More than 60% of all US hyperscaler AI capacity lives inside five metro areas — and every one of those markets is now constrained on power.
Northern Virginia's Loudoun County — called "Data Center Alley" — hosts over 130 operational facilities and handles an estimated 35% of North America's internet traffic. Dominion Energy has a multi-year backlog for large commercial power connections in that county. New development is already spilling 10–20 miles south.
This matters because geography isn't just a logistics decision for AI companies. When Microsoft, Google, Amazon, and Meta are each committing $60–80B per year in capex, deciding where to build is a trillion-dollar call that shapes power grids, real estate markets, and regional economies for decades.
AI Data Center Location Geography: The Top US Markets
The five largest US data center markets by operational capacity, as of mid-2026:
| Market | Est. Capacity (MW) | Key Operators | Primary Advantage |
|---|---|---|---|
| Northern Virginia | 3,700+ | AWS, Google, Microsoft, Equinix | Legacy fiber, low natural disaster risk |
| Phoenix / Scottsdale | 1,600+ | Microsoft, Google, CyberNAP, QTS | Land availability, energy costs, tax incentives |
| Dallas–Fort Worth | 1,400+ | Google, AT&T, Equinix, Digital Realty | ERCOT grid, central US location |
| Chicago | 1,000+ | Microsoft, Equinix, Aligned | Midwest transit hub, water access |
| Silicon Valley | 850+ | Google, Meta, Oracle | Talent density, network effects |
| Atlanta | 650+ | Google, Microsoft, Compass | Energy costs, Georgia incentives |
These numbers shift quarterly. Amazon alone added 2 GW+ of contracted US capacity in 2025, and Microsoft's $80B capex plan includes buildouts in 18 states. Track real-time AI infrastructure spending at the AI Spending Dashboard.
What Makes a Data Center Site Win: The PILE Framework
Every hyperscaler uses some version of the same four-factor site selection model. I call it PILE:
Cheap, abundant, and reliable. NOVA runs at $0.04–0.06/kWh industrial rates. A hyperscale campus pulls 100–500 MW continuously — the equivalent of a mid-sized city. Proximity to transmission infrastructure and signed PPAs for renewable power now dictate site selection above all else.
Fiber density and proximity to internet exchange points (IXPs) matter most for inference latency. NOVA benefits from having the world's highest concentration of IX traffic. Submarine cable landing stations bias where AI APIs get served internationally.
Hyperscale campuses are not office parks. Meta's DeKalb, Illinois facility covers 1.4 million square feet. Google's Papillion, Nebraska campus spans 300+ acres. You need land that is zoned, cleared, flat, available, and expandable. In NOVA, suitable parcels now exceed $1M/acre.
Flood plains, earthquake zones, and hurricane-prone regions get disqualified immediately. Water availability for cooling is now a deal-breaker in Phoenix and Las Vegas — a 100 MW facility can consume 3–5 million gallons per day in warmer climates requiring evaporative cooling.
The Emerging Markets: Where New AI Capacity Is Going
The constrained established markets are pushing real investment to a second tier of geographies. These aren't backup options — they're where the actual gigawatts are getting permitted and built:
Ohio
Microsoft's Columbus-area footprint now exceeds 800 MW contracted. Google has major facilities in New Albany. Ohio combines cheap power ($0.05/kWh average industrial), abundant Great Lakes water, and central US position.
Iowa
Meta's Altoona campus is among the largest in North America. Iowa generates 60%+ of its electricity from wind — ideal for carbon-neutral commitments. Cold winters mean free economizer cooling for 5–6 months/year.
Indiana & Kentucky
AWS has 2+ GW of planned Indiana capacity. Low industrial power rates, pro-business state governments, and access to the Midwest fiber corridor. Kentucky's coal-to-grid transition is creating stranded transmission capacity that data centers are filling.
Wyoming & Montana
Cold air means free economizer cooling for 7–8 months/year. Abundant hydro power. Low land cost — parcels at $3,000–10,000/acre vs. $1M+ in NOVA. Sparse population means minimal opposition.
Texas (with caveats)
ERCOT's deregulated grid makes power cheap during normal hours but volatile during peak demand. The 2021 freeze caused widespread outages. Still: Meta has a 2 GW campus in Eagle Mountain, UT area and Google's San Antonio facilities show real appetite.
International: FLAP-D and Beyond
Europe's data center geography is dominated by the FLAP-D markets — Frankfurt, London, Amsterdam, Paris, and Dublin — each of which is grappling with its own version of the constraints hitting NOVA:
Deutsche Telekom, Equinix building at scale. 1,000+ MW. Germany's energy transition creating some instability.
AWS, Meta, Microsoft EU campuses. EirGrid unable to connect new large loads until 2027+ in some areas.
Paused new permits 2019–2023. Reopened with renewable energy requirements. AMS-IX is world's second largest internet exchange.
Moratorium 2019–2023. Now open with strict sustainability requirements. Southeast Asia hub for US hyperscalers.
Equinix, NTT, AWS. Fiber-dense, high power cost. Primary Japan market for hyperscalers.
AWS, Equinix, AirTrunk (acquired by BlackRock). ANZ market anchor, growing 25%+ annually.
Power and Water Are Reshaping the Map
Loudoun County has run out of power. Dominion Energy's transmission queues have a multi-year backlog for large commercial connections. New data center development in NOVA is heading 10–20 miles south to Prince William County or west toward the Shenandoah Valley — adding latency and land cost.
Phoenix faces a different constraint. Maricopa County is one of the fastest-growing regions in the US while sitting on a depleted Colorado River basin. Intel, TSMC, and now hyperscale AI campuses are all drawing from the same stressed aquifer. Arizona has begun requiring data center developers to document water sourcing before permits are issued.
The economics make the geography shift inevitable. Land in NOVA runs $1M+ per acre for data center parcels. Wyoming land costs $3,000–10,000 per acre. Industrial power in NOVA is $0.04–0.06/kWh; Iowa wind power is $0.02–0.03/kWh via PPA. Over a 20-year data center lifecycle with 200 MW average draw, the geography decision is worth $400–800M in operating costs alone — before land.
The AI infrastructure build isn't going where the headlines say it is.
The real capacity is quietly being permitted in Ohio, Iowa, and Wyoming — where the power is cheap, the water flows, and the grid can actually handle it.
Track real-time AI infrastructure investment and hyperscaler capex at the AI Spending Dashboard. For AI startup valuations tied to this infrastructure wave, see the AI Valuations Dashboard at Value Add VC.