AI’s power appetite just got a much bigger number attached to it. Data centers will consume roughly one-fifth of all electricity generated in the U.S. by 2035, four times what they use today, according to a new BloombergNEF report covered by TechCrunch AI. That’s not a gentle climb. It’s a step change driven almost entirely by the race to build and run AI models.
The headline numbers
BloombergNEF projects U.S. data center capacity will reach nearly 200 gigawatts over the next decade. Here’s how the forecast breaks down:
- Nearly half of that capacity goes to AI training and inference.
- By 2033, the U.S. will host 64% of the world’s AI chips measured by power demand.
- Globally, data centers could add 1,935 terawatt-hours of new electricity demand by 2033 on an aggressive AI adoption path, nearly what all of India uses in a year.
What stands out here is how fast the estimates are moving. BloombergNEF’s new 2035 demand figure is 83% higher than what the same firm predicted back in December. It’s not alone. EPRI, an electrical industry nonprofit, more than doubled its 2024 estimate, and S&P raised its forecast by more than a third between October and April. When multiple independent forecasters all revise up this sharply in a matter of months, that tells you the ground is shifting faster than the models can keep up.
The grid problem is already here
The part practitioners should pay attention to isn’t the 2035 number. It’s what’s happening on the grid right now. BloombergNEF expects most new data centers to connect to networks that are already strained.
Two regions carry an outsized load:
- PJM Interconnection (Virginia to Illinois): 34% of its electricity will go to data centers.
- ERCOT (most of Texas): 22% of its generating capacity devoted to data centers.
PJM is the cautionary tale. It already hosts a large chunk of the country’s data centers and got so overwhelmed by connection requests that it paused new source applications for four years. It reopened the queue to new generators in April, but the strain is real. One utility, American Electric Power, has threatened to pull out of the interconnection entirely. And the supply-demand imbalance pushed electricity prices up 76% over the past year.
Even with that congestion, demand isn’t cooling. Data centers made up 38% of charges in PJM’s most recent capacity auction. Everyone still wants in.
Why this matters for you
If you’re building, buying, or budgeting anything that runs on AI compute, this report is a signal worth acting on.
- Compute costs won’t just track chip prices. Power is becoming the bottleneck, and a 76% jump in one grid’s electricity prices flows straight into what you pay for inference at scale.
- Location is now a strategic decision. Where your provider sits, and which grid it draws from, affects both cost and reliability. Concentrated regions like PJM carry more risk of connection delays and price spikes.
- Capacity planning has a longer lead time than you’d expect. A four-year pause on grid connections means the infrastructure behind your compute can’t just spin up on demand. Factor that into long-term roadmaps.
My read: the AI industry has been talking about chips and models as the scarce resource. The real constraint is turning out to be electrons and the wires that carry them. That reframes a lot of the competitive map, and it explains why hyperscalers are suddenly signing nuclear deals and chasing their own power generation.
The caveat
Forecasts this far out come with wide error bars, and the report itself hints these figures could be conservative given how every revision has trended upward. The global 1,935 terawatt-hour figure assumes AI adoption keeps its aggressive pace, which isn’t guaranteed. Treat the numbers as direction, not precision.
Still, the direction is unmistakable. The question shifting into focus isn’t whether we can build the models. It’s whether the grid can keep the lights on while they run. You can find the full breakdown in the original TechCrunch AI report.