A Heat Wave Just Pushed 100 Million People's Power Grids to the Brink. AI Data Centers Aren't Helping.

Late September isn't supposed to be when the power grid gets nervous. But this week, temperatures running more than 20 degrees Fahrenheit above normal across the Midcontinent grid region, with triple-digit heat forecast across the South through the weekend, have pushed power systems serving more than 100 million people across the United States and Canada into what grid operators are calling a genuinely tight operating window — arriving just as equipment scheduled for routine fall maintenance was still needed online. It isn't just a weather story, and grid analysts covering it keep circling back to a second cause layered directly on top of the heat: the electricity appetite of AI data centers, which has grown large enough on its own to reshape how regional power markets behave.

Exponential Appetite: The Projections Confronting Reality

The scale of that appetite is no longer a rounding error in anyone's forecast. 451 Research, part of S&P Global, projects that US data-center grid demand will climb to 75.8 gigawatts in 2026 and to 134.4 gigawatts by 2030. The Electric Power Research Institute's most recent modeling suggests data centers could account for as much as 17% of total US electricity generation by 2030 — more than double current levels, and 60% higher than the same institute's own estimate just two years earlier. Globally, the International Energy Agency expects data-center electricity demand to more than double to 945 terawatt-hours by 2030, a figure roughly equivalent to Japan's entire annual electricity consumption, coming from a sector that barely registered as a distinct line item in national energy planning a decade ago.

Grid Bottlenecks: PJM Spikes, Texas Pauses, and Fault Cascades

What makes this particular moment worth pausing on is where the strain and the growth are actually colliding. In the PJM market alone, which serves 65 million people from Illinois to North Carolina, data-center demand accounted for an estimated $9.3 billion of the increase in the 2025–2026 capacity market, with residential bills in parts of Ohio and Maryland already climbing by $16 to $18 a month as a direct consequence. Texas has taken the more dramatic step of pausing new grid connections for large-load facilities entirely, after regulators discovered that data centers made up roughly 90% of the 474 gigawatts of new capacity sitting in ERCOT's interconnection queue — a number large enough that state officials openly questioned how much of it represented real, financeable demand rather than speculative reservations. And it isn't only a demand-and-price story. The North American Electric Reliability Corporation flagged something more technical and more alarming in July: several 2025 incidents where more than 1,000 megawatts of data-center load disconnected suddenly during transmission faults, a failure mode grid engineers didn't have to worry about at this scale before AI-optimized racks, which can draw 30 to 110 kilowatts against the 5 to 15 kilowatts of a traditional server rack, became common.

Nuclear Power and the Timeline Mismatch

The industry's own answer to this collision has increasingly become nuclear power, and not for the reasons the marketing usually leads with. Baseload reliability, not carbon accounting, is what a hyperscaler actually needs from a power source that has to run at 99.999% uptime, and nuclear is one of the few options that can promise that without the intermittency problem solar and wind still carry. That's part of why interest in small modular reactors has picked up sharply among the same companies racing to build data-center capacity, and why utilities near existing nuclear plants are fielding an unusual volume of inquiries from operators looking to lock in direct power-purchase agreements years in advance.

The Real Ceiling on Artificial Intelligence

None of this resolves the underlying tension, though. Grid capacity, water availability for cooling, and transmission infrastructure all move on multi-year timelines measured in permits and construction schedules. AI compute demand is moving on a timeline measured in product cycles. Every forecast in this piece assumes the demand curve keeps climbing roughly as projected, and every one of the fixes on the table — new gas turbines, new transmission lines, new reactors — takes years longer to build than the compute clusters they're meant to power. Power, not chips and not capital, has quietly become the actual bottleneck determining how fast the AI buildout can grow anywhere in the world. This week's heat wave was a preview of what that bottleneck looks like under ordinary seasonal stress. It wasn't a worst case. It was a Tuesday in September.