Why AI Data Centers Could Change America's Power Grid

At 7:56AM on July 22, 2026, something unusual happened in Northern Virginia. More than three gigawatts of electricity demand suddenly disappeared from the power grid.

A power plant hadn't failed. Millions of people hadn't simultaneously turned off their air conditioners. Instead, a transmission-line fault caused numerous data centers in Northern Virginia's massive data-center corridor to transfer to backup power. According to Reuters, the resulting drop exceeded 3 GW—about 3% of the total demand on PJM's system at the time. The disturbance was significant, but the grid didn't collapse. Dominion Energy restored normal conditions within minutes, and PJM reported no lasting reliability impact. (Reuters)

That's what makes the incident so interesting. Sometimes the challenge created by enormous data centers isn't that they need too much electricity. It's that several of them can stop needing it at almost exactly the same time.

And that sent me down a rabbit hole.

Why I Started Looking Into This

Most conversations I've encountered about AI infrastructure focus on a few familiar concerns. How much electricity will AI require? How much water will be needed to cool data centers? Where will these enormous facilities be built? Will farmland or undeveloped land be converted into server campuses?

Those are important questions. But they're mostly questions about how much infrastructure AI requires. I became interested in a slightly different question: What happens when enormous, highly sensitive new electrical loads start behaving differently from the customers America's power grid was originally designed to serve?

That's a more complicated story. To understand why, we first need to talk about what happens when you flip on a light switch.

How Does the Power Grid Actually Work?

Electricity isn't quite like water. When you open a faucet, you're allowing water maintained under pressure in the plumbing system to flow through an opening. It's tempting to imagine electricity similarly: somewhere there's an enormous reservoir of electricity, and plugging something in opens a tiny electrical faucet.

That's not really how the bulk power grid works.

Electricity generation and electricity consumption have to remain balanced essentially in real time. Electricity also doesn't necessarily travel along one predetermined route from a specific power plant to your house. Power flows through the interconnected transmission system according to the physical characteristics of the network. A change in generation or demand in one place can therefore affect conditions elsewhere. Practical Engineering's excellent explanation of the 2003 Northeast blackout is one of the resources that originally helped me understand this distinction. (Practical Engineering)

The EIA's Real-Time Operating Grid provides a great way to actually see this system in action. It tracks electricity demand, forecast demand, generation and interchange across regions and balancing authorities rather than depicting America as one giant power plant connected to millions of outlets. (U.S. Energy Information Administration)

One important indicator of whether generation and demand are balanced is frequency. Most of the North American grid operates around 60 hertz. If generation exceeds demand, frequency tends to rise; if demand exceeds available generation, it tends to fall. Grid operators and automated controls continually work to keep the system stable.

This also helps explain why very hot or very cold weather can create reliability challenges. On a hot afternoon, millions of air conditioners may be running simultaneously while generators and transmission equipment are already working under difficult conditions. During the 2003 blackout, for example, high air-conditioning demand combined with unavailable generation, transmission constraints, equipment problems, inadequate situational awareness and eventually a series of line outages. It wasn't one giant failure—it was a series of conditions interacting inside an interconnected system. (Practical Engineering)

So the useful question isn't simply: Does America produce enough electricity?

It's: Can the grid produce and deliver the right amount of electricity to the right places at the moment it's needed—and respond when that demand suddenly changes?

AI is introducing some extraordinarily large new customers into that balancing act.

Three Terms You Need to Know

Before going further, there are three terms worth separating because discussions about data-center electricity often use them interchangeably.

Capacity describes the potential power scale of a facility. If a company announces plans for a 1 GW data-center campus, that describes how large its potential electrical demand could become. It does not necessarily mean the campus will consume exactly 1 GW every second of every day.

Load is the amount of electrical power something is actually demanding at a particular moment. Your home's load changes throughout the day as appliances, heating, air conditioning and other devices turn on and off. A data center's load can change too.

Energy consumption measures electricity use over time. That's why your electric bill is measured in kilowatt-hours rather than simply kilowatts.

The distinction becomes especially important later in this story:

• kW, MW and GW describe power

• kWh, MWh, GWh and TWh describe energy consumed over time

That means saying a data center has a 100 MW capacity and saying it consumed a certain number of MWh during a day are related statements, but they aren't the same statement.

How Big Is a Data Center?

Megawatts and gigawatts aren't particularly intuitive measurements for most people, so let's start somewhere familiar with your house. According to the U.S. Energy Information Administration, the average American household consumes approximately 10,500 kWh of electricity annually. Spread across all 8,760 hours in a year, that's an average continuous power demand of roughly 1.2 kW. Actual household demand varies enormously throughout the day and by region, home type and weather, but the average gives us a useful scale comparison. (U.S. Energy Information Administration)

The International Energy Agency describes a conventional data center as roughly 10–25 MW, while a hyperscale AI-focused facility can exceed 100 MW. The IEA says a 100 MW AI data center can consume approximately as much electricity annually as 100,000 households, and projects increasingly large facilities reaching into the gigawatt range. (IEA) That gives us a simple way to visualize the scale.

Putting Data-Center Power in Household Terms

Scale	Approximate equivalent average U.S. household demand
Average U.S. home	1 home
10 MW data center	~8,300 homes
100 MW AI/hyperscale facility	~83,000 homes
1 GW AI campus	~830,000 homes
3 GW July 22 load loss	~2.5 million homes

Household equivalents above compare power with the approximately 1.2 kW average continuous demand calculated from EIA's annual household electricity consumption. They illustrate scale; they do not mean data centers literally divert electricity from this number of homes. The IEA's annual-energy comparison uses different assumptions and estimates a 100 MW hyperscale facility at roughly 100,000 households.

Graphic sources: U.S. Energy Information Administration — Electricity Use in Homes; International Energy Agency — Understanding the Energy-AI Nexus; Reuters — July 22 PJM Data-Center Event.

That final comparison is the one that made the July event click for me. When more than 3 GW of data-center demand disappeared from PJM, the grid experienced the rough equivalent of the average electrical demand of about 2.5 million American households disappearing almost simultaneously.

Again, those homes didn't actually disconnect. It's a scale comparison, but it helps turn "3 GW" from an abstract electrical-engineering number into something comprehensible.

AI Is Changing the Scale

Data centers aren't new. Every cloud application, streaming service, website, online banking transaction and enormous corporate database has to run somewhere. What's changing is the scale and density associated with AI computing.

The IEA estimates that data centers represented about 1.5% of global electricity consumption in 2024 and expects global data-center electricity demand to more than double to approximately 945 TWh by 2030. It also notes that AI-focused hyperscale facilities can exceed 100 MW and that the largest projects are moving into multi-gigawatt territory. (IEA) In the United States, the latest Lawrence Berkeley National Laboratory update estimates that data centers could account for 11.8% of total U.S. electricity consumption by 2030, with modeled scenarios ranging from 9.5% to 15.3%. (LBL ETA Publications)

That range is important. Nobody knows exactly how quickly AI electricity demand will grow. Models may become more efficient. New models may become larger. Companies will announce data centers that never get built. Others may expand beyond their original plans. Computing hardware will improve, while AI adoption may accelerate.

That uncertainty creates an infrastructure-planning problem because electrical grids aren't built overnight. Transmission lines, substations, transformers and new generation can require years of planning, permitting and construction. The Belfer Center at Harvard highlights the resulting risk of both insufficient infrastructure and overbuilding around data-center demand that may not materialize as expected.

That raises a question far beyond electrical engineering: if utilities build expensive infrastructure around projected AI demand and some of that demand never arrives, who ultimately bears the cost? (Belfer Center) The data-center operator, the utility, other businesses, or local residents? AI's relationship with the grid is therefore also a forecasting, policy and cost-allocation problem.

The Problem Isn't Just How Much Electricity AI Uses

This brings us back to Northern Virginia. Data centers aren't simply large electrical loads. Their electrical behavior can also be different from conventional loads because facilities use sophisticated uninterruptible power supplies, batteries, power electronics, backup generation and automated protection systems to protect extremely expensive computing equipment.

Those protections exist for good reasons, but when many large facilities respond similarly to the same grid disturbance, individually rational decisions can create a much larger system event.

NERC has been examining exactly this issue. Its Large Loads initiative says rapidly connecting loads such as AI data centers have electrical characteristics that can be distinctive and less predictable than conventional loads, creating new forecasting, planning and reliability challenges. (NERC)

This concern predates July 2026. In a 2025 filing with FERC, NERC discussed an earlier system fault in which approximately 1,550 MW of voltage-sensitive load, including data centers, disconnected without utility action. (NERC) NERC technical materials also document a Dominion event involving roughly 1.5 GW across about 60 data centers, noting that from the data-center perspective, UPS systems were operating as intended while grid operators faced over-frequency and voltage concerns from the collective load drop. (NERC)

That's a crucial distinction. The data center can do exactly what it was designed to do while simultaneously creating a new problem for the larger system.

When Doing the Right Thing Creates a Different Problem

Imagine a restaurant preparing dinner for 100 people. The kitchen knows roughly how many meals it needs to produce. Then, without warning, 30 customers stand up and leave. The restaurant can't instantly un-cook the meals already being prepared.

The electrical grid faces a vastly more sophisticated version of that balancing problem. Immediately before the July 22 disturbance, generators across the interconnected system were supplying the electricity being demanded by customers. Then more than 3 GW of that demand disappeared. For a brief period, generation and demand were no longer where the system expected them to be. Reuters reported a frequency increase and a disturbance lasting roughly ten minutes before conditions normalized. (Reuters)

Ting's sensor network provides another perspective on the event. The company reported detecting voltage and frequency effects across its network of household electrical sensors and argued that the incident demonstrates the importance of understanding how large data-center loads actually respond to disturbances, rather than relying solely on assumed behavior in grid models. See Ting's July 22 sensor analysis

It's important not to overstate what happened. July 22 wasn't an example of America's power grid failing. The system absorbed an unusually large disturbance, operators responded, and conditions stabilized. (Reuters) It demonstrated why planners are paying attention—particularly as individual data-center campuses grow toward the same gigawatt scale as the entire July 22 disturbance.

Grid Planners Saw This Coming

One of the most important things I discovered while researching this story is that July 22 wasn't the first warning. NERC established its Large Loads Task Force in August 2024, and by 2025 it was formally raising large-load reliability concerns with FERC. Its work now addresses not only how much electricity large facilities require, but also forecasting, load modeling, interconnection studies, transmission planning, resource adequacy, operational flexibility and the behavior of large loads during system disturbances. (NERC)

Its May 2026 Risk Mitigation for Emerging Large Loads guideline takes the issue even further, covering areas including model verification and validation, interconnection, long-term planning, resource adequacy and operational constraints. (NERC)

So the question isn't simply whether the grid has enough electricity. It's whether grid operators understand how these enormous new customers will behave.

The Model Says One Thing. The Sensor Says Another.

This part of the research felt surprisingly familiar to me. I've spent part of my marketing career working around manufacturing technology and production-monitoring systems. One of the fundamental ideas behind that technology is simple: A model tells you what a system should be doing. Sensors tell you what it actually did. Both matter.

Grid planners can model how they expect a large data center to respond to an electrical disturbance. But what if its real-world protection settings behave differently? What if several facilities use similar equipment? What if they're all responding to the same disturbance?

What if they disconnect simultaneously? Or reconnect simultaneously?

Suddenly, understanding customer behavior becomes part of operating the grid. That's one reason NERC's guidance places so much emphasis on data collection, model verification and coordination around emerging large loads. (NERC) You can't accurately model behavior you can't see.

Another Problem: What If We Build Too Much?

There's another side of this conversation that has little to do with blackouts. Infrastructure is expensive. If utilities expect enormous new AI loads, they may need new generation, transmission lines, substations, transformers and other equipment. AI development is moving much faster than utility infrastructure planning.

Suppose a technology company announces a massive data-center campus. The utility plans around it. Infrastructure gets approved and construction begins. Then the economics of AI change. Models become dramatically more efficient. The company chooses another location. The project gets downsized. Or it never gets built.

The Belfer Center identifies this mismatch between rapid data-center development and slower grid planning as a significant challenge, including the potential for stranded infrastructure and debates over who should bear the resulting costs. (Belfer Center) That's not merely an electrical-engineering problem. It's a policy problem.

Are AI Data Centers Going to Break the Power Grid?

I think that's the wrong question. AI isn't a giant appliance someone is plugging into America's electrical outlet. The grid itself is constantly evolving, and AI can potentially be used to improve grid operations as well. The IEA identifies applications including improved weather forecasting, transmission monitoring and optimization, and battery research. (IEA)

Meanwhile, utilities and data-center operators have multiple options available: new generation, transmission expansion, battery storage, better grid monitoring, on-site generation, microgrids, more sophisticated interconnection standards and potentially more flexible computing loads.

NERC's work suggests the solution will also require better coordination. Grid planners need more accurate information about how enormous loads will respond to voltage and frequency disturbances, and data-center operators need to protect highly sensitive and expensive equipment. Those aren't inherently opposing goals. (NERC)

In fact, AI data centers may eventually offer something valuable to power grid operators precisely because they're unusual customers. Facilities with enormous batteries, backup generation, sophisticated controls and potentially flexible computing workloads may have considerable ability to control when and how they consume electricity. The question is whether those capabilities are coordinated with the grid around them.

The Bigger AI Conversation

AI conversations tend to revolve around the things we can see: Will AI replace workers? How many tokens did the model use? Can we build repeatable workflows? What will the next model be able to do?

Those are understandable questions because they're the parts of AI we interact with, but every prompt has a physical story behind it. Servers, chips, cooling systems, buildings, transmission lines, transformers, power plants—and people operating an electrical system that has to remain balanced every second of every day.

That's what fascinated me about this rabbit hole. The cloud isn't actually in the clouds - it's physical infrastructure. As AI becomes more integrated into the economy, we're going to have to think about that infrastructure with the same seriousness we're beginning to apply to the software running on top of it.

Professor Sherwood's Practical Wisdom

I don't think the lesson here is that we should stop building AI. That would be a particularly strange conclusion for me to reach while using AI to help organize the research for this article. I think the lesson is about asking better questions—and following them far enough to understand the consequences.

How much electricity will AI use? Scale matters, but timing and location matter too.

Can we generate enough power? Only if we can also deliver it where and when it's needed.

What happens as massive data centers connect to the grid? We also need to understand what happens when several gigawatts suddenly disconnect.

How much infrastructure should we build? Demand forecasts have consequences, especially when someone has to carry the financial risk if those forecasts are wrong.

Data centers and utilities aren't opposing sides of this story. They're interconnected systems. Both need better visibility, better information sharing, and a clearer understanding of how the other will respond when conditions change.

The takeaway isn't that AI is breaking the power grid, it's that AI is changing the assumptions we've used to plan and operate it. That's a less dramatic headline—but a much more useful conversation.

What Happens Next?

The July 22 event didn't knock out the Eastern United States. That's actually why it's worth studying. The system experienced an unusual disturbance. Protection systems activated. Operators responded. The grid recovered. Engineers received another real-world example of how enormous computational loads interact with an electrical system built long before anyone imagined individual computing campuses potentially demanding a gigawatt or more of power. (Reuters)

The 2003 Northeast blackout demonstrated that complex systems don't necessarily fail because one enormous thing breaks. Multiple smaller technical and organizational conditions can interact until the system reaches a state nobody intended. (Practical Engineering) AI gives us an opportunity to apply that lesson while much of the next generation of infrastructure is still being planned. Not by panicking about data centers or pretending their electricity demand doesn't matter. And not by assuming technological progress will automatically solve the infrastructure challenges it creates.

By measuring, modeling, coordinating and building—and by asking better questions before gigawatts of new demand arrive. Or disappear.

Keep Exploring

For readers who want to keep pulling the thread, I'd start with the EIA Real-Time Operating Grid to see how electricity demand, generation and interchange change throughout the day; NERC's Large Loads Action Plan for the reliability questions grid planners are actively examining; and Lawrence Berkeley National Laboratory's 2025 Data Center Energy Usage Update for the latest U.S. demand projections. For the July 22 event specifically, Reuters' independent reporting establishes what happened, while Ting's sensor analysis provides an unusual look at how the disturbance appeared through distributed household electrical sensors.

If you're starting from scratch on how the grid works, I strongly recommend Practical Engineering's explanation of the 2003 Northeast blackout. It's one of the resources that inspired the approach I'm taking with Field Notes in the first place.

Bibliography

• International Energy Agency. (2025). Understanding the energy-AI nexus. In Energy and AI. International Energy Agency

• Mural, R., Pherwani, D., Gupta, C., Yu, Y., Takahashi, A., Kim, D., Majumder, S., Lee, H., Yu, M., & Xie, L. (2026, February 10). AI, data centers, and the U.S. electric grid: A watershed moment. Belfer Center for Science and International Affairs, Harvard Kennedy School. Belfer Center

• North American Electric Reliability Corporation. (2025). Comments of the North American Electric Reliability Corporation regarding large loads. Federal Energy Regulatory Commission. NERC filing

• North American Electric Reliability Corporation. (2026). Large Loads Action Plan. NERC

• North American Electric Reliability Corporation. (2026, May). Reliability guideline: Risk mitigation for emerging large loads. NERC

• Practical Engineering. (2022, February 15). What really happened during the 2003 blackout? Practical Engineering

• Reuters. (2026, July 22). Massive disconnect of power roils largest U.S. electric grid. Reuters

• Smith, S. J., Hubbard, A., Newkirk, A., Ganeshalingam, M., Holecek, B., Sartor, D. A., Mills, M., & Shehabi, A. (2026, June). United States data center energy usage report: 2025 update. Lawrence Berkeley National Laboratory. Lawrence Berkeley National Laboratory

• Ting. (2026). What happened when 3 gigawatt of data center demand disappeared from the grid? Ting

• U.S. Energy Information Administration. (n.d.). Electricity use in homes. U.S. Energy Information Administration

• U.S. Energy Information Administration. (n.d.). Real-Time Operating Grid. U.S. Energy Information Administration

Why AI Data Centers Could Change America's Power Grid