There are two stories about AI and electricity, and they are both true, which is exactly why the argument never resolves. One says AI’s energy appetite is a manageable rounding error on a global grid that already powers billions of air conditioners and cars. The other says AI is spiking people’s power bills and straining the grid to the point of blackout warnings. The reason both camps have data is that they are describing different altitudes of the same problem — and the honest version of this story is the one that holds both numbers at once instead of picking the scarier or the more soothing one.
The global number is smaller than the panic
Start at 30,000 feet, where the reassuring story lives, because it is genuinely well-founded. The International Energy Agency — the most authoritative source there is on this — projects that global electricity consumption from data centres will roughly double to around 945 terawatt-hours by 2030, which would be just under 3 percent of total global electricity consumption. Three percent. For all the apocalyptic framing, data centres are on track to remain a small slice of the world’s power draw.
It gets less dramatic the closer you look. That roughly 530-terawatt-hour increase by 2030, according to the IEA’s own accounting, amounts to only about 8 percent of the total projected growth in electricity demand — less than the growth expected from electric vehicles (838 TWh) or from air conditioning (651 TWh). If you are looking for the single biggest new driver of global electricity demand this decade, it is not AI. It is people cooling their homes. That fact deserves to travel much further than it does, because it punctures the laziest version of the doom narrative.
The local number is brutal
Now descend to ground level, where the alarming story lives, and it turns out to be just as well-founded — because the one thing a global average erases is location. Data centres do not sprinkle themselves evenly across the planet. They cluster, hard, in a handful of regions with cheap land, fiber, and permissive politics. And where they cluster, the abstract 3 percent becomes a very concrete line item on a household bill.
The clearest evidence comes from PJM, the largest electricity market in North America, covering all or part of 13 states. In its annual capacity auction — essentially the price paid to keep enough power available for peak demand — the clearing price soared from $28.92 per megawatt-day for 2024/25 to $329.17 for 2026/27, an increase of roughly tenfold. That is not a typo and not a gentle trend. It is the kind of jump that would have looked like a modeling error a couple of years earlier.
And the cause is not a mystery, because PJM’s own independent market monitor put a number on it: data centres were responsible for 63 percent of that price increase, translating to about $9.3 billion in costs to be recovered from customers across the region in a single year. That money does not come from the tech companies. It comes from everyone’s bill. Reuters reported analysts projecting that residential rates in the PJM region could climb between 30 and 60 percent by 2030, driven largely by exactly this dynamic. For a family in northern Virginia or Ohio, “AI is only 3 percent of global electricity” is cold comfort when their own bill is the one absorbing the buildout next door.
Why the averages lie in both directions
Here is the part worth slowing down on, because it is the actual insight buried under the dueling headlines. A global average is a machine for hiding concentration. When something is spread evenly, the average describes everyone’s experience. When something is clustered — and data centre load is about as clustered as infrastructure gets — the average describes no one’s experience. It is simultaneously true that AI is a small share of world electricity and that AI is the dominant force on specific regional grids, in the same way it can be simultaneously true that a lake is safe to wade in on average and that one end of it is over your head.
This is why the debate is so maddening and so easily weaponized. A tech optimist can quote the IEA’s 3 percent and be completely accurate. A local ratepayer advocate can quote the tenfold auction spike and be completely accurate. They are not contradicting each other; they are standing at different altitudes of the same mountain. The mistake — made constantly, in both directions — is to take the number from your preferred altitude and pretend it settles the argument at every other altitude too.
What the honest version demands
So resist the tidy takes. “AI’s energy use is overblown” is true globally and false in northern Virginia. “AI is breaking the grid” is true in specific markets and false as a planetary claim. The accurate framing is narrower and more useful: AI’s energy footprint is a distribution problem before it is a total problem. The aggregate is manageable; the concentration is not, at least not under current rules that let the costs of serving a handful of enormous customers get spread across millions of ordinary ones.
That reframing also points at where the real fight is, and it is not really about kilowatt-hours. It is about who pays for them. The engineering question — can we generate 945 terawatt-hours — has a boring, affirmative answer. The political question — when a data centre lands in your region and your bill goes up to underwrite its power, is that a cost you agreed to carry — is the one actually being litigated right now in statehouses and utility commissions. The energy story that matters in 2026 is not whether AI uses too much power in the abstract. It is whether the people benefiting from the compute and the people paying for the electricity are the same people. Increasingly, they are not — and no global average will make that gap go away.
Related reading: Everyone’s Adopting AI. Almost Nobody’s Profiting. That’s the Real 2026 Story
