Transcript
Introduction
Power quality issues cost industries billions every year, and most of them never get properly diagnosed. If you work in the field, you know the frustration. This podcast is for you. Welcome to Power Quality Decoded, brought to you by Power Monitors Incorporated. Find us at powermonitors.com. Let’s dig in.
Welcome to another Power Monitors white paper webinar. Today we’re gonna be looking at our new white paper on data center power consumption, creatively titled Data Center Power Consumption Two. You may be familiar with our last white paper on data center power consumption, and if you’re not, we’ve got a nice link for you right in there.
But while we were working on this brief little white paper that gives an overview of concepts of data center power usage, the IEEE was also working on a much, much longer paper, it’s 55 pages or so, about power consumption and grid reliability. So I read that so that you don’t have to. And we’re gonna give you a brief overview of a couple points there.
Updated Data Center Energy Consumption Estimates
When we wrote our white paper, the best estimates were that data centers had used 150-ish terawatts in 2023. Based on new Berkeley lab projections, that number was actually short a mere 26 terawatt hours. And the total annual energy consumption just from data centers should reach 325 to 580 terawatt hours by 2028, which is just a crazy amount of power.
Note that Federal Energy Regulatory Commission thinks those are too aggressive. It will only go from 176 to 306 terawatt hours by 2030. So, plenty of time.
In our last white paper we noted that data centers consume enough power to run the average American home for 14 million years, or 14 million American homes for one year. These demand forecasts, while it’s a big giant window, we’re somewhere between, by the end of the decade, looking at between running an American home between 28 to 54 million years.
Bear in mind, there are only about 120 million homes in the United States, which tells us that if we hit the high-end numbers, data centers will be consuming as much power as half of all residences in the United States. Truly, truly incredible growth.
Data Center Investment Trends
Now, I’m not a fortune teller. I can’t tell you where exactly we’re going to land in that massive, massive gap between 306 terawatt hours, 580 terawatt hours, but I can tell you that data centers are being built as quickly as they can put them in the ground. We’ve got a nice little graph there of the investment in data centers. And as you can see, that line is aggressively spiking skyward.
Not slowing down at all. We’ve sort of hit an inflection point of increase back a few years ago, and we don’t see that slowing down anytime soon.
Regional Examples and PJM Forecasts
This IEEE report includes some regional examples, including our home district of PJM, the RTO serving the mid-Atlantic. Includes data center alley and Power Monitors, where we do not use as much power as a data center.
They’ve got some graphs and charts there that they have produced for public consumption with their peak forecasts, which are looking aggressive. You can see their forecasts there for 2024 and 2025 taking a massive spike forward. I don’t think I need to convince anybody any further that data centers are coming, they’re coming quick, and they consume truly tremendous amount of power.
Grid Reliability Concerns
IEEE then went on to discuss some more grid reliability concerns. We mentioned briefly in our last paper that you can put in a data center or an entire street full of data centers, as happens back in my hometown, much faster than you can put in distribution to actually service these centers.
There’s an average of two to three years in the United States for how quickly you can put a data center from buying the land to data center’s totally up and running. Transmission upgrades in the United States run somewhere between five to ten years from the planning of, “Yes, we want to buy this equipment,” service to actually being able to scale that up and have it running in production. So that’s two to three times as long.
Mitigating Power Demand
We’ve highlighted a couple interesting ideas of how people are working on this and how they’re trying to mitigate the massive needs for power. Of course, you may have heard Microsoft is working to try to restart one of the Three Mile Island nuclear turbines. We’ll keep an eye on that, see how that story progresses. Obviously, nuclear power gets complicated getting regulatory approval, but they are pushing forward with a partner.
Over in Pittsburgh, west side of Pennsylvania, working on industrial parks with small natural gas generation facilities, power generation facilities around some high-load industrial parks. So you minimize how much you need to build out distribution. You’re putting a much smaller, and these are 25 megawatts. A number that is certainly decent. If you proposed it to people a century ago, they’d think you were insane.
Co-Location
This is what’s called co-location. There’s a pretty big report and review process going on right now to determine what the regulations are around co-location. That is still ongoing, so we’re not going to delve into that at this time ’cause it’s largely speculative. We’ll see what the government comes back with regarding co-location.
That would also apply to not just something like a natural gas, but if you’re using a major solar installation at one of these big industrial parks, how you would handle… Well, yes, they can generate their own power, but they also need to be grid tied, so you still need to have distribution lines built out to the facility even if the facility is expected to largely provide its own power.
It gets complicated from a regulatory perspective. Very few industries are set up currently to generate their own power independent of major electrical providers. So it’s an emerging field. We’re all excited to see what it looks like.
Data Center Power Usage Characteristics
We did mention in the last white paper, and again, you can see more of how data centers define their power usage and power consumption. Main thing to remember here is that they run at full bore all the time. There’s very little ramping up and ramping down like you would see in a traditional industry or certainly a motor-based industry.
Now, one advantage is that pretty much everything they’ve got has to comply. Electronics are, of course, very sensitive, so most of their equipment actually gives you a pretty clean power profile, but it’s never slowing down.
Black Start Situations
This paper went into the idea of, okay, what happens in a black start situation? Because you are bringing your data center on. If something has happened, your data center’s coming on from nothing to full bore, how quickly does it start up? How does it ramp up? Are you just all on immediately, kicking on demand?
A lot of the IEEE paper talked about coordinating between who owns the data center, who controls the data center, and the utility, saying, “Look, in an emergency event,” let’s say you’ve got a major weather event, and you’re bringing back on the data center, maybe you need to develop together a plan of bringing things back on, kind of a ramping up of the power demand instead of full blast on immediately.
There was a recent event in Northern Virginia where there was a massive one and a half gigawatt load shedding incident where a major data center pretty much went offline immediately, which caused some problems for the upstream generation to suddenly have this load that you’re expecting to be providing for just immediately gone.
And again, IEEE TR-131 goes back and says, “Hey, you need to be talking with your data center companies as they’re building data centers in your area about how they’re coming off, how they’re coming back on. How they’re powering up and powering down impacts the grid around them.”
Power Monitoring at Data Centers
IEEE TR-131 at the end talks about power monitoring devices, saying that many devices are not well set up to measure power quality and power issues at data centers because they only record RMS measurements averages across 12 cycles, which could average out important characteristics. You’ve got real quick switching on these electronics with nonlinear loads, high switching characteristics. You’re gonna miss a lot of it.
Fortunately, here at Power Monitors, we are taking many, many, many samples per cycle. Take a look at your specific model for how quickly we’re sampling. But we offer the ability to do single cycle interval recordings and do waveform captures where you’ll get a nice clear chart of, hey, here’s your exact waveform at the time of this capture.
You get a few cycles before, a few cycles after. So your default, I believe, is, for those waveform captures, is 13 cycles, the cycle that triggered it, three cycles before and nine after. So you’re looking at that window where some of the other guys just average that all to one point. Boom, here you go. You get the waveform.
And if you’re interested in how to work with waveform captures to see that kind of information and get detail, we have a nice handful of links there in the white paper, how to capture waveforms and how to set up thresholds for those to capture what you need to look at.
Wrap-Up
This Data Center Power Consumption 2 has been a very brief overview of a much, much longer report put out by IEEE. If you’re interested in that, definitely go check it out. But be advised, it is over 50 pages of information, so budget your time as you will.
Q&A: How Might Data Centers Affect PQ for Utilities?
We have a little bit more information on that on our first white paper. But data center equipment is often very sensitive to recloser events. Their equipment has very, very fast switchover time. Many of them are backed up by generators, and so during a recloser event, they may get kicked over back onto a generator.
For instance, that one example was a one and a half gigawatt load, that was a recloser event, and the recloser functioned properly and came back on the reclose. Things were okay again. They were ready, but suddenly one and a half gigawatts is just gone. Caused some issues to suddenly have a massive load you’re expecting to be servicing just gone.
Data centers also have nonlinear loads and high switching characteristics, which increase harmonic distortion and cause voltage flicker. While we talked about how data centers are typically on full bore, the AI workloads that we’re seeing increasing, and I don’t think are going anywhere, have more of a swing to them than some of the traditional data center loads where things are just kind of always on very, very flat.
It’s still a tremendous amount of data, and you’re still not seeing the kind of… It’s still not like motor starts, but you could have some more swings than you’re perhaps used to before. Great question.
Closing
If anybody thinks of any other questions, you can of course shoot us an email. And until next time, this has been a Power Monitors webinar, tools you need from people you trust. Thanks everybody. Bye.
