The Real Cost of AI: How Data Centers Are Consuming More Power Than Entire Countries
Peak of Trending
Global data center electricity consumption is set to reach 565 terawatt-hours in 2026, according to Gartner's June 2026 forecast, a jump large enough that the increase alone rivals what mid-sized nations use in a year. AI-optimized servers are the reason: their draw grew 84 percent in a single year and is on pace to overtake every conventional server on the planet by 2027. The cost is not staying inside the data hall. In Virginia, wholesale power prices spiked past $1,000 per megawatt-hour this July, and residents in a county with 37 data centers were asked by their own government to turn off the lights. This piece tracks where that electricity is going, who is paying the difference, and which numbers in the public conversation about AI's energy appetite hold up and which don't.
- The Number That Changed the Argument
- What a Single AI Query Actually Costs
- Where the Electrons Go Once They Reach the Rack
- The Bill Is Already Landing on Your Doorstep: Virginia's Grid Under Strain
- The Nuclear Pivot: Big Tech Becomes a Power Company
- The Gas Turbines Nobody Voted For
- Who This Reporting Is For
- The Verdict
The Number That Changed the Argument
Global data center electricity use will hit 565 terawatt-hours in 2026, up 26 percent from 447 TWh the year before, according to Gartner's data center energy forecast. That single-year increase is roughly the annual electricity consumption of a country the size of Portugal, materializing inside twelve months. Worldwide power demand tied to these facilities is projected to reach 132 gigawatts this year, up from 104 GW in 2025, and Gartner expects it to hit 290 GW by 2030.
AI-optimized servers are doing almost all of the pulling. They consumed about 95 TWh in 2025 and will draw 175 TWh this year — an 84 percent jump — while conventional servers grew barely 1 percent. Gartner puts the crossover point in 2027, the year AI hardware starts consuming more electricity than every other server in every data center combined. The International Energy Agency's own base case, published in its Energy and AI analysis, projects global data center consumption doubling to roughly 945 TWh by 2030, with accelerated servers growing 30 percent annually against 9 percent for everything else.
What makes this period distinct from prior tech buildouts is the timeline mismatch. A data center can go from groundbreaking to operational in two to three years. A transmission line or a new gas plant takes longer, often much longer, and that gap between switch-flip demand and shovel-in-the-ground supply is where every downstream problem in this article originates.
The number nobody uses: cooling systems alone are forecast to consume 195 TWh in 2026, a 22.6 percent jump in a single year — more electricity than Sweden's entire population uses annually, spent purely on keeping chips from overheating.
What a Single AI Query Actually Costs
Three separate figures circulate for the energy cost of one ChatGPT query, and they disagree by a factor of ten. The oldest, from researcher Alex de Vries in 2023, put it near 3 watt-hours. OpenAI's Sam Altman stated 0.34 Wh in a 2025 blog post, offering no methodology behind the figure. Epoch AI's independent 2025 analysis landed close to Altman's number, at roughly 0.3 Wh for a typical GPT-4o query, and attributed the gap with the older estimate to genuinely more efficient hardware plus an inflated token-count assumption in the original math.
Both numbers can be true at once, which is the part most coverage skips. Epoch AI's own researchers note that reasoning-heavy queries and longer outputs push consumption far past the simple-query baseline — University of Rhode Island researchers separately estimated GPT-5 averaging 18.35 Wh per query and spiking to 40 Wh for complex responses, according to reporting compiling both figures. A simple factual lookup and a multi-step reasoning task are not the same transaction, and treating "AI query energy" as one number flattens a genuine tenfold spread.
This is where the research pulled the argument somewhere unexpected. The per-query panic is mostly noise. Even at the high end of 3 Wh, a single query costs less energy than running a microwave for ten seconds. The real number is the multiplier: OpenAI reports roughly 2.5 billion queries a day, and at even the conservative 0.3 Wh figure, that is 750 megawatt-hours daily before a single training run is counted. Scale, not the individual prompt, is what shows up on a grid operator's forecast.
Where the Electrons Go Once They Reach the Rack
Servers account for roughly 60 percent of a data center's electricity draw, with cooling and environmental control taking up to 30 percent in less efficient facilities and closer to 7 percent in the best-run hyperscale sites, per IEA's infrastructure breakdown. Networking and storage split the remainder. That gap between 7 and 30 percent is not a rounding error — it is the entire argument for liquid cooling, and it explains why Nvidia's Dion Harris describes efficiency work inside data centers as, in his words, a constant search for every remaining unit of power that can be squeezed out.
Rack density is the other half of the story, and it has moved faster than most facilities were built to handle. A traditional enterprise rack draws 10 to 15 kilowatts. Nvidia's GB200 NVL72 rack draws 120 to 140 kW, and the industry is bracing for chips that push individual racks toward a full megawatt — enough to power roughly 750 average U.S. homes from a single server cabinet, according to Bloomberg's reporting on data center redesigns. Nvidia itself estimates that around 30 percent of the power flowing into a modern AI data center never touches a chip doing AI work; it is lost to cooling overhead and long-distance transmission across sprawling campuses.
You are three sentences into troubleshooting a stalled model deployment, the dashboard shows normal utilization, and the actual bottleneck turns out to be the facility's power delivery, not the code. That scenario is now common enough that Gartner names power availability, not chip supply, as the binding constraint on AI expansion in 2026.
"The current supply of capacity in PJM is not adequate to meet the demand from large data center loads and will not be adequate in the foreseeable future."
The Bill Is Already Landing on Your Doorstep: Virginia's Grid Under Strain
PJM Interconnection's capacity auction price rose 833 percent between the 2024/25 and 2025/26 market cycles, according to the Institute for Energy Economics and Financial Analysis. The following year's auction climbed another 22 percent and hit the federally approved price cap. Wholesale power across the entire 13-state PJM territory rose 75.5 percent year-over-year in the first quarter of 2026, according to Monitoring Analytics, the grid's independent market monitor, which stated plainly in its report that the price impacts on customers have been very large and are not reversible.
Northern Virginia's Dominion zone, home to the densest cluster of data centers on Earth, felt this first and hardest. Summer peak load there hit 23,905 megawatts in 2025, 23 percent above 2019 levels, and winter peak load rose 45 percent over the same span, per U.S. Energy Information Administration data. On July 2, 2026, PJM's total grid demand was expected to hit a record 166.3 gigawatts as spot power prices in Northern Virginia briefly spiked past $1,000 per megawatt-hour during a heat wave.
Henrico County, Virginia, spent years courting data center operators with tax breaks and fast permitting. It now hosts 37 operating facilities with 17 more approved. On June 26, 2026, the county manager emailed thousands of government employees asking them to turn off the lights, after electricity rates for county buildings jumped 24.9 percent overnight — a request the manager compared to Great Recession-era austerity measures, as reported by a local account of the county's electricity crisis.
Data centers now consume roughly 40 percent of Virginia's total electricity, up from under 5 percent in 2010. Nearly three-quarters of Virginia voters surveyed in January 2026 blamed data centers for their rising bills. The mechanism is a familiar one dressed in new clothes: someone has to pay for new transmission lines and peaker plants, and utility rate structures were never built to isolate the cost of a small number of enormous customers from everyone else on the line.
The Failure Advocates Skip: Cost Allocation
Virginia's State Corporation Commission approved a new rate class in November 2025 requiring large-load customers — meaning data centers — to commit to 14-year contracts and cover at least 85 percent of transmission costs regardless of whether they ever use the full capacity they reserve. The measure exists because the alternative already happened: ordinary ratepayers across Maryland, Ohio, D.C., and New Jersey absorbed part of a $9.3 billion capacity cost increase tied directly to data center load growth, according to IEEFA's analysis of the 2025/26 auction. Six other states have introduced construction moratoriums. Oregon became the first to create a dedicated data center rate class specifically to stop this cost-shifting from repeating.
The Nuclear Pivot: Big Tech Becomes a Power Company
Faced with interconnection queues stretching past five years in some regions, the largest AI operators stopped waiting for utilities and started buying power plants directly. Thirteen nuclear deals across the four major hyperscalers now commit more than 9.8 gigawatts of capacity, according to a tracker compiling every announced agreement. Microsoft signed a 20-year, $16 billion power purchase agreement with Constellation Energy to restart Three Mile Island Unit 1, the reactor now renamed the Crane Clean Energy Center, with full power expected by 2027 after FERC approved a transmission waiver in June 2026 that removed the last permitting obstacle. Amazon has committed more than $20 billion to nuclear-powered data centers in Pennsylvania, including a 1.92 gigawatt power purchase agreement tied to the Susquehanna nuclear plant.
Meta leads on raw capacity, with up to 6.6 gigawatts committed across TerraPower's Natrium reactor design, Oklo's Aurora small modular reactors, Vistra, and Constellation. Google went the small-modular-reactor route with Kairos Power, committing to 500 megawatts across six or seven reactors that won't come online until 2030 to 2035 — a timeline that undercuts the idea that nuclear power solves this decade's demand growth.
Here is the part that deepened under scrutiny rather than resolving cleanly: nuclear looked, at first pass, like the clean answer to the grid problem. It is low-carbon, dispatchable, and unlike solar or wind, runs at the constant load AI training clusters actually need. But the earliest of these reactors will not deliver a single electron before 2027 at best, and Kairos Power's SMR fleet won't finish until the mid-2030s. The nuclear pivot is a bet on the back half of this decade. It does almost nothing about the demand spike happening right now, which is exactly why the next section exists.
[No relevant internal article found]
The Gas Turbines Nobody Voted For
While nuclear gets the press releases, natural gas is doing the actual load-bearing work of the 2026 buildout. Data center developers have announced roughly 101 gigawatts of behind-the-meter gas generation in the United States, with more than 57 gigawatts already tied to disclosed equipment orders, according to RBC Capital Markets' May 2026 strategy note. Close to 80 percent of planned behind-the-meter projects use natural gas technology, per a Rabobank analysis. Microsoft's contract with Nscale for an 8-gigawatt gas microgrid in West Virginia is projected to raise the company's annual emissions by 40 percent on its own, according to research firm Cleanview.
The consequences are not abstract for the people living next to these installations. In Memphis, Tennessee, more than 30 natural gas turbines were installed at xAI's Colossus data center in a historically Black neighborhood already facing elevated asthma rates; local residents and the NAACP filed notice of intent to sue under the Clean Air Act, according to World Resources Institute reporting on community health impacts. In Southaven, Mississippi, 27 gas turbines at another xAI facility run around the clock rather than only during emergencies, producing continuous noise that off-grid diesel backup generators — used solely for outages — do not, since those emergency units emit 200 to 600 times more nitrogen oxides than a natural gas plant when they do run.
Nobody in this buildout campaigned on Clean Air Act lawsuits. That is the unspoken part of the AI infrastructure story that rarely makes it into an earnings call.
The scale keeps compounding. Global Energy Monitor tracks more than 1,000 gigawatts of gas-fired power now in development worldwide, a 31 percent jump in a single year, with the United States responsible for about a quarter of that pipeline and more than a third of new U.S. gas capacity earmarked specifically for data centers, per Grist's investigation into the gas buildout. Two-thirds of U.S. gas project developers have not yet identified who will manufacture the turbines they've ordered — a supply chain gap that mirrors the same disconnect between announced capacity and delivered power running through the nuclear deals.
More than 75 data center projects worth $130 billion were blocked or delayed in the first months of 2026 alone, driven by community opposition over power and water costs — a scale of local resistance the industry did not face two years ago, according to Data Center Watch figures cited in coverage of the AI power crisis.
Who This Reporting Is For
- Anyone in a PJM, ERCOT, or similarly affected utility territory who wants to understand why their bill jumped and whether a data center in their region is part of the reason.
- Investors evaluating hyperscaler capital expenditure who need to weigh nuclear timelines against near-term gas dependency rather than taking sustainability pledges at face value.
- Local officials and planning boards weighing tax incentives against the cost-shifting risk that Virginia, Oregon, and half a dozen other states are now trying to legislate away after the fact.
- Readers who want a grounded answer to "how much energy does my AI use" that accounts for the real spread between a simple query and a reasoning-heavy one, instead of a single viral number.
This is not an argument for abandoning AI infrastructure, and it is not a brief for building it anywhere, anyhow, without scrutiny. It is a case for reading capital expenditure announcements and clean-energy pledges with the same skepticism now being applied by state regulators, and for recognizing that the electricity bill for a chatbot response is currently being paid, in part, by a retiree in Maryland who has never opened ChatGPT.
The Verdict
The 565 TWh figure for 2026 is not a projection to argue with — it is close to locked in, given the capital already deployed. What remains genuinely unresolved is who absorbs the cost of the gap between when demand arrives and when clean generation catches up. Nuclear deals answer that question for 2030 and beyond. Gas turbines are answering it right now, in Memphis, in Southaven, and in the queue of 101 gigawatts of behind-the-meter capacity racing to fill a hole the grid cannot fill fast enough on its own. Track the ratio between announced nuclear capacity and delivered nuclear capacity over the next eighteen months; that gap, not the aggregate TWh figure, is where this story's next chapter gets written.
Nobody involved in this buildout — not the hyperscalers, not PJM, not the state regulators scrambling to write new rate classes — has actually solved the underlying mismatch between how fast a data center can be built and how slowly a power grid can grow to meet it.
Frequently Asked Questions
How much electricity do AI data centers use in 2026?
Global data center electricity consumption is projected to reach 565 terawatt-hours in 2026, a 26 percent increase over 2025, according to Gartner. AI-optimized servers account for 175 TWh of that total, up 84 percent from the prior year, and are on pace to surpass conventional servers in total consumption by 2027.
How much energy does a single ChatGPT query use?
Estimates range from roughly 0.3 to 0.34 watt-hours for a typical GPT-4o query, per Epoch AI and OpenAI's own disclosure, though older estimates put it near 3 Wh. Reasoning-heavy queries on newer models can use 18 to 40 Wh, so the honest answer depends heavily on which model and query type is being measured.
Why are data center electricity prices rising in Virginia?
PJM's capacity auction prices rose 833 percent between the 2024/25 and 2025/26 cycles, driven largely by data center load concentrated in Virginia's Dominion zone. Utilities are building new transmission and generation to serve that load, and those infrastructure costs are currently spread across all ratepayers in the region, not just the data centers causing the growth.
Are tech companies building nuclear power plants for AI?
Yes. Microsoft, Google, Amazon, and Meta have signed 13 nuclear deals committing over 9.8 gigawatts of capacity, including Microsoft's restart of Three Mile Island Unit 1 and Meta's 6.6 gigawatts across four separate reactor developers. Most of this capacity will not come online before 2027, and some small modular reactor projects are not expected until the early 2030s.
What are the downsides of gas turbines at data centers?
Behind-the-meter gas turbines run continuously rather than only during emergencies, producing sustained noise and air pollution in surrounding communities. Facilities in Memphis and Southaven, both in historically underserved neighborhoods, have prompted Clean Air Act legal action and NAACP involvement over asthma and air quality concerns.
Does AI use more energy than a Google search?
The commonly cited "10 times more than a Google search" claim traces to a 2023 estimate that current research considers outdated. Epoch AI's 2025 analysis puts a typical AI query at roughly the same energy cost as a modern Google search, though complex reasoning queries remain substantially more energy-intensive.
Will data center energy use keep growing after 2026?
Yes. The International Energy Agency's base case projects global data center consumption doubling to about 945 TWh by 2030, growing roughly 15 percent annually overall, with AI-driven accelerated server demand growing around 30 percent per year over the same period.