested. Target length approximately 4,500 words. Interviews pending with SELC (Memphis litigation), Brian Janous (Cloverleaf Infrastructure, ex-Microsoft), Jesse Jenkins (Princeton ZERO Lab), Michael Webber (UT Austin, flexible-load reality check), Monitoring Analytics (PJM IMM), and either Sana Ouji or Ariel Horowitz at Anthropic.