The Grid Was Not Built for What We Are Doing to It
- Rich Washburn
- 1 hour ago
- 8 min read
On Wednesday morning, July 22, a transmission line went down in northern Virginia.
That happens. Transmission lines go down. It is a normal event on an electric grid. Equipment fails, weather interferes, animals get into places they shouldn't. Grid operators deal with this constantly. What happened next is not normal.
The data centers connected to that section of grid did what their control systems are designed to do. They detected the voltage disturbance, disconnected from the grid, and transferred to backup power. Automatically. In milliseconds.
A massive amount of load — the kind of power consumption that runs entire cities — simply vanished from PJM Interconnection, the largest electric grid in the United States. The voltage disturbance propagated outward and could be felt from Washington, D.C., to Chicago.
Dominion Energy, the utility that serves northern Virginia, said its system operations team stabilized the situation within minutes. That is probably true. They are good at their jobs. But "stabilized" is doing a lot of work in that sentence. The grid experienced a sudden, massive loss of demand that sent frequency spiking in the opposite direction from what operators normally manage. The tools for handling this are limited. The recovery was manual. And the data centers stayed off the grid for hours, because reconnection is not automatic. This is not a one-time event. It is a pattern. And the pattern is accelerating.
In July 2024, the same thing happened.
A lightning strike arrestor failed on the Ox-Possum 230kV transmission line near Fairfax, Virginia. Six voltage drops rippled through the system as controls tried to compensate. Sixty data centers in northern Virginia automatically switched to their uninterruptible power supply systems and internal generation.
1,500 megawatts of demand disappeared from the grid in an instant. That is roughly equivalent to a third of all households in Virginia suddenly turning everything off at the same time.
PJM-wide frequency spiked to 60.047 Hz. That sounds small. It is not. The NERC target band is plus or minus 0.036 Hz. The grid was outside its safe operating envelope. Emergency adjustments prevented cascading outages. NERC documented the event. Reuters investigated it.
That was 2024. Sixty data centers. 1,500 megawatts.
Now it is 2026, and it happened again. The Reuters report does not specify the exact megawatt figure this time, but the language — "a massive amount of data center power" — and the geographic extent of the disturbance, from D.C. to Chicago, suggest it was larger. The grid operator and the utility are still analyzing the data.
Here is what should concern you: this is the same failure mode, two years apart, in the same place, getting bigger.
Northern Virginia is home to the largest concentration of data centers in the world. Nearly 600 facilities. They handle an estimated 70 percent of global internet traffic daily. Data centers now consume somewhere between 25 and 40 percent of all electricity used in the state of Virginia, depending on whose numbers you use and how you count.
Dominion Energy projects that peak power demand from data centers in Virginia could reach 13.3 gigawatts by 2038. That is nearly a fivefold increase from current levels.
PJM's 2026 long-term load forecast projects summer peak demand rising from 160 gigawatts in 2025 to 253 gigawatts by 2046 — a 58 percent increase driven primarily by data centers. Summer peak load is projected to grow at an average of 3.6 percent per year over the next decade.
To put that in perspective: for decades, U.S. electricity demand was essentially flat. Energy efficiency improvements offset growth. The grid was sized for a stable world. Now the demand curve looks like a hockey stick, and the blade of the stick is made of GPUs.
The technical problem is not just that data centers use a lot of power. It is that they can stop using power instantly, and in enormous quantities, for reasons that have nothing to do with grid operations. When a power plant trips offline, the grid loses generation. Frequency drops. Operators have decades of experience managing this. They have reserves, demand response programs, and time-tested procedures for bringing generation back online. When a data center disconnects, the opposite happens. The grid loses load. Frequency spikes upward. And the tools for dealing with high frequency are far more limited than the tools for dealing with low frequency.
You can ramp generation up. Arresting a rapid oversupply of power is harder. You can take capacitor banks out of service, which is what Dominion did. You can try to rebalance manually. But you are working against physics and the clock simultaneously.
The 2024 event illustrates the asymmetry. After 1,500 megawatts vanished, PJM's fast-responding frequency regulation resources barely moved. The battery storage that would have been ideal for absorbing the spike barely exists in PJM — roughly 400 megawatts of installed battery capacity, compared to 8 gigawatts in ERCOT and 12 gigawatts in CAISO. PJM had to rely on slower traditional generation resources to rebalance. And the data centers did not come back quickly. UPS systems can activate automatically. Reconnecting to the grid requires manual intervention. In 2024, the data centers stayed offline for hours.
That is hours during which the grid is operating without its largest loads, with generation still running as if those loads were there, with frequency elevated, and with operators doing manual work to prevent the kind of cascading failure that takes down an entire interconnection.
The grid operators know this is a problem. They are saying so publicly.
PJM's CEO, David Mills, has been in the role for about a month. He published a letter describing a "credibility gap" between the need for high prices to incentivize power plant construction and the need to protect consumers from unaffordable bills. He called the current situation "not tenable."
FERC Chairman Laura Swett told PJM's annual membership meeting in May that the organization faces "historically unprecedented demand with historically potentially unprecedented, catastrophic failure." She has suggested that PJM may need to be broken into smaller, more manageable pieces if reform does not work. A senior White House official agreed that a breakup should be considered.
American Electric Power, one of the largest utilities in PJM's territory, has threatened to leave. Pennsylvania has threatened to pull out. Maryland's governor told PJM they "failed to get ahead of it."
In early July, with a heat wave bearing down, the Department of Energy issued emergency orders allowing PJM to curtail data centers with backup generation as a last resort. PJM forecast electricity demand of 166,147 megawatts. The grid was minutes away from rolling blackouts during an earlier heat event, according to reporting from Virginia.
Wholesale power prices on PJM jumped 76 percent in the first quarter of 2026 compared to the same period in 2025. Capacity costs — the payments that ensure enough generation exists to meet peak demand — rose by almost 400 percent.
This is not a story about one bad day. This is a story about a system that was designed for a world that no longer exists.
There is a strange irony operating underneath all of this.
The same AI acceleration that is driving the data center buildout — the one that produced the Codex escape, the Jacobian counterexample, the Glasswing zero-days, the singularity debate — is also driving the physical infrastructure toward its limits. The software is getting smarter faster than the grid is getting bigger. The models are scaling while the transformers, transmission lines, and substations that power them are aging.
We are building intelligence at a pace that the physical world cannot match.
The grid was engineered for a specific kind of demand profile: industrial loads that run steadily, residential loads that follow daily patterns, commercial loads that track business hours. Operators could predict demand, schedule generation, and maintain margins. The system worked because the behavior of the load was understood.
Data centers break that model. They draw enormous, constant power — and then, when something goes wrong on the grid, they vanish. Instantly. All at once. Not because the grid operator told them to, but because their internal systems decided to protect themselves.
The technical term for what grid operators want data centers to do is "ride through." When voltage dips, stay connected. Absorb the disturbance. Help the grid recover. The proposed standard is to ride through voltage dips of up to 10 percent for 1 to 2 seconds.
Data centers do not want to do this. Their equipment is expensive. Their uptime guarantees are contractual. A voltage spike can damage servers. So their control systems disconnect at the first sign of trouble, switching to UPS and backup generators. That is rational behavior for the data center. It is dangerous behavior for the grid.
This is the core tension: the data center's incentive is to protect itself. The grid's need is for the data center to stay connected and help absorb the disturbance. Those incentives are not aligned. And right now, the data center's control systems win every time, because they act in milliseconds and the grid operator has no authority to override them.
The scale of the problem is starting to produce creative responses.
In late June, Tesla, Sunrun, and Renew Home announced a partnership to create virtual power plants across the country. The idea is to use the existing fleet of millions of home batteries and smart thermostats to free up grid capacity during peak demand periods. The program aims to unlock 16 gigawatts of capacity — not by building new power plants, but by coordinating distributed energy resources that already exist.
New Jersey signed the Data Center Fair Share Act in early July, which incentivizes data centers to fund home efficiency upgrades in exchange for priority in utility interconnection queues. Other states are considering similar legislation. These are real responses. The virtual power plant concept is genuinely useful. It can be deployed in months rather than the years it takes to build new generation. And it turns the demand side into a flexible resource instead of a rigid one. But 16 gigawatts of virtual capacity, spread across the entire country, is a fraction of what is needed. PJM alone is projecting demand growth of nearly 100 gigawatts over the next two decades. The VPP program is a bridge, not a destination.
The harder truth is that the grid needs new generation, new transmission, new storage, and new market structures to handle loads that behave in ways the system was never designed to accommodate. All of that takes time. The data centers are not waiting.
I keep coming back to the same framing.
The week before the grid event, the AI world was debating whether we had entered the singularity. A model escaped its sandbox and attacked another company's infrastructure. AI systems solved three open mathematical problems. A security model found ten thousand zero-days. Elon Musk posted four words and the internet lost its mind.
The day after those conversations, the physical infrastructure that makes all of it possible wobbled.
The connection is not incidental. It is structural.
Every capability acceleration in AI — every new model, every breakthrough, every agent deployment — resolves to electricity. Compute requires power. Power requires infrastructure. Infrastructure requires years of permitting, engineering, construction, and capital deployment that cannot be compressed by a software update.
The AI curve is bending upward. The infrastructure curve is not keeping pace. And the gap between them is where the real risk lives.
The grid was not built for what we are doing to it. It was not built for loads that can draw gigawatts and then vanish in milliseconds. It was not built for a single county in Virginia to consume more power than some countries. It was not built for the possibility that the second-largest demand resource on the system might decide, autonomously, to protect itself at the exact moment the grid needs it most. This is not a prediction. It is a description of something that has already happened twice. The question is whether the third time is the one that cascades.
Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital and CIO of Data Power Supply.
Comments