I’ve been involved with the electric power industry since 2008 and there have been a number of scares. Yet, the past two or three weeks marked the first time I’ve seen genuine concern about an existential threat to the grid itself. This concern didn’t originate from a bunch of wild-eyed bloggers like yours truly. It originated from NERC, the organization charged with ensuring the reliability of the North American electrical grid. On May 4, NERC put out just the third Level 3 alert (the highest level) in its history, about this threat.
What was the threat behind this alert? Even if you hadn’t seen the title of this post, you would already know the answer to this question: It’s the rapid (and even more rapidly expanding – i.e., the second derivative is also strongly positive) growth of data centers. However, I’m willing to bet a lot of money (for me, anyway) that you don’t know the exact nature of the threat I’m referring to. You probably think it’s based on one of these two well-documented facts:
1. The immense needs of data centers (and AI data centers in particular) require a huge buildout of new generation and transmission capacity in the grid – mostly paid for by…you’re right!...you and me, Mr./Ms. Rate Payer!
2. Setting aside the question of who will pay for all this, it’s not clear that it’s even possible to build out the required amount of generation and transmission resources as quickly as is needed. This means it’s possible there will be regular brownouts, and maybe even blackouts, in some parts of the country in a few years.
The good news is that, at least on one level, there’s a fairly simple solution to these two problems: BYOP, which stands for Bring Your Own Plant. This means that any new data center must provide its own generation, not rely on someone’s rosy projection about how the US can magically put in place all required generation, without much cost increase, when it’s required. The White House recently secured a BYOP pledge from the major beneficiaries of the new data centers (while overlooking the fact that the usual suspects who took the pledge, including Amazon, Microsoft, Meta and Google, aren’t the ones in a position to keep it. But we can put that concern aside for the moment).
This intuitively makes a lot of sense: At first, it seems there should be a hard and fast rule that no new data centers can be built unless they meet 100% of their own power needs. Moreover, these plants won’t even be attached to the grid; they’ll just be connected to the data center(s) they power. This means their costs will be borne entirely by the operators of the data centers, as well as their customers. What could possibly go wrong with this?
If you read this post, you know the answer: BYOP plants need to start quickly and generate power. Therefore, wind and solar plants need not apply to be BYOP. In fact, clean energy in general need not apply, unless you can find a currently non-working nuclear plant lying around and can spend a billion or so to get it working again, or in case you can say build a huge new dam where there’s a lot of hydro power potential and few people – say, in Northern Ontario – plus pay for the massive transmission line that would be needed to bring that power to your data center in the US.
The only type of plant that makes sense for BYOP is natural gas – and a simple cycle gas plant, not a combined cycle one. The former takes just a few months to build, while the latter takes 2-3 years and is much more expensive. However, simple cycle plants produce much more pollution than combined cycle – both greenhouse gases and air pollutants like nitrogen oxides and carbon monoxide.
Of course, this hasn’t stopped the BYOP movement. After all, the people who make the decisions on BYOP don’t need to worry about pollution, since whatever happens, these plants won’t be built anywhere near where they live. Instead, they’re being built in – surprise! - some of the poorest areas, like the two BYOP plants (one 400 MW and the other over 900 MW) that Elon Musk has built – and is expanding – in Memphis and northern Mississippi respectively. Currently, his AI company is getting away with saying – presumably with a straight face – that these plants aren’t subject either to EPA or NERC/FERC regulation, since they’re not directly connected to the power grid. Thus, even though ratepayers aren’t paying for those plants, the inhabitants of the areas around them will pay an enormous cost in poor health outcomes.
Is it even possible for a BYOP plant to be completely disconnected from the grid? In the longer term, I doubt it. This is because data centers aren’t very tolerant of long outages. If a data center is completely dependent on BYOP power and the plant itself goes down (not just part of it, like one or two turbines), that’s exactly what the data center might get – a long outage. My guess is most BYOP data centers will maintain some connection to the grid, if for no other reason than to serve as a backup to the primary backup, even though the chance of needing that “backup squared” is small.
This “half in/half out” connection to the grid can result in two big problems, which might at first seem contradictory: 1) the plant might not be drawing power from the grid and then suddenly draw a lot of power, and 2) the plant might be drawing a lot of power from the grid and then suddenly not draw any power at all. Problem 1) is a nuisance that can be dealt with through “weapons” currently in the arsenals of grid operators, but Problem 2) is why big data centers now pose a huge threat to the grid – and the solution to this problem is still very much up in the air.
Regarding the first problem, if a large data center with BYOP has been relying entirely on its own plant for power but suddenly has to switch to the grid for all or most of its power (probably because the onsite plant failed), this shouldn’t pose a problem to the grid at all – that is, unless the grid operator (perhaps PJM or SPP) didn’t anticipate that happening and therefore didn’t have enough reserve generation available. And if they didn’t do that, shame on them.
However, what if a data center with BYOP, that has been running on power from the grid for a long time, suddenly drops off the grid? This can happen – and has happened recently, as I discussed in this post - because a protective relay that sits between the data center and the grid senses a disturbance that could damage the equipment in the data center; as a result, it immediately disconnects the data center from the grid, so it relies entirely on the onsite plant.
If just one or two data centers drop off the grid at the same time, that’s an annoyance that can be dealt with by the grid operator. But what if 40 data centers drop at the same time? That’s not an academic question. In February 2025, 40 data centers in Virginia simultaneously dropped off the grid due to failure of a high-voltage line; the same thing happened (also in Virginia, which is currently the Saudi Arabia of data centers) in July 2024, when 70 data centers dropped simultaneously.
In both cases, the grid operator (PJM) acted quickly and averted disaster. However, the total amount of lost demand (aka “load”) from the data centers was less than 2,000 MW. But, as more and more data centers are added to the grid, it’s possible that 4-5,000 MW of data center load might drop at the same time. How will the grid operator deal with so big a drop?
You might wonder why this should be an issue. After all, there have been huge blackouts before, including the 2003 Northeast Blackout, in which over 23,000 MW of load was lost. But there’s a big difference between previous blackouts and what we’re concerned about today. The previous blackouts were almost all caused by deficient power supply in one area, which produced a cascading effect as more and more power plants shut down to protect themselves from the disturbance. As supply kept falling, load (demand) fell even faster due to cascading blackouts. Ultimately, at least in some areas, supply again exceeded the much reduced load and the grid started building back up again.[i]
The current data center buildout has literally flipped the script. Now, the problem is that losing thousands of megawatts of data center load could literally become a self-reinforcing trend, since as the imbalance between supply and demand grows, more and more “large loads” (which includes data centers but also other large power users like steel and aluminum plants) will withdraw from the grid and run from their off-grid backup power.
This means a huge oversupply of power will develop; the imbalance will only be exacerbated by the decreasing load. Since the data centers, each with their own plant, will keep happily running during all of this, this means demand won’t snap back (or even crawl back) to match the oversupply for a long time. Unlike a blackout caused by deficient supply, one caused by deficient demand doesn’t ultimately self-correct. Left to itself, it could lead to a spiral down to a point like what is described in Ted Koppel’s very readable book Lights Out. That book has a number of problems, but it does a great job of describing what could happen if the power grid falls into a ditch and can’t find its way back out again. Spoiler alert: It’s not pretty at all.
Of course, this isn’t an insoluble problem, but the solution requires a lot of thought and planning; that’s what NERC is engaging in now, with its Large Load Action Plan. There’s a 4-minute video available that will give you a very high-level idea of what’s going on.
Tom Alrich’s Blog, too is a reader-supported publication. You can view new posts for two months after they come out by becoming a free subscriber. You can also access all my 1300 existing posts dating back to 2013, as well as support my work, by becoming a paid subscriber for $30 for one year (and if you feel so inclined, you can donate more than that or become a founding subscriber for $100). Whichever option you choose, please subscribe.
If you would like to comment on what you have read here, I would love to hear from you. Please comment in my chat or email me at [email protected].
[i] “Rebuilding” the grid isn’t an automatic process and needs to be done very carefully; this is called “blackstart”. If blackstart isn’t done carefully, rotating generating equipment might be damaged and restoring the gird could literally take months.