D-Keine
Episode Description
The world remains abuzz over artificial intelligence, but rapid advancement and adoption of the technology is poised to drive a significant increase in power demand, and this demand could redefine energy consumption as we know it. Today we ask the critical question: is the energy sector equipped for the AI power revolution?
Will Su, of BlackRock’s Fundamental Equities team, is one of BlackRock’s leading voices on all things energy. Will walks us through the sector’s pivotal role in the build-out and future of AI and digs into the potential investment opportunities and challenges.
Transcript
Oscar Pulido: Welcome to The Bid, where we break down what’s happening in the markets and explore the forces changing the economy and finance. I’m Oscar Pulido.
The world remains a buzz over artificial intelligence, but rapid advancement and adoption of the technology is poised to drive a significant increase in power, demand, and this demand could redefine energy consumption as we know it today, we ask the critical question. Is the energy sector equipped for the AI power revolution?
Today I’m joined by Will Sue from BlackRock’s Fundamental Equities team. Will is one of BlackRock’s leading voices on all things energy. He’ll walk us through the sector’s pivotal role in the build out and future of AI, as well as dig into the potential investment opportunities and challenges Will, thank you for joining us on The Bid.
Will Su: Thank you, Oscar. Great to be here.
Oscar Pulido: So, Will, we’ve talked about artificial intelligence on the podcast a lot, and it seems like there’s no limits to the growth of this technology except the fact that it consumes a lot of energy and maybe that’s the constraint. Tell us a little bit about why AI consumes so much power.
Will Su: The simple answer to that extremely complicated question is that information processing is energy, and we are processing more information today than we’ve ever thought of, even from just a few years ago. At its most fundamental level, computations are just moving electrons around a semiconductor chip, but when you multiply that very small electric current by trillions of calculations, the energy demand adds up very, very quickly.
I think Rob Goldstein mentioned this the concept of AI is not really anything new. The MIT AI lab was started in the late 1950s, we did have a breakthrough moment in 2017 when a team of researchers wrote a paper about the transformer, which then became the architecture for today’s large language models or LLMs. Now, these models are being trained on trillions of parameters and tokens that make them high quality, high capacity, and able to contextualize the questions that they’re being asked.
And just to give you an idea of how big the computational power we’re talking about is here. ChatGPT4 was trained on about 70,000 Zetta flops of compute power. That’s 70 trillion trillion operations per second. Mind bending numbers. And as that number grows over time, that’s why we’re seeing this recent interest in meeting the power demand of AI.
Oscar Pulido: Did you say Zetta flops? ‘Cause I’m going to need a glossary. I think as we talk more about artificial intelligence, it feels like the terminology is new to a lot of people. And when you talk about power and the quantity, help us understand like, how much are we talking about on a global scale?
Will Su: So as anyone who tried to model this out can tell you it’s very hard to have a lot of confidence for 10, 20 years down the road when you’re looking at something with such exponential growth. That being said, we did build our own model because as they say, all models are wrong, but some are useful. In building this model, it’s helped us understand what the key variables are and maybe how the shape of that future power demand might look like.
And the punchline is, we think there could be up to 1000 terrawatt hours of incremental electricity demand for AI by 2030, and that would be about 3% of global electricity. And keep in mind that the internet today already consumes 2 to 3% of global electricity for things like data centers, networking transmissions and increasingly for blockchains. In aggregate you could see total internet demand, including AI, make up 6 to 7% of global electricity demand by 2030.
Oscar Pulido: And how is the world going to manage that power demand because it’s incremental on top of what is already the demand for power, right?
Will Su: Right. I think we can first dig a little bit into what is driving that AI demand. There’s really three roughly equal buckets in our 2030 outlook.
One is for training. So that’s the power that it takes to train these very large models. And again, just to give you an idea of the scale in 2022 Chat GPT-3 came out. It was trained on 175 billion parameters and 300 billion tokens. And the amount of energy it took to train could power about 90,000 US homes for a year.
Now you fast forward to 2023 when Chat GPT-4 came out, that model was reportedly trained on 1.8 trillion parameters and 13 trillion tokens. And the energy it took to train that could power 2.5 million U.S. homes for a year, and these models are getting bigger by the day.
And the good news there is with each generation of semiconductors, each generation becomes about 50% more power efficient. So, it takes half the amount of power for one calculation. It’s not enough to offset just how quickly the models are getting bigger, and then remember, more players are entering this game, globally, not just in the US but also in Europe and Asia. So, you add it all together and training really represents the bulk of the power growth that we see for AI in the coming few years.
The second bucket for demand is something called querying. So that’s when consumers, businesses, and other computers start to ask questions to these trained large language models. And in our model, we think you could see up to 30 billion AI queries per day by 2030. For comparison today, we make about 10 billion internet searches per day. But you have to remember that not all queries are created equally, right? A text-based query takes about the same amount of power as an internet search, but an AI generated photo takes up to 30 times more power, and a 60 second AI generated video takes up to 7,000 times more power than a text query. And video is big, it’s 57% of all internet traffic today. So how the consumer adapts to AI video is really one of the key variables that’ll determine just how much energy we’re going to require to power AI.
And then the third bucket is really for data center operations, mainly for cooling, because when you’re doing trillions of calculations per second, these chips run really hot.
So yes, 1000 terrawatt hours by 2030. That is a big number. I think it’s a challenging task to meet that demand, but not an impossible one.
Oscar Pulido: And maybe you can expand there because you shared a lot of numbers. you said the word trillions a couple times. the percentage increases that you’ve cited, particularly when you talked about how we use artificial intelligence to query, was quite large. So, what role do renewables play in this energy demand? I’m thinking about things like wind solar, are they the major component or are there other, sources of energy that we’re going to rely on?
Will Su: So, renewables are by far the fastest growing source of power generation. In the last 20 years. They’ve gone from almost nothing to 13% of global power generation. And they will continue to grow at a very fast pace.
Without a doubt, renewables are going to play a big part, in powering AI, but also in powering this overall theme of electrification of our energy systems. Now renewables have one really big drawback when it comes to powering AI, which is intermittency. Right? Let’s zoom into the Ercot grid in Texas, which is the largest wind market and the second largest solar market in the U.S.
So, it has a lot of renewables, and if you just zoom in on a typical day, the solar power tends to peak out between 8:00 AM and 7:00 PM when the sun’s shining. And the wind peaks when the wind speeds are the highest, which is usually from midnight to 7:00 AM when you wake up. Peak demand really happens in the hours of 8:00 PM to midnight. That’s when people are at home relaxing, watching TV, streaming, checking their social media. And you’ll see that during that period, natural gas demand really increases to meet that gap that can’t be met by wind and solar.
And this is probably a good time to talk about nuclear, which people don’t think of a lot, but it’s actually today the largest source of carbon free power generation. It makes up about 9% of global power.
But I think as governments around the world start to realize how much electricity growth there’s going to be, there’s starting to be a change in thinking. And in countries like South Korea, Japan, Italy, and here in the U.S. you’re seeing regulators extending previously planned shutdowns of nuclear plans, and even in some cases, allowing them to restart after they’ve already been shut. So definitely don’t count nuclear out in this low carbon way to power AI going forward.
Oscar Pulido: So, it sounds like the, the demand is so significant that it is causing even some sources of energy that in the past that felt like, were becoming less of a priority to reenter the focus. Ultimately what you’ve said is there’s a lot of different sources of energy that are going to help, power the AI demand. You mentioned nuclear gas, but also renewables. And if I could focus you on the US for just a second, artificial intelligence is not just the US topic, but it is the part of the world where the build out is really gaining a lot of momentum and therefore, how is the U.S. thinking about the power supply for artificial intelligence?
Will Su: we really should talk about one of the biggest unsung triumphs in the energy transition so far, which is the U.S. Power grid has decarbonized itself by a third. Over the last 20 years, and about 60% of that came in the form of cheap and abundant natural gas as a result of the shale revolution that allowed us to substitute out much more polluting coal.
You saw a coal share in the last 20 years drop from 50% to 16%. Natural gas went up from 19% of U.S. Power generation to 42. The other 60, the other 40% came from renewables. So, renewables, again, grew from almost nothing 20 years ago to 14% of the US power grid today. So, there’s already a really strong track record of partnerships between natural gas and renewables to combine and help us decarbonize.
Now, when you think about AI, and you think about data centers. The U.S. has about one third of the total data center capacity in the world, and I’m very confident that share will grow over time because we have the leading technology companies that are leading this AI revolution. And then we are also blessed with abundant resources, both traditional and renewable.
If you look at a map of where these data centers are located in the U.S., you’ll see that they’re mostly in these big clusters that are located close to population centers. So almost half of all data centers in the U.S. are in Virginia. They’re almost all in this six square mile tiny area called Data Center Alley near Arlington.
There’s other big clusters like Hillsborough, near Portland, Oregon. there’s also growing clusters around Ohio, and you’ll see a problem if you juxtapose that map onto one where the renewable resources are best in this country. The source of greatest solar radiation is in the southwest U.S., so that’s places like Southern California, Nevada, New Mexico, and where the wind speeds are the highest are down the middle of the U.S. In the windy corridor that goes from the Dakotas down to Northern Texas.
And they don’t really overlap with where the data centers are located today and where the most growth is likely to happen in the coming years. And then to make matters worse, this country’s really falling behind in making long distance transmission investments. We’re making one eighth the mileage of new transmission lines than we did 10 years ago.
That’s a result of a number of regulatory and economic challenges with interstate infrastructure, and this is where natural gas is going to come in. It’s a proven, mature technology. It’s much cleaner than coal. It plugs easily into different grids, so it shapes my view that I think at least half, if not more of the incremental power for AI in the U.S. will come from natural gas and the balance will mostly be met by new renewable developments.
Oscar Pulido: Data Center Alley doesn’t sound as glamorous as like Silicon Valley, but it seems like it’s also very important. Let’s come back to your role as an investor. You spend your day thinking about companies to invest in, and if you follow the markets over the last couple years, it’s been all about technology. But we’re having a discussion about the energy space, and so presumably that means there’s investment opportunities in the energy sector. Where are those?
Will Su: Absolutely. So, as a value minded income investor, I have thought for a long time that energy is an undervalued sector because the market under appreciates both the volume and the duration for which the world needs oil and gas for the decades to come.
And I think this recent focus on how do we power AI just shines yet another spotlight on how power hungry our world really is. And over time, I think that will help this sector rerate higher. Now, aside from that, I think the energy sector actually might be one of the most underappreciated beneficiaries for all the technological gangs that’ll come with better generative AI.
Some of the world’s largest supercomputers are actually owned by large energy companies. Why? Because they perform a number of very computationally intensive tasks. Things like asset optimization, algorithmic trading, four D seismic imaging for new resource discoveries.
And I’ll give you one specific example, which is the industry is using more and more of what’s called a digital twin. So, this is like a virtual replica of a real-world asset, something like a refinery or an offshore platform. It’s just got so much data inside of it that you can do a lot of really interesting and exciting things. Things like predictive maintenance, fixing things before they break, things like stress, testing them for severe weather events or identifying methane leaks and reducing emissions that way. So, I think there’s more than one way to win with energy when it comes to the theme of AI that’s greatly underappreciated by the market today.
I think the other sector that deserves some airtime here is utilities. So, utilities are a yield driven high dividend paying sector that’s been somewhat out of favor in the last few years in a rising rate environment. But as the U.S. Grid goes from not having much growth for the last 20 years to needing to grow one to 2% per year going forward, there’s a big opportunity for these utilities.
It’ll come after an initial period of heavy investments now, which utilities will win depends very strongly on what regulatory regime and what geography they operate in.
Oscar Pulido: And it’s interesting just to hear you talk about energy and utilities. I’m reminded we spoke to your colleague Carrie King, who reminded us that while it has been a very tech-driven market, in the last couple of years, there are opportunities that are starting to appear. And you’re zooming in on the energy and utility sector as a function of artificial intelligence and power demand. But for an investor who is looking at this space, what should they be considering as they think about investing?
Will Su: The energy sector contributes about 10% of the S&P 500’s net income, but it makes up less than 4% of the index by market cap. And I think that valuation disconnect is driven by this persistent, and in my view, misplaced fear that this sector has no long-term growth. Because I think as we sit here talking about breakthrough technologies like generative AI, it is important for us to remember that there’s many different poles for incremental energy demand in this world, and all or nothing approach to energy just isn’t going to work.
We have to find ways to help the traditional energy sources become cleaner and more responsibly sourced. At the same time, we scale up our renewables portfolio together, and only together will they be able to power the world forward in a pragmatic energy transition.
Oscar Pulido: Right, the world is evolving, where the demand for energy will come from is changing. With the number of statistics that you’ve been able to cite about the energy sector and artificial intelligence, where does this passion come from? How did you get interested in this space?
Will Su: Oscar, I’m having flashbacks to 16 years ago when I started my career at a large investment bank in the equity research department, and my recruiter said, you can either join the internet team or the energy team. And I had no hesitation. I said, energy, it’s supply-demand driven. It’s quantitative. The world needs this stuff. And you fast forward to today, and I think the internet index has outperformed energy by about 1,100%. But if you gave me a time machine to go back, I will make the same choice over again.
This job has taken me to really exciting places all over the world. Offshore Norway, the Permian Basin in Texas, the Bakken in North Dakota, or deep into the Amazon jungle in Guyana. That’s a country that’s going to go from the second poorest in South America to having the same GDP per capita as Brazil in less than a decade because of their resource development. So, it’s been a really thrilling ride so far, and I look forward to more of what’s to come.
Oscar Pulido: We’re glad you made that decision 16 years ago and that you would make it again, if you went back in time. Thanks for sharing all this insight on the energy sector, on artificial intelligence, and thank you for doing it here on The Bid.
Will Su: Thank you, Oscar,
Oscar Pulido: Thanks for listening to this episode of The Bid. If you’ve enjoyed this episode, check out our episode with Rob Goldstein and Lance Bronstein. Where they discuss AI through a COO lens and what business leaders are considering as AI is advancing.
Sources
“Electricity Mix” Our world in energy, January 2024;
“What is U.S. electricity generation by energy source?” Energy Information Administration, “OpenAI Presents GPT-3, a 175 Billion Parameters Language Model” Nvidia, 2020;
GPT-4 Details Revealed, Patrick McGuinness, 2023; Data Centers Around The World, United States International Trade Commission 2021;
“North America Data Center Trends H2 2023”, CBRE 2024;
“Electric power sector CO2 emissions drop as generation mix shifts from coal to natural gas” EIA, 2021;
“Electravision” JPMorgan, March 2024;
“Fuel Mix” Ercot, March 2024; “Television, capturing America’s attention at prime time and beyond” US bureau of Labor Statistics, September 2018.
Written Disclosures
This content is for informational purposes only and is not an offer or a solicitation. Reliance upon information in this material is at the sole discretion of the listener.
For full disclosures go to Blackrock.com/corporate/compliance/bid-disclosures
MKTGSH0524U/M-3571844
This post originally appeared on BlackRock.
Editor’s Note: The summary bullets for this article were chosen by Seeking Alpha editors.
