In July 2026, the North American Electric Reliability Corporation issued its highest-level alert. An AI data center had triggered a 1,800-megawatt drop on the regional grid. That is the output of a nuclear reactor vanishing in minutes.
The same month, PJM Interconnection — the grid operator serving 65 million people across 13 states and Washington, D.C. — held its capacity auction for power starting June 2028. The auction failed for the third consecutive time to secure enough electricity to guarantee reliability. The shortfall was 6.8 gigawatts, roughly the output of seven nuclear reactors.
The price hit the auction cap of $325 per megawatt-day. Without the cap, it would have cleared at $554.72. PJM ratepayers are on the hook for nearly $30 billion across recent auctions, with data centers accounting for roughly $6.3 billion of that total.
The AI revolution has a power problem. It is the problem nobody in the pitch decks talks about, because the pitch decks are about software. The constraint is hardware — specifically, the kind that generates electricity.
How Much Power AI Actually Uses
The numbers are staggering once you dimensionalize them.
The International Energy Agency projects global data center electricity consumption will roughly double from 485 terawatt-hours in 2025 to 950 TWh by 2030. The U.S. Department of Energy’s Lawrence Berkeley National Laboratory found that data centers consumed 4.4% of total U.S. electricity in 2023 and projects that figure to reach between 6.7% and 12% by 2028.
To put that in perspective: if data centers were a country, they would be the fifth-largest energy consumer on Earth, sitting between Japan and Russia.
A single Nvidia Blackwell chip draws up to two kilowatts. That is more than the average American house uses at peak load. A hyperscale AI data center packs thousands of them into server racks, drawing 50 to 100 megawatts continuously. The newest AI-focused facilities are being designed for up to 5,000 megawatts — the output of five nuclear reactors, dedicated to one building complex.
S&P Global’s 451 Research forecasts U.S. data center grid load at 61.8 gigawatts in 2025, rising to 75.8 GW in 2026 and roughly tripling by 2030. NERC’s 10-year peak demand forecast jumped 24% in a single revision cycle, driven almost entirely by data centers. Data centers are now expected to account for more than half of all forecast U.S. load growth over the next five years.
The grid was built for a world where electricity demand grew 1 to 2% per year. It is now growing at 22% per year in the data center segment alone.
Why the Grid Cannot Keep Up
The grid is a physical infrastructure problem. You cannot deploy more megawatts with a firmware update. Building a transmission line takes 7 to 10 years from proposal to energization. A new natural gas plant takes 3 to 5 years. A new nuclear plant takes a decade or more, if it finishes at all.
I have watched enough infrastructure projects to know that the timeline always wins. You can accelerate software. You cannot accelerate concrete.
Virginia’s Data Center Alley — the densest concentration of data centers on Earth — has a grid interconnection queue running 4 to 7 years. Texas processes connections in 12 to 18 months, which is why Texas and Louisiana are leading 2026 construction starts. The 33% of hyperscalers planning on-site gas turbines or small nuclear reactors are doing it because they cannot afford to wait in the queue.
Power is the primary bottleneck, responsible for roughly 50% of all data center project delays. The global data center construction pipeline stands at $2.3 trillion. Thirty-five gigawatts are under construction in North America, with 92% already pre-leased. The demand is there, but the electricity supply is not.
The fossil fuel reality is uncomfortable. As of 2024, natural gas accounts for more than 40% of U.S. data center electricity. Fossil fuels collectively supply about 56% of the sector’s power. New demand growth is being met by gas peaker plants. The AI boom, at the infrastructure level, is currently a fossil fuel boom. Virginia alone permitted 27 gigawatts of diesel generator capacity by end-2025 — the equivalent output needed to power 20 million American homes. During a June 2025 heat wave, operators ran their diesel fleets for hours to reduce grid load during peak demand-response events.
The Nuclear Bet
Every major hyperscaler signed a nuclear deal in 2024 or 2025.
Microsoft signed a 20-year agreement with Constellation Energy to restart Three Mile Island Unit 1 — the reactor that sits on the same site as the 1979 partial meltdown, though a different unit. The restart costs $1.6 billion. Target: online by 2028, generating 835 megawatts.
Google partnered with Kairos Power for up to 500 megawatts of small modular reactors, starting with a 50 MW reactor targeted for 2030 in Tennessee.
Amazon invested over $20 billion in a data center campus adjacent to the Susquehanna nuclear station in Pennsylvania, while backing X-energy’s SMR program for up to 5 gigawatts across multiple future projects.
Meta signed a 20-year deal for 1.1 gigawatts from Illinois’ Clinton nuclear plant and issued requests for proposals for an additional 1 to 4 gigawatts of new nuclear capacity.
Collectively, Big Tech signed contracts for more than 10 gigawatts of U.S. nuclear capacity in roughly 18 months. That is a structural shift. The companies building the AI infrastructure have concluded that the grid cannot give them what they need, so they are building their own power plants.
The nuclear bet faces a timing tension that Harvard Business Review flagged plainly: small modular reactor deployments target 2030 to 2035, but AI data center demand is accelerating now. Nuclear is the right long-term answer running on the wrong timeline for the near-term problem.
The Historical Parallel
In the early 1900s, the automobile industry faced a constraint that had nothing to do with engines or design. The constraint was roads. There were almost no paved roads in the United States. The car existed. The infrastructure to use it did not.
The Federal Aid Road Act of 1916 committed federal money to road construction for the first time. The Federal-Aid Highway Act of 1956 created the Interstate System. Between those two dates, the auto industry grew, stalled, grew again, and eventually transformed the country. The companies that survived the gap between invention and infrastructure were the ones that could last long enough for the roads to catch up.
AI is in the same gap. The models work, the demand is real, but the infrastructure to power them at scale does not exist yet, and the timeline to build it is measured in years, not quarters.
The parallel extends to electrification itself. In the 1930s, the Rural Electrification Act brought electricity to farms that private utilities had refused to serve because the return on investment was too low. The government had to step in because the market alone would not build the infrastructure fast enough to serve everyone who needed it. The PJM auction failures and NERC alerts suggest we are approaching a similar inflection point — where the market alone cannot build power infrastructure fast enough to serve the AI buildout.
What This Means
The AI story has two halves. The first half is about intelligence — models, chips, software, applications. That half gets all the attention.
The second half is about power — megawatts, grid capacity, nuclear restarts, transmission lines, cooling water. That half gets almost none. But it is the half that determines whether the first half actually works.
When a wave of investment analysts start pitching energy stocks — geothermal, nuclear, critical minerals, grid infrastructure — they are pointing at the same structural fact from different angles. The AI buildout has hit a physical constraint. The constraint is electricity. The companies that solve it, or the companies positioned to profit from the solution, are where the next chapter of the AI story gets interesting.
The data center pipeline is $2.3 trillion. The grid cannot serve it at current capacity. Something has to give — and that something is the energy infrastructure.
Read the grid auction results. Watch the capacity shortfalls. Power is the constraint that determines whether the AI buildout succeeds, and the grid is telling you it cannot keep up.