Why artificial intelligence could cause the lights to dim

Tom Temin And energy looks at the power consumption, specifically of the data center industry, why?

Arman Shehabi So, actually, Lawrence Berkley Labs and my research team looks at many different sectors in the economy. We look at transportation, we look at residential and commercial buildings and we look at different industrial sectors. Data centers is one of the new industrial sectors that we’ve been looking at, because it’s really part of the emerging economy and new technologies. And it’s an area where over the last 20 years, we’ve seen times where it’s really started to grow. And understanding where that growth is going is has been important to understand what’s causing the growth and what efficiency opportunities there might be.

Tom Temin Right. And what are the trends right now? What have you noticed? Because there’s almost a hockey stick curve in your chart for this year.

Arman Shehabi Yeah. So we’ve done these reports a few times in the past. The way that we do these analyzes takes a long time. It’s a lot of data collection and rigorous analysis because it’s what we call a bottom up model. It’s like uses sort of all estimates based on all the IT equipment that’s going into data centers and then builds up the model that way. So when we did this report back in 2016, we saw that electricity use, which was quite flat in the teens, from 2010 to when we did it in 2016. And it look that way going forward, even though if you think back then, in the teens, the data center industry was just growing like gangbusters, you can think of all that that social media was growing. It was just expanding and being bigger part of the economy. But the electricity use was pretty flat. What you are seeing now, looking back is around 2017, we started seeing a new type of piece of IT equipment coming into the data centers. And what that is, is a server, which is like a computer, except it doesn’t just have a CPU, it has its other pieces of IT equipment in them called graphical processing units (GPUs). And those extra pieces of equipment use a lot of power. And so we saw that the new servers going into data centers are using a lot more power than before, and the number of those servers was increasing really rapidly.

Tom Temin That kind of reverses an historical trend, because early in the age they used what’s called bipolar technology for transistors, which uses a lot of power for the transistor to stay in the on state. And then when they went to, I guess, metal oxide semiconductors, it changed that equation. And they use a lot less than bipolar transistors. It seems like it’s going back to the way it was a long time ago.

Arman Shehabi Well, one thing we’re seeing is, there’s been a general trend of like increased computational efficiency throughout sort of the history of computations of the transistor and of of CPU’s and integrated circuits. But the amount of computation that’s being desired is constantly increasing as well. And so there’s move to this new type of computing which is really used for machine learning and that’s used for AI, for example, is using a lot more computations. And that’s what these new type of servers are providing.

Tom Temin And I imagine then this would at some point back up to the cost for commercial computing, say cloud computing services and whatever services people use from these providers. Do we know what percentage or what portion of the operating expenses of a data center are represented by electricity?

Arman Shehabi Yeah. Well, it’s changing as the electricity use has been growing. Overall it’s the amount of electricity being used is pretty large for the industry and the cost is in the millions of dollars. But when you look at it overall being the entire cost of operating these huge companies, it’s not that large. So it’s not a main consideration at this point, with this growth in AI. Really what we’re seeing, I’ll just add to that is that it’s a bit of an arms race right now with with sort of who’s going to be the first company to have the killer app with AI. And so there is a lot of investment going in from all these companies. The cost is not really the primary sort of concern yet. It’s sort of this is the time when it’s, they’re investing, hoping that they’re going to be able to recoup this later on.

Tom Temin We are speaking with Arman Shehabi. He’s staff scientist at the Lawrence Berkley National Laboratory. And does energy, as a entity, the Energy Department, let’s say, looking at energy trends across the entire economic sector of the United States. Do you envision maybe a mitigation strategy where if demand is low at midnight to 4 a.m. or something, that the parts of the data centers could power down? Or do they do that already?

Arman Shehabi Well, the power for data centers is fairly constant when we think of like how the load, like the power demand changes throughout the day. That’s different for every type of service or building out there. You can think in your own home the power demand is going to increase like in the early morning when you’re getting up, and then it’s going to go down when you go to work. And then it’s going to pick back up when you get home at the end of the day. Data centers, it’s fairly constant throughout those 24 hours with a look, can change a little bit from data center to data center. But because of that, I think there’s one of the things that is being looked at as we’re building out data centers, is there are ways that data centers can be flexible in the amount of power they’re drawing from the grid. And can that actually help us expand out our ability to meet the power need and data centers quicker?

Tom Temin Right. Because some of the large operators of these, like Microsoft, has said, they want to restart nuclear power plants that have been mothballed or invest in technology where you have a nuclear reactor on site just for your data center. That may be another part of the Energy Department. But that’s an emerging trend, also fair to say?

Arman Shehabi Yeah. Right now there’s quick growth in the AI sector and a desire for power and getting power quickly is a top priority. I think it’s important to put this growth in in data center electricity use for AI into context to the overall electricity growth that we’re really going to be having over the next 20 years. It’s important to think that we are trying to shift our economy to more of an electrified economy. And so data centers are already electrified. So that that part’s good. But as we think about like electric vehicles are going to be coming online at some point. Heat pumps for heating or something, that’s going to start happening. We’re on shoring a lot of manufacturing, and a lot of manufacturing of other types of manufacturing is going to be electrified. So there is this large anticipated growth that the Department of Energy is expecting. And data centers really are kind of the first wave of that growth. And and it’s really an opportunity now for Department of Energy and sort of broadly for everyone to think about how do we meet not just this immediate demand for data center electricity use, but how do we build that demand for electricity use while thinking about the additional demand that we’re going to be needing in the future? So this investment in new types of of clean baseload electricity use is something that is good to see.

Tom Temin And we’re talking not just kilowatts or gigawatts, but terawatts at this point.

Arman Shehabi Yeah. When we think about how much electricity use like the whole country uses, right currently we consume about 4000 terawatt hours of electricity a week every year. Data centers are a small part of that, but it’s increasing. So in the past, they’ve been about 1% of that. Currently, what we’re seeing now, based on our recent report is that it’s gone from about 1.5% to about 4.5%. And looking to the future, just in the next few years, we see it increasing again in the range of roughly seven to 12% of the electricity use. And that’s considering other demand that’s going to be increasing as well.

Tom Temin And does this report include these Bitcoin mining operations, which are another kind of data center and people don’t like them because they’re noisy and so forth, but it’s kilowatts gigawatts, terawatts, maybe essentially chasing nothing but operating all the time.

Arman Shehabi Yeah. It’s important to note that the growth that we put in this report and the numbers that I’ve been talking about this morning and that we see in sort of our main figures that have kind of been going around the internet, those are not including Bitcoin. But we do have a section in our report. We sort of treat Bitcoin separately because it’s kind of a different kind of data center because they just use different kind of equipment. So just from a definitional point of view. You treat them differently. But yeah, that’s another addition of electricity use on top of the growth that we’re seeing in AI.

Tom Temin All right. So who should read this report and what should the government do about it, if anything?

Arman Shehabi Well, I think really the important thing that we’re doing here, there’s so little understood about what electricity and the amount of electricity that’s actually consumed in the data center sector. It’s just something that’s kind of like behind the scenes that’s not understood very well. So this report really what it does is give a lay of the land and to understand how much electricity is actually being consumed, what’s roughly expected in the future, what’s driving that electricity use. And all of that is something that can be used to identify where there are efficiency opportunities and start building out strategy for building out new electricity generation, new transmission and where and how data should be built and operated.