AI data centers may use 4x more electricity by 2035 — and that already touches bills on some grids

BloombergNEF now projects that U.S. data centers could consume about one-fifth of the electricity generated in the country by 2035 — roughly four times today's share. That is not a startup press release. It is a revised demand forecast that jumped sharply in under a year.

We read TechCrunch's July 21, 2026 report on the BloombergNEF outlook, Latitude Media's same-day write-up of the capacity numbers, and TechCrunch's earlier May coverage of Monitoring Analytics' report on PJM wholesale power prices.

If you use ChatGPT, Claude, or Gemini the way most people do — a few times a week for writing, planning, or explaining something — this does not mean your next prompt will raise your bill tonight. It does mean the physical cost of "free-feeling" AI is becoming a utility and grid story, not only a Silicon Valley one.

The source type here is independent research, filtered through news coverage. Forecasts are not meters. BNEF is projecting capacity and demand under an aggressive build-out path; co-author comments in trade coverage note that U.S. construction is hard, interconnection is slow, and some pipelines may never finish. Still, the direction of the revisions matters: other groups have also been raising their numbers. BNEF's own 2035 electricity-demand estimate is 83% higher than what it published in December.

What the July outlook says, in plain terms: AI compute is expected to push U.S. data-center capacity toward roughly 194–200 gigawatts over the next decade. Nearly half of that capacity is tied to training and inference — the two big jobs of building models and then answering your prompts. By 2033, the U.S. is still expected to host most AI chips by power demand. Globally, aggressive AI adoption could create on the order of 1,935 terawatt-hours of new data-center electricity demand — roughly in the same ballpark as India's annual electricity use.

Where this stops being abstract is regional grids that are already tight. BloombergNEF expects many new centers to land on strained systems. In its framing, PJM — the big interconnection from Virginia through much of the mid-Atlantic and Midwest — could see about 34% of its electricity go to data centers; ERCOT in Texas, about 22%. That matters because PJM is not a theoretical map. It already hosts a dense cluster of data centers, especially in Northern Virginia. Earlier in 2026, Monitoring Analytics, PJM's independent market monitor, reported that the total cost of wholesale power in the region rose about 76% year over year in the first quarter — from $77.78 to $136.53 per megawatt-hour — and pointed at data-center load and the grid operator's struggle to keep supply and demand aligned. Wholesale is not the same as the line item on your residential bill, but wholesale strain is how pressure begins.

The significance assessment, measured against this reader's life: this is more significant than a typical AI news week, and for a different reason than the industry scoreboard. You do not need to become an energy analyst. You do need an honest mental model: every chatbot reply runs in a building full of chips. When those buildings cluster faster than power plants and transmission can catch up, households and small businesses on the same grid can feel it — first as politics and rate cases, then sometimes as higher bills. Practical AI choices still matter at the margin, but the big levers are policy, siting, and who pays for new capacity. Individual thrift alone will not solve a 4x demand story.

One honest reaction: part of me finds the scale impressive. Another part is quietly irritated that the industry sold weightlessness for a decade and is only now having a public conversation about concrete, water, and kilowatt-hours. The irritation is useful. It keeps the story attached to real life.

This is the situation as of August 1, 2026. Forecasts move; interconnection queues and rate decisions move faster in some states than others. Treat any single 2035 number as a directional signal, not a fixed appointment.

If you live in a region with heavy data-center growth — Virginia, parts of the Midwest, Texas, and a growing list of other states — the next time your utility or local paper mentions rate pressure, AI infrastructure is a fair question to ask about, not a conspiracy theory. Before you assume your ChatGPT habit is the villain on next month's bill, separate three things: household usage, wholesale market stress in your grid region, and who regulators make pay for new load. The honest frame for everyday AI use is: the tool is light on your desk and heavy somewhere else.

The cloud has an address, a substation, and a power bill. Pretending otherwise was always the spin.