Energy Future: Powering Tomorrow’s Cleaner World

Power Grab: AI and the Electric Grid - Part 1

Peter Kelly-Detwiler Episode 23
Prepare to be shocked by the truth behind the AI-driven data center boom and its colossal impact on our electricity landscape. With predictions suggesting a staggering leap in power consumption by 2030, this episode promises to unpack the complexities of this rapid development. From exclusive insights gathered from the Electric Power Research Institute and various utility companies, you'll gain a comprehensive understanding of why this exponential demand for data center hookups is blindsiding even seasoned industry insiders. Journey with us through pivotal regions like Virginia and Texas, where the grid is under unprecedented pressure, and explore the surprising predictions from McKinsey that could redefine our expectations.

Join our expert guests as we explore the implications of this "power grab" for sustainability goals, our communities, and, inevitably, our electricity bills. As we dissect how AI is at the heart of this transformation, we'll visit the extraordinary developments in Ohio and consider what the future holds for energy consumption in the digital age. Discover why this surge in data center demand has significant climate ramifications, and why it might just be the biggest game-changer in recent memory. This episode, the first in our multi-part series, sets the stage for understanding the monumental shifts underway and the challenges that lie ahead.

Support the show

Speaker 1:

The past few weeks, every second story seems to be about AI and data center load. The topic has certainly dominated my recent weekly videos, but since we're mostly getting isolated data points from the media, perhaps it's time to construct the big picture. This week, instead of a summary of top stories, we're going to go deep in the first of a multi-part series on the 800-pound gorilla. That's rather abruptly changed the electricity supply and demand landscape AI and data centers. It's not that electric vehicle load or power demand from Bitcoin doesn't matter. It's just that the potential future demand from AI-driven data centers is so much larger. How big, that remains to be seen. But let's consider some recently available numbers. In May, the Electric Power Research Institute, epri, published a report saying that data centers might consume up to 9% of US electricity generation by 2030. Epri then followed that up with a May to July survey of 26 utilities, of which 60% said they had requests for new hookups of 500 megawatts or larger. 48% of those utilities indicated they had requests exceeding 1,000 megawatts and almost half said that current data center requests exceeded 50% of current peak demand. This load has come out of nowhere. None of the 26 utilities surveyed had current data center connections over 500 megawatts. Then there's the one utility sitting in the Virginia data epicenter, dominion. Virginia is the world's data hub because of the high-speed fiber backbone located there. Google estimates 70% of the world's internet traffic goes through northern Virginia and in 2022, dominion served about 2,800 megawatts of data center load. That was before the AI boom. That number has since soared to 4,000 megawatts by the end of last year and Dominion will add another 1,000 megawatts or more this year, with total consumption in Virginia now equaling one quarter of the entire state's consumption. 50,000 megawatts of additional data centers are waiting in line, but many won't get built. Dominion now says it won't service new interconnection requests for data centers over 100 megawatts for seven years.

Speaker 1:

Meanwhile, in Texas, encore faces 59,000 megawatts of data center connection requests. For context, the Texas grid saw a record peak of 85,000 megawatts this past summer. Then there's AEP Ohio, in the middle of a tussle with data center giants there. As I noted last week, there's quite the dialogue going on in Ohio and it gets to the heart of the issue. As of May, when the utility filed a proposal for how to address data center load, aep Ohio served 600 megawatts of data centers with about 4,400 megawatts of interconnection requests Fast forward to today, and those now exceed 40,000 megawatts.

Speaker 1:

One last note before we talk about why this is happening and what it all means. Mckinsey thinks EPRI is underestimating the demand and projects an 11 to 13 percent of total load growth by 2030, totaling 80,000 megawatts. Hmm, encore, aap and Dominion alone tot up well north of 80 gigawatts. So what's that all about? More on that later as well. It's pretty obvious that this very recent data center demand development, which seemingly came out of nowhere, has some serious implications for our climate and sustainability goals. Nobody really saw this coming, with the exception of a few industry insiders in the artificial intelligence space, and even most of them have been surprised at the speed with which this has blossomed. This dynamic is also likely to meaningfully impact electricity bills of anybody being served in markets where these data centers are growing. Why? Let's dive into the complexities and find out. First let's talk about AI and why the literal power grab. In March of 2016, something very unusual and somewhat earth-shaking happened, now known as Move 37.

Speaker 1:

We already knew that computers were getting better and perhaps smarter when IBM's Deep Blue beat world chess champion Garry Kasparov in 1997. But at some level, that was simple math. Then, in 2011,. Jeopardy's reigning champion, ken Jennings, accepted a challenge from IBM's Watson At the time. Jennings expressed confidence, saying later that he'd taken some AI classes, and quote I knew there were no computers that could do what you need to do to win on Jeopardy. People don't realize how tough it is to write that kind of program that can read a clue in a natural language like English. To understand the puns, the red herrings, to unpack just the meaning of the clue, I thought, yes, I will destroy the computer. Unquote. Watson, however, was able to decipher the puns and nuances and destroy Jennings, but it still couldn't reason.

Speaker 1:

The next highly publicized demonstration of Ari's prowess in a contest was AlphaGo's joust with top Korean Go player Lee Sedol. Go is way harder than chess. Chess has 64 squares, go has 361, and you place black or white stones strategically on the board to capture territory. In chess there are 30 possible next moves, while in Go you have over 200. So within two moves, in chess you have about 400 potential outcomes, versus around 130,000 in Go. Within just two moves, the AlphaGo team taught its computer to play against humans and then released it to play millions of games against itself. So on that much publicized joust in Korea, on move 37 of that first game, alphago placed a stone in a location that nobody would have expected appearing to demonstrate creativity, and it was the move that eventually won the game and demonstrated the growing capabilities of AI.

Speaker 1:

But AI still couldn't demonstrate prowess on tests or deal with any complexities of language Until recent years. Now it can, and it's getting better with startling speed. Consider two standard exams as proof. In 2022, chat-gbt 3.5 scored in the 40th percentile on the law school LSAT exam. In 2023, it jumped to 88th. How about the standard SAT? Ah, there we went from 87th to 97th in a year, and within a few years we were likely to achieve what is known as superintelligence. That's where computers exceed the thought processes of the best of us. How and why is this occurring so quickly? Quite simply, the machines are training more rapidly on more powerful chips. Next week, we'll discuss those chips and the electricity they devour. Thanks for watching and we'll see you then.