Infrastructure Intelligence

A data center is a bet on geography. The power contract, the cooling architecture, the grid it connects to: these decisions lock in cost structure for a decade or more. This page maps the factors that determine where intelligence gets built, who controls the power underneath it, and where the next concentration clusters are forming right now.

Framework page · estimates from public sources · market structure lens, not an asset registry

$0.04–$0.28
Power price range per kWh across US grid regions
1.58 vs 1.10
Industry avg PUE vs best-in-class · 43% effective cost gap
$25B+
Capital committed to Indiana data centers in 2026 alone

The Portability Asymmetry

Silicon
Globally portable
A B200 can be shipped to any facility on the planet. Hardware supply chains are global. The asset moves.
Power
Geographically fixed
You cannot relocate a hydro basin, a wind corridor, or a nuclear plant. Power procurement is a decade-long commitment to a specific geography. The asset does not move.
As hardware cost per token declines (NVIDIA's stated objective with each generation) power becomes a larger share of the total compute cost structure. The providers who solved power procurement early hold a structural advantage that compounds precisely because silicon depreciates and power infrastructure does not. Geography is not a secondary consideration. It is the primary one.

Site Selection Framework

Ten factors determine where a data center gets built.

Factor Signal Best Case Constraint Risk ⓘ
Power availability Interconnection queue 5–7 yrs in constrained markets Stranded renewable · nuclear PPA High
Power cost Largest long-run controllable input · PPAs lock in advantage <$0.04/kWh High
Water access Cooling at hyperscale · millions of gallons annually Abundant · cold climate Medium
Fiber density Latency and bandwidth · legacy hubs compound Existing backbone node Low
Land cost & availability Hyperscaler campuses require 300–500 MW footprint Rural adjacent to grid Medium
Tax incentives States compete · abatement deals shift project economics Full abatement + grants Low
Latency to demand Inference is latency-sensitive · training is not <20ms to major metros Medium
Climate Cold climates enable free air cooling · eliminates chiller plants Nordic · Pacific Northwest Low
Political stability Long-lived assets require decade-plus horizon underwriting Tier 1 OECD jurisdictions Medium
Disaster risk Seismic · flood · storm exposure drives insurance cost Low seismic · inland · elevated High in coastal markets

US Grid Regions & Power Story

The US power grid is not one market. Seven major ISOs/RTOs operate independently with different price structures, fuel mixes, and capacity constraints. Where your data center sits determines your commodity exposure.

ISO / RTO Geography Avg Price Fuel Mix Data Center Significance
PJM Mid-Atlantic · Midwest
13 states + DC
$0.08–0.12/kWh Gas · Nuclear · Coal Largest ISO by load. Northern Virginia (Ashburn) is the world's largest data center market. Heavy nuclear base load. Capacity market exists.
ERCOT Texas
Isolated grid
$0.04–0.08/kWh Gas · Wind · Solar Lowest average prices. Most volatile. No capacity market means price spikes during stress events (2021 winter storm). Fastest-growing data center market. Abundant wind in West Texas.
CAISO California
90% of state load
$0.18–0.28/kWh Solar · Gas · Hydro Most expensive major ISO. Renewable-heavy but intermittent. High regulatory burden. Silicon Valley legacy concentration but new builds increasingly moving to cheaper markets.
MISO Midwest · South
15 states
$0.05–0.08/kWh Wind · Gas · Coal Abundant cheap wind generation. Emerging data center destination. Chicago corridor benefits from fiber density and moderate power costs.
SPP Central US
14 states
$0.04–0.07/kWh Wind · Gas Among cheapest power in the US. Wind-heavy. Underbuilt fiber historically but improving. Emerging for stranded energy compute arbitrage.
NYISO New York
Full state
$0.12–0.20/kWh Gas · Nuclear · Hydro Expensive. Strong nuclear base (Constellation). Financial services proximity drives some demand. Not a growth market for new large-scale builds.
ISO-NE New England
6 states
$0.14–0.22/kWh Gas · Nuclear · Renewables High cost, constrained gas pipeline infrastructure. Limited growth potential for power-intensive workloads. Existing facilities serve financial and academic demand.

The Efficiency Stack

Two data centers with identical power contracts can have materially different effective costs. The difference is operational efficiency: how much of the power bill actually reaches the GPU.

Technology What It Does PUE Impact Cost
Free air cooling Uses ambient outdoor air directly for cooling. Eliminates or reduces mechanical chiller plants. Only viable in cold climates (Nordic, Pacific Northwest, high altitude). PUE 1.05–1.15 Near zero opex
Liquid cooling (direct-to-chip) Coolant circulates directly across GPU heat spreaders. Required for H100+ power densities. Eliminates hot-aisle/cold-aisle air management overhead. PUE 1.03–1.10 High capex
Evaporative cooling Uses water evaporation to cool condenser coils. More efficient than mechanical chillers. High water consumption, problematic in drought-prone regions. PUE 1.15–1.30 Low capex
Air-cooled (traditional) Raised floor or overhead air distribution. Cheap to build, inefficient to operate. Insufficient for modern GPU power densities without significant infrastructure investment. PUE 1.40–1.80 Lowest capex
Immersion cooling Servers submerged in dielectric fluid. Theoretical best-in-class efficiency. High operational complexity, limited vendor ecosystem, not yet at hyperscale. PUE 1.02–1.05 Very high capex
Waste heat recovery Captures exhaust heat for district heating or industrial processes. Converts a cost center into a revenue stream. Viable in dense urban areas with heat demand (Nordic cities). Effective PUE <1.0 Site-specific
Nebius operates Finnish data centers on Nordic hydro with free air cooling and a reported PUE near 1.10, among the lowest in the industry. The industry average is approximately 1.58 (Uptime Institute, 2023). That gap means a provider at industry average PUE pays 43% more in effective power cost than Nebius for the same GPU workload, before any difference in power price.

Known Concentration Clusters

Major compute concentration follows infrastructure decisions made years or decades ago. These clusters compound. Each new facility lowers the cost of the next.

Region Grid Why Here Power Story Constraint
Northern Virginia
Ashburn corridor
PJM MAE-East fiber exchange (1992) created incumbent advantage. Government proximity. Dense peering. Nuclear base · $0.07–0.09/kWh Power constrained · long waits
Dallas / Fort Worth ERCOT Low power cost, no state income tax, central US geography, large labor market. Gas + Wind · $0.05–0.07/kWh Grid volatility risk
Phoenix / Mesa WECC Low land cost, tax incentives, stable geology. Amazon, Microsoft, Google all have major presence. Gas + Solar · $0.07–0.10/kWh Water scarcity · drought risk
Portland / Hillsboro WECC Cheap Columbia River hydro. Cool climate enables free air cooling. Intel manufacturing presence. Hydro dominant · $0.05–0.07/kWh Grid capacity limits growth
Chicago MISO / PJM Major fiber hub. Financial services demand. Lake Michigan cooling. Midwest wind access. Gas + Nuclear · $0.07–0.09/kWh Limited · room to grow
Nordic Region
Finland · Sweden · Norway
NordPool Cheapest renewable power in the world. Free air cooling year-round. Political stability. Waste heat revenue potential. Hydro + Wind · $0.03–0.05/kWh Latency to US demand
Middle East
UAE · Saudi Arabia
Isolated grids Sovereign AI ambition driving massive investment. Microsoft, Google, AWS all announced major commitments 2024–2025. Gas dominant · subsidized Heat · water · geopolitical

Where Intelligence Is Being Built Next

The US Midwest is becoming the new data center belt. The geographic shift is already underway, driven by cheap land, available power, state tax incentives, and the constraints choking existing clusters. Three tiers: announced commitments, emerging pipelines, and locations where the site selection factors converge but major operators haven't yet arrived.

Announced · Active Builds or Committed Capital
Location Grid Operators Scale Power Story
Texas · Abilene · Milam · Shackelford ERCOT Stargate (OpenAI · Oracle · SoftBank) 1.2 GW operational · 9 GW by 2029 Cheap ERCOT power · wind abundant · grid volatility risk · water constrained
Wisconsin · Mount Pleasant · Port Washington MISO Microsoft ($13B+) · Stargate Multi-GW planned Former Foxconn site. Cold climate. MISO wind access. Fairwater AI campus online early 2026.
Indiana · New Carlisle · Jeffersonville · Fort Wayne MISO Amazon ($15B) · Meta ($10B) · Google $25B+ committed MISO grid. Cheap wind. Central geography. Low land cost. All three under active construction.
Ohio · Lordstown · Sunbury · New Albany PJM Amazon ($10B+) · Stargate · Microsoft · Meta · Google 80+ facilities projected PJM grid with nuclear base. Central US fiber routes. State actively recruiting. Stargate Lordstown active.
Michigan · Saline Township MISO Stargate · Hyperscale Data Early stage Adjacent to Indiana belt. MISO grid. Hyperscale Data acquired 48.5 acres southwest Michigan.
Tennessee · Lebanon TVA Meta ($10B+) 1,500 acres · 13 buildings TVA power (nuclear + hydro mix). Broke ground February 2026. Largest Meta campus announced to date.
Wyoming · Cheyenne WECC Microsoft 3,200 acres · April 2026 Cold climate. Wind power. No state income tax. Tripling existing Microsoft footprint.
New Mexico · Doña Ana County WECC Stargate Active build Solar abundant. Cheap land. National lab proximity (Sandia, Los Alamos). Water scarcity risk.
Quebec · Montréal · Québec City Hydro-Québec Microsoft ($7.5B CAD) · AWS · Google 600 MW pipeline Among cheapest clean power in North America. Cold climate. ⚠ Bill 69 (2024): government approval required for new grid connections >5 MW · regulatory overhang.
Emerging · In Motion, No Major Permits Yet
Location Grid Status Why It's Moving
Scotland · 17 identified sites National Grid UK In planning system · no permits granted 2–3 GW of proposed capacity identified by researchers. Scottish government actively promoting. Renewable dominant grid, cold climate, political stability. Constraint: latency to US, planning system slow.
Utah · Box Elder County WECC County vote pending · 2026 Hyperscale proposal with Shark Tank investor backing. Cheap land. WECC grid. Still in local approval process.
Strategically Underserved · Framework-Based · No Major Announcements
The following locations are identified by applying the site selection framework above, not by operator announcements. They are observations, not predictions.
Location Grid Factors That Converge Primary Constraint
SPP Corridor
Kansas · Oklahoma · Nebraska
SPP Cheapest wind power in the US. Abundant land. Low disaster risk. Central geography. Underbuilt fiber · limited labor market
West Texas · Permian Basin
Beyond existing Stargate footprint
ERCOT Stranded solar and wind at scale. ERCOT grid. Massive land availability. Low cost. Water scarcity · dry cooling required

We built this framework into an interactive map. Current clusters, announced builds, and the locations where the structural conditions converge but operators haven't yet arrived.

View the Map