The US map of AI infrastructure is being redrawn in real time. The legacy "data center alley" framing describes only the rear-view mirror.
Northern Virginia hosted approximately 38% of US data center capacity in 2022. By 2030 its share of new AI training builds above 100 MW will fall below 15%. That collapse is not a story about Loudoun County zoning fatigue or Dominion's transmission queue. It is a story about the variables that select sites changing underneath the industry. The map drawn for the workload of 2015 is being replaced by a map drawn for the workload of 2027.
Six structurally different geographies are absorbing the redirected capital. They share almost nothing with the old map. They were chosen for reasons the 2015 site-selection playbook did not score for.
What The Old Map Was Selected For
The legacy clusters were optimized for the workload of the prior decade: cloud storage, enterprise SaaS, content delivery, hyperscale general compute. The selection variables were low fully-loaded land cost, fiber proximity to a major exchange, sales tax abatement, mild ambient temperature for air cooling, and metro labor supply for the operations workforce. Power was a constraint, not a binding one. A 30 MW cloud campus could be sited anywhere a utility could deliver service inside 18 months on a 230 kV feeder.
Northern Virginia won that contest because Equinix sat in Ashburn, MAE-East cross-connect economics compounded, and Dominion was willing to build distribution. Dallas, Phoenix, Santa Clara, and Chicago repeated the formula at smaller scale. The optimization was real. It is also obsolete.
Why The Variables Changed
A modern AI training campus is not a 30 MW building. It is a 300 to 1,500 MW industrial complex with water demand on the order of a mid-sized municipality, a power signature that destabilizes a regional control area if interconnected naively, a national security profile that pulls in federal review, and a construction timeline measured against competitors building the same thing at the same time. The variables that pick sites at that scale are different in kind, not degree.
The new binding constraints are 500 kV interconnect ownership or signed CIR transfer rights, water rights senior enough to survive a drought year, jurisdictional speed measured in interconnection study cycles rather than zoning hearings, brownfield CIR transferability from retired thermal generation, and defense-adjacency for projects that will eventually carry classified or export-controlled workloads. Fiber and tax abatement still matter. They no longer pick the site.
ERCOT energizes a 300 MW industrial customer in 12 to 24 months. PJM energizes the same customer in 60 to 84 months when the queue clears at all. That single delta, four to five years of revenue on a $4 to $8 billion campus, is worth more than every state tax incentive program combined.
The Six New Geographies
The Texas Power Belt. The Abilene-San Angelo-Midland triangle is the densest industrial-power arbitrage zone in North America. ERCOT's energy-only design eliminates capacity payments, the Permian provides behind-the-meter natural gas at wellhead economics, and rural West Texas counties approve industrial plats inside 90 days. Dyess AFB at Abilene and Laughlin AFB at Del Rio create a defense-proximity overlay that the Pentagon's emerging sovereign-compute posture will reward. Stargate's announced Abilene campus is the first signal. It will not be the last.
The Ohio-Pennsylvania Brownfield Corridor. PJM has the deepest brownfield CIR opportunity in the United States. More than fifteen retired or retiring coal stations along the Ohio River and in western Pennsylvania carry transferable interconnection capacity rights between roughly 2,000 and 4,500 MW each: Conesville, Homer City, Bruce Mansfield, Shippingport, Cheswick, Cardinal, Mansfield, Hatfield's Ferry, Mitchell, and the Beaver Valley adjacency to operating nuclear. CIR transfers bypass the 60-month PJM new-service queue entirely. The corridor's labor base is the displaced thermal workforce, which is the closest analog the country has to a turnkey data center construction crew.
The TVA Corridor. TVA carries the lowest wholesale industrial rates among major eastern markets, with large-load contracts pricing inside the $0.04 to $0.05 per kWh band on multi-year structures. Oak Ridge concentrates the national lab and HPC supply chain. Memphis sits at the FedEx logistics nexus. North Alabama runs from Huntsville's Redstone Arsenal cluster through to the TVA generation fleet, and Paducah's West Kentucky position absorbs the legacy gaseous diffusion footprint and its surviving transmission. TVA's announced Clinch River SMR signal, even if delivery slips, is doing what announcements are supposed to do: pulling forward capital that needs a credible 2030s baseload story.
The Carolinas Defense Corridor. Charleston, Greenville-Spartanburg, the Charlotte outer ring, and Wilmington combine Duke Energy's operating nuclear fleet, an industrial base anchored by Boeing, BMW, Michelin, and Volvo, and the densest defense-industrial footprint east of San Antonio. The Carolinas have the workforce, the regulated-utility predictability the brownfield corridor lacks, and a political environment that has not yet hardened against large-load interconnect. Duke's V.C. Summer site and its expansion optionality is the asymmetric asset in the region.
The Pacific Defense Corridor. California's Central Valley, the Inland Empire, the Mojave-Edwards AFB adjacency, and Reno-Sparks in Nevada form a single geography defined by defense-adjacency, federal land, and proximity to the largest concentration of AI demand on the continent. CAISO is slow and CEQA is real, but the federal-land overlay around Edwards and the China Lake corridor moves projects out of state environmental review entirely. The Inland Empire's logistics-industrial base already absorbed the warehouse buildout of the 2010s; the same parcels, the same substations, and the same labor pool absorb the inference buildout of the late 2020s.
The Hidden-Asymmetric Tier. Four geographies sit outside the conventional ISO map and will surprise the consensus by 2028. Bakken stranded gas in western North Dakota offers behind-the-meter generation at flare-economics for operators willing to build in winter. Wyoming's coal-to-gas conversion programs at Jim Bridger and Dave Johnston create CIR-equivalent opportunities under a different regulatory regime. New Mexico's federal-land corridor between Los Alamos, Kirtland, White Sands, and Holloman is a sovereign-compute footprint waiting for the procurement vehicle. Eastern Washington's Hanford adjacency carries surviving 500 kV infrastructure from the legacy nuclear complex and a Columbia River water position that no other Western site can match.
What The Map Looks Like By 2030
By 2030 these six geographies will host more than 70% of US AI training capacity above 100 MW. Northern Virginia's share of new builds falls below 15%, and most of what Virginia retains is inference and enterprise compute, not training. The Inland Empire becomes the largest single sub-cluster by megawatt count, anchored by federal-land projects and Mojave-adjacent campuses that route around CAISO through behind-the-meter generation. The Ohio-Pennsylvania brownfield corridor is the second-largest, driven by CIR transfers that compound as more coal units retire through 2028.
The flagship campuses are already in motion. Stargate's Abilene buildout sets the Texas Power Belt benchmark. Meta's Richland Parish Louisiana site, on the SERC-MISO seam, signals the southern brownfield play. Amazon's Susquehanna nuclear-adjacent campus in Pennsylvania is the template for thermal-to-AI conversion. Google's Dothan, Alabama announcement sits inside the TVA corridor's southern reach. Each of these is a leading indicator, not a one-off. The hyperscalers are building the new map in public; the trailing capital has not yet caught up.
By 2028 the marginal training campus is sited on transferred CIR or federal land. By 2030 the marginal inference campus is sited inside one of the six geographies above and almost never inside the legacy four.
What This Means For Capital Allocation
The asymmetry is in the variables that have not yet repriced. Brownfield CIR positions in PJM are still tradable at fractions of what equivalent greenfield interconnect will cost once the queue resets. West Texas 500 kV-adjacent parcels are still pricing on agricultural comps rather than industrial-arbitrage comps. TVA wholesale contracts remain available to credit-qualified large loads on terms that will not exist once 2027 demand prints. Federal-land Section 7 and military-installation Enhanced Use Lease pathways are open to operators willing to do the procurement work, and closed to everyone else.
The trade is not buying land in the new geographies at consensus prices. The trade is buying the specific structural assets, CIR transfers, water rights with senior priority dates, defense-adjacent ground leases, brownfield substation positions, before the rest of the market reprices them. The window is the next 18 to 30 months. After that the new map is the consensus map and the asymmetry is gone.
The legacy four clusters described where compute happened in the rear-view mirror. The six new geographies describe where it will happen next. Capital that confuses the two will fund the wrong substations.