AI-driven data center electricity consumption is on a trajectory to reach 945 terawatt-hours globally by 2030, nearly double today's figures, according to a United Nations University report published this week — a finding that has sharpened debates in Washington and in state capitals from Virginia to Arizona about who pays for the grid upgrades required to keep pace with the artificial intelligence boom, and whether the industry can grow this fast without destabilizing regional power networks.
What 945 Terawatt-Hours Actually Means
Nine hundred and forty-five terawatt-hours is difficult to make intuitive without context. It is roughly equivalent to the combined annual electricity consumption of Pakistan, Bangladesh, and Nigeria — or nearly triple the total electricity use of the 1.3 billion people who live across Sub-Saharan Africa. The UN University report, titled "Environmental Cost of AI's Energy Use: Carbon, Water and Land Footprints," found that accelerated server consumption driven primarily by AI model training and inference is growing at 30 percent annually in a base case scenario, while conventional server electricity growth proceeds at 9 percent per year.
The International Energy Agency independently arrived at the same 945 TWh figure in its own assessment, adding that by 2030, data centers would account for just under 3 percent of total global electricity use — up from less than 1.5 percent today. That share, while seemingly modest, arrives at a moment when electricity systems around the world are already contending with surging demand from electric vehicles, heat pumps, and industrial electrification programs. Grid operators were not planning for this much load this fast.
Northern Virginia at the Center of the Storm
No geography in the United States feels this pressure more acutely than Northern Virginia, which hosts more data center capacity than any other region on earth. Loudoun County alone — known in the industry as "Data Center Alley" — houses more than 300 facilities and accounts for roughly 70 percent of the world's internet traffic routing on a given day. Dominion Energy has filed for regulatory approval for several new high-voltage transmission lines to serve planned expansions, projects now drawing opposition from environmental groups and rural communities concerned about visual impacts and agricultural land use.
"We cannot build the infrastructure fast enough," a utility executive familiar with the mid-Atlantic grid planning process said, declining to be named given ongoing regulatory proceedings. "Every major hyperscaler has interconnection requests in that collectively exceed what we can realistically deliver in the next five years. We are managing a queue, not a plan."
The tension is not unique to Virginia. Georgia, Texas, and the Pacific Northwest are all absorbing large new data center footprints, each with its own grid planning constraints. In the Pacific Northwest, where cheap hydroelectric power once made the region a natural draw for large computing facilities, transmission capacity limitations are creating bottlenecks that prevent new loads from connecting even when generation is technically available.
Carbon, Water, and Land: The Hidden Footprints
The UN report's contribution beyond electricity figures is its attempt to account for the full environmental cost of the AI buildout. Data centers require enormous quantities of water for cooling systems — the report estimates that AI infrastructure could consume between 4.2 and 6.6 billion cubic meters of water annually by 2030. This places significant pressure on water resources in exactly the arid states — Arizona, Nevada, New Mexico — where land is cheap enough to attract large campus developments.
Semiconductor fabrication adds further upstream emissions. Producing a single advanced graphics processing unit generates a carbon footprint roughly equivalent to driving an average American car for several months, according to estimates cited in the report, before the chip has processed a single inference query. The full lifecycle carbon cost of the AI infrastructure buildout, the report argues, is substantially underreported when only operational energy use is measured.
The Policy Response, Slow as It Is
Congress has begun, fitfully, to engage with the energy dimensions of the AI buildout. A bipartisan bill introduced in the Senate in March would require large-scale AI facilities to purchase renewable energy credits equal to their consumption — a requirement that tech industry groups have lobbied against as punitive and likely to push new development to jurisdictions with fewer environmental regulations. The measure has not yet received a committee vote.
Several states have introduced their own disclosure requirements for data center energy use. California's Public Utilities Commission is weighing whether AI-driven load growth should be classified separately from general commercial demand for resource planning purposes. Oregon has enacted a data center siting law that conditions permitting on demonstrated grid-integration plans.
Grid planners from the PJM Interconnection in the mid-Atlantic to the Midcontinent Independent System Operator have now formally flagged data center load growth as one of the top drivers of near-term reliability risk in their regional capacity assessments. The UN report, with its 2030 projection, lands with a clear implication: the decisions being made about permits, transmission lines, and grid integration today will determine whether the AI boom runs into a power wall — or whether the world manages to build its way around one.