AI data centers are, above all, large electricity customers. The IEA estimates that data centers used about 415 terawatt-hours globally in 2024, around 1.5 percent of world electricity, and projects that figure to reach roughly 945 terawatt-hours by 2030. In the United States, it expects data centers to account for nearly half of electricity demand growth to 2030. [1]

US government-funded research points the same way. Lawrence Berkeley National Laboratory found that US data centers used about 176 terawatt-hours in 2023, 4.4 percent of national electricity, and projected 6.7 to 12 percent by 2028, with growth driven largely by AI servers. [4]

Because a single AI campus can require hundreds of megawatts, the first question in site selection is no longer only where land or customers are, but where that much power can be delivered within a few years. That reframes the map: regions with spare generation and transmission capacity gain, while traditional hubs with congested grids struggle to grow.

Getting connected is often the slowest step. The IEA notes that grid connection queues for both generation and large loads are long and complex, that new transmission lines take four to eight years to build in advanced economies, and that waiting times for transformers and cables have doubled in three years. It estimates that about a fifth of planned data-center projects could be delayed unless grid constraints are addressed. [1]

Texas shows the scale of demand. In August 2026, Utility Dive reported that ERCOT faced about 474 gigawatts of large-load interconnection requests, roughly 90 percent from data centers, and more than five times the grid's record peak demand, and that ERCOT had delayed the first batch study under its new large-load process while the governor called for an audit of data centers' power, water use and incentives. [6]

In Europe, Ember reported in 2025 that new facilities in the main hubs wait seven to ten years on average for a grid connection, and CBRE's June 2026 report said power timelines in Chicago extend into 2032 or later and grid upgrades in Frankfurt and London are delayed until the early 2030s. [7][8]

Some developers respond by bringing their own power. The IEA notes that some operators invest directly in co-located renewables and power purchase agreements, and expects natural gas to be the largest source of electricity for US data centers, while renewables meet nearly half of global additional demand to 2030. [2]

Cooling is a major energy use in its own right. The IEA puts its share at around 7 percent of consumption in efficient hyperscale sites and more than 30 percent in less efficient enterprise data centers. Lawrence Berkeley National Laboratory notes that evaporative systems such as cooling towers raise local concerns about water availability, while liquid cooling paired with dry coolers is designed to minimise water consumption. [3][5]

Local climate and infrastructure therefore matter, and some sites can put waste heat to use: Nebius reports that its Mäntsälä site in Finland supplies up to 72 percent of the town's annual heating demand. Stargate Norway plans closed-loop direct-to-chip liquid cooling and renewable hydropower. [9][10]

Water is increasingly a condition of approval. In Malaysia, Malay Mail reported in September 2026 that the national Data Centre Task Force approves projects only where power and water supply are secured, and that reclaimed wastewater is being supplied for data centers in Johor. [11]

Not every AI workload needs to be near users. Equinix notes that model training is not latency-sensitive and can run in large, often remote facilities with cheaper power and land, whereas inference frequently needs recent data and low latency and so benefits from sites close to users and data sources. [12]

Established hubs grew on connectivity as well as power. Virginia's legislative audit commission, JLARC, attributes Northern Virginia's lead to a strong fibre network, reliable and inexpensive energy, available land and proximity to major customers and government agencies. [13]

Tax policy has long shaped the map. JLARC found that Virginia's sales and use tax exemption for data-center equipment, adopted in 2010, cost the state about 683 million dollars in fiscal year 2023 and is used by about 90 percent of the industry. [13]

Other governments now allocate capacity deliberately. Singapore's IMDA launched a second data-center call for application in December 2025, offering at least 200 megawatts, with more available through green energy pathways. Ireland's regulator requires new data centers to bring matching generation or storage and source most of their electricity from new Irish renewables. [14][15]

The direction of travel is clear: incentives are giving way to conditions. We expect site selection to depend increasingly on whether a developer can demonstrate its own power, water and efficiency plans, not only on tax breaks.

Northern Virginia remains the largest data-center market in the world, according to JLARC, and CBRE's June 2026 report recorded vacancy of just 0.3 percent there and more than 1,100 megawatts of absorption in the first quarter of 2026, alongside continuing power supply challenges. Dallas-Fort Worth, at 1.8 percent vacancy, had 88 percent of space under construction already preleased. [13][8]

Texas also hosts some of the largest single AI campuses. Crusoe's Abilene campus, which serves Oracle Cloud Infrastructure, energised its first buildings within a year of starting construction in June 2024 and is planned to reach eight buildings supporting hundreds of thousands of GPUs. [16]

In the Nordics, Ember expects capacity in Sweden, Denmark and Norway to triple by 2030, and projects such as Stargate Norway target 100,000 GPUs by the end of 2026. In the Gulf, The National reported that the first 200 megawatts of Stargate UAE, part of a planned 5-gigawatt campus in Abu Dhabi involving G42, OpenAI, Oracle, Nvidia, Cisco and SoftBank, was due for completion in the third quarter of 2026. [7][10][17]

In Asia, Singapore remains tightly constrained, with 2 percent vacancy and the highest colocation rates in CBRE's survey, while Malaysia approved RM95.8 billion of data-center and cloud investment in the first half of 2026, with Johor and the Klang Valley as the main hubs. [8][11]

Where data centers are built ultimately shapes where GPU capacity is listed and what it costs. Kovara's data-center and regions views show where the roughly 100 providers it tracks list GPU capacity, and the GPU prices page lets readers compare the same accelerator across locations. Kova can help relate a workload's latency and data-residency needs to candidate regions, and buyers can request a quote to confirm capacity in a specific location.