Grid Talk - Watt's New in Power | Sep 2026 Edition

Abhinav Gupta

Sep 7

11 min read

Updated: Sep 18

A big thank you to all the folks who have stuck with GridTalk over the years and to everyone who signed up after the DTech edition in February (we are now at 850+ subscribers which still blows my mind). In February I promised that the next edition would take less than two years and here we are at a little under seven months, so I am going to call that progress before anyone does the math on what “monthly newsletter” was supposed to mean :D

Before we jump in, Happy Labour Day to all my Canadian readers and Happy Labor Day to the folks south of the border (yes, I am spelling it both ways and no, I am not picking a side). I also wanted to give a shout out to the folks who did not get the long weekend off, i.e., the control room operators, lineworkers and substation crews across North America who work to keep the lights on for all of us. A special note on the eastern side of the continent, PJM called a Pre-Emergency Load Management action on September 1st in five zones with a preliminary peak of 152,518 MW and had Maximum Generation and Load Management alerts out for the 3rd (details -> here).

One more quick update, Grid Transformation is a sponsor at GridFWD 2026 in Snowbird, Utah on September 29th and 30th (details -> here) and the theme, “Reprioritizing Affordability”, could not be timelier as you will see below. We are also sponsoring the gondola ride, so if you are attending, come find me on the way up the mountain (or on the way down, I am not picky) and let’s talk shop.

A Tale of Two Gigawatts (Part 1)

For the uninitiated, the title is a play on ‘A Tale of Two Cities’, the 1859 novel by Charles Dickens, the 19th century English novelist (born in Portsmouth, spent most of his life in and around London) who set the story between London and Paris in the years leading up to the French Revolution and opened it with probably the most famous first line in the English language, “It was the best of times, it was the worst of times”. Let me walk you through why this made sense to me as the title for this edition.

Similar to Dickens, our industry is living through two very different stories at the same time. On one side, there is more money flowing into the electric utility business than at any point in my career (the big tech companies alone spent more than $400 billion on capital expenditure in 2025 and the IEA expects that number to grow another 75% this year!) and the phone at every utility, ISO and consulting firm (including ours) does not stop ringing. On the other side, the largest grid operator on the continent just ran a capacity auction that came up 6,831 MW short of what it needs and on the morning of July 22nd, roughly 3,800 MW of data center load in Northern Virginia dropped off the grid because of a fault that the system cleared exactly the way it was designed to. Hence the two gigawatts in the title, the one that shows up in every load forecast and the one that vanishes when a data center does not like the voltage it is seeing.

This topic ended up being way too big for one edition (my first draft was twice the length of anything I have published before), so I have split it into two parts. This one covers the load itself and what it looks like from the control room, Part 2 will cover who is paying for all of this and whether these loads can actually be flexible, and it will drop in about a month or so, or whenever I get the next edition out, whichever comes first :D

How much load are we actually talking about?

For folks who have been around since Edition #1 (February 2024, link at the bottom), you will remember that I described data centers as “one small sliver” of global electricity demand, about 240-340 TWh in 2022 or 1-1.3% of the total, that was going to grow exponentially with the new AI chips. Well, two and a half years later the IEA puts global data center consumption at 485 TWh for 2025, which is a 17% jump in a single year with the AI-focused facilities growing 50%, and they expect it to roughly double again to 950 TWh by 2030 (report -> here). Closer to home, the Lawrence Berkeley National Lab put out its 2025 update in June and estimates that US data centers used about 4.7% of all the electricity in the country in 2024 and could be at 649 TWh or 11.8% by 2030, with a range of 521 to 843 TWh (report -> here). Now, there are a lot of different numbers floating around on the internet (and every single one of them will turn out to be wrong in some way) but they are all in the same zip code and the direction is not in question.

So, what does one of these campuses actually look like from the grid’s point of view? A “1 GW data center” is 1 GW of IT load, i.e., the servers, upon servers racked up in those Walmart-sized buildings, and once you add the cooling, the UPS and the power conversion on top (a PUE of 1.2 is a reasonable number for a new build) you are at 1.2 GW at the point of interconnection, which at a 90% load factor works out to about 9.5 TWh a year or roughly the annual consumption of 875,000 average American homes. That is one customer at one substation and utilities are getting asked for these in bunches.

Which brings me to the part that took me a while to get my head around. ERCOT has more than 438,000 MW of large load connection requests sitting in front of it, roughly 89% of them data centers (here), on a grid that just set an all-time peak record of 91.1 GW in July! The amount of large load that has actually been energized and observed on the ERCOT system as of April was 5.9 GW (here). Closer to home, the IESO said in August that it has roughly 7,000 MW of data center related connection requests, double what it reported in March, in a province that peaks at about 25,000 MW (here). So, the lesson every planner has learned the hard way over the last couple of years is that a request is not a forecast, ERCOT literally cut new data center demand in its 2025 forecast to about half of what was requested and PJM now only counts the near term loads that have signed something and put money down.

What a gigawatt looks like from the control room

Now, this is the part of the newsletter that is closest to my heart as an OT person and it is the part that I think most folks outside our industry do not appreciate yet, a data center is not a factory. NERC’s Large Loads Task Force put out a white paper last summer (here) that walks through what these facilities look like electrically, 60-95% of the load is IT equipment and almost all of it sits behind uninterruptible power supplies and power electronics that can decide in milliseconds whether they want to stay connected to the grid or not. The paper includes a recorded example of a North American data center ramping down from about 450 MW to about 40 MW in 36 seconds (yes, 36 seconds!) and the AI training clusters are worse, the same paper measured a training block changing its demand at a rate of almost twice its rating per second as the GPUs synchronize and save their progress. There is a joint paper from Microsoft, OpenAI and NVIDIA (here) that goes into this in a lot more detail, the power of a single training data center can swing by tens of megawatts a few times a second, which happens to be right around the natural frequencies of turbine-generator shafts and long transmission lines, and their fix for now is to make the GPUs burn power while they wait for each other, which costs about 10% extra energy. I will admit that when I first read that a room full of chips could interact with the torsional modes of a steam turbine somewhere else on the grid, I had to read it twice.

And then there is the ride-through problem. On July 10th, 2024, a lightning arrester failed on a 230 kV line in Northern Virginia and the line reclosed a few times (six faults in 82 seconds), about 1,500 MW of data center load responded by transferring itself to backup generation, most of it did not come back for hours and the operators had to pull capacitor banks to get the voltages back down (NERC’s incident review -> here). Then on the morning of July 22nd this year, a mechanical failure on a 230 kV line caused a normally cleared fault, i.e., the protection did exactly what it is supposed to do, and approximately 3,800 MW of data center load dropped off in two waves, 2,970 MW and then 1,099 MW, and it took the operators nine minutes to get the area control error back within limits (PJM’s write up -> here). The chair of PJM’s Operating Committee summed it up in one line, “This was a normally cleared fault. They should not disconnect from the grid.” For the folks who lived through the inverter ride-through story over the last decade, this should all sound very familiar, only this time it is the load that is dropping off instead of the generation.

The good news is that 2026 is the year the rule book started to catch up. NERC found through its Level 2 alert last fall that roughly three quarters of the 33 GW of operating data centers it looked at were represented in planning models as generic loads that cannot reproduce any of the behavior above (here), FERC has since ordered NERC to file mandatory reliability standards for computational loads by December 31st of this year (here) and ERCOT’s ride-through requirement for computational loads above 75 MW went into effect on August 1st (here). Here at home, the IESO published version 3 of its technical requirements for large computational loads on September 3rd (here) and the AESO published a similar guide in June. If you have spent any time around ADMS and DERMS implementations you know that the platform is only as good as the network model underneath it, and the same is now true at the transmission level, the load model in the EMS has become as important as the generator model and most of them are not there yet.

Now, the obvious next question is who is paying for all of this and whether these loads can bend when the grid needs them to, and that is Part 2. To give you a preview, PJM’s market monitor thinks data center load is responsible for about $29 billion of the capacity charges over the last four auctions, Ontario and Alberta both rewrote their connection rules over the summer (no freebies and curtail the data centers first, respectively) and a study of 155,000 GPUs in production found that only about 6% of a facility’s power can actually be shed on short notice. There is a lot to unpack there and I did not want to do it in 400 words.

So, if I had to sum up where we are in one sentence, it would be this, two years ago the industry was arguing about whether this load was even coming and today it is arguing about ride-through curves, minimum bills and who gets curtailed first, which is a much better argument to be having. For us at Grid Transformation, this is the same story we have watched play out with ADMS and DERMS over the last decade, the technology shows up faster than the operating model and the hard problems are addressed over time; the load models in the EMS and ADMS, the PMU data nobody is looking at yet and the LDC being asked to connect a 100 MW customer on a feeder built for 20, and that is exactly the space we have been building towards. I would like to end on a positive note, the first set of NERC standards for these loads will be on FERC’s desk by the end of this year, which is about as fast as I have ever seen a standards process move.

“To infinity and beyond!”

Folks who know the reference, feel free to comment below with the name of the movie. Coming back to my personal learnings, this summer has been all about animals again with my daughter, we spent a day at Reptilia (Canada’s largest reptile zoo -> here) and she has not stopped talking about it since. So, here are my two fun facts for this edition, both courtesy of the 5 and a half year old.

1. Whether an alligator hatches as a boy or a girl is decided by the temperature of the nest, 31 °C or below produces females, 32 °C produces mostly males and 32.5 °C and above goes back to mostly females (here). Who knew?

2. Snakes do not have eyelids, so they cannot blink and they sleep with their eyes open, and their most acute sense is smell, which they do with their tongue (here).

And finally, on a lighter note, in keeping with the title, I have picked up A Tale of Two Cities again for the first time in a very long time. It is a very different book when you are older and, given what Part 2 is about, Dickens had a few things to say about who pays for what too.

For those who have been around from the early days of GridTalk, thank you so much for sticking with me and for all the new subscribers, welcome aboard! My humble request to folks who are subscribed is that if you find this newsletter useful or entertaining in any way, then please feel free to repost to your network and engage with me by liking and commenting on the post. As always, feel free to reach out to me here or via email abhinav@grid-transformation.com and follow Grid Transformation to keep up to speed with what we are doing and the latest in the industry as a whole.

Previous editions: Edition #1 | Edition #2 | Edition #3

GridTalk will be back with Part 2 (and a recap from GridFWD) in about a month. Till then, be safe and stay curious!

Grid
Transformation

Grid Talk - Watt's New in Power | Sep 2026 Edition

A Tale of Two Gigawatts (Part 1)

How much load are we actually talking about?

What a gigawatt looks like from the control room

Top Stories from Grid & Tech

“To infinity and beyond!”

Recent Posts

Comments

Let’s transform the grid together with solutions for the future.