From GPUs to Gigawatts: The Race to Power AI

Bloomberg

Watch From GPUs to Gigawatts: The Race to Power AI - Bloomberg on Dailymotion

Transcript

00:00Today, I want to talk to you about our belief that AI is one of the most important new paradigms

00:11in the energy world, perhaps comparable to the LNG revolution or the breakthrough that we saw

00:21with U.S. shale, oil, and gas, or the stunning rise of renewables that we've seen in recent years.

00:30Earlier this month, Sam Altman was speaking before the U.S. Congress, and he said two things that I think

00:37are particularly relevant for the conversation today. One is he said that the cost of AI will

00:45ultimately converge to the cost of energy, and he also said the abundance of AI will be limited by

00:53the abundance of energy. Now, when I first heard Sam Altman speaking about the energy aspects of AI,

01:01I have to admit I was a bit skeptical. Sam and other AI luminaries were talking about this idea

01:08that energy was going to be such a constraint on AI. The CEO of Arm, microprocessor company,

01:17talked about the idea that U.S. power demand, about 20 to 25 percent of it, could be consumed by data

01:27centers by 2030. So I set out on a mission to disprove that, because I didn't think it felt right.

01:35I've worked in the energy business a couple of decades. But after doing some substantive work on

01:42the space, I think that the neighborhood is roughly correct. So first slide up here on the screen that

01:47I've got may help to illustrate this a little bit. So my colleague Mandeep Singh from our technology

01:56team at Bloomberg Intelligence, his group has sized the generative AI market. And the takeaway from the

02:03chart that you'll see on the left is that by 2030, this is potentially going to be a trillion dollar

02:11a year market. And so we're seeing hyperscalers, Alphabet, Microsoft, Meta, on down the line,

02:21really increasing CapEx. And over the last couple of years, CapEx has more than doubled for most of

02:30the players. They're building data centers, and those data centers consume quite a lot of power. So if you take

02:39the ideas from this group and translate it into power demand, here's what the figures look like. So if you

02:47use 2023 as a baseline, we think, at least in the U.S., data centers could, power demand could surge

02:59anywhere from 4X to 10X by 2030. These are profound numbers. Typically, you don't see things moving

03:10this quickly in the utility world. This is a slow-moving segment. And perhaps to that point,

03:20my colleagues at Bloomberg NEF, they just put out a great new report on this topic as well. Everybody

03:27wants to talk about AI these days. And they're closer to the lower end of our projection. Now,

03:34why is that? It's because if you do a bottom-up projection and you look in the U.S., the utilities

03:41simply can't get electrons on the grid as fast as these AI companies want to move. And so that leaves

03:52us perhaps with two options. One is that the AI revolution goes global, and I think we're seeing

04:01evidence that that's happening. The other is that maybe the industry will find a way. If you go read

04:07about XAI and what they've done with their facility in Memphis, I mean, they're building very quickly

04:16there, and they're being very creative about how they achieve that outcome. For those who don't think

04:24too much about power, I'll just quickly note that beyond AI, there are lots of other things that are

04:31shaping the narrative in that space. In the U.S. right now and globally, we are seeing a surge

04:38in manufacturing spending. This is partly propelled by the IRA trade policy. All sorts of things are

04:47contributing to this. But we are electrifying what we can electrify. So manufacturing, very electricity

04:57intensive vehicles increasingly moving in this direction as well. And so we are increasingly living

05:05in a world where we need more electrons. Now, what kind of electrons are getting built? This chart that

05:12you see up on the screen is the amalgamation of all of the power plants that want to connect to the grid

05:22in the U.S. Now, the takeaway here is that solar and storage are by far the two biggest types of

05:33power that folks are trying to build in the U.S. It's pretty analogous globally, too, by the way.

05:40We have about one terawatt of solar projects out there in the ether. For reference, the U.S.'s total

05:50installed power capacity is about 1.25 terawatts. So a lot of solar is out there, a lot of storage,

05:57a lot of wind. Of course, there is gas in the mix as well. But when you think about gas, there's a time

06:06constraint on the matter. If you want a new gas turbine, well, you have to get in line. If you think about a

06:13company like GE Vernova, they're currently taking orders for 2029, 2030. And so there's definitely a gas piece

06:23to the story. But because of the pace with which you can move with solar storage these days, we think they are the

06:33types of technologies to potentially benefit from this big build out in AI data centers.

06:42Quickly, I have a chart up on the screen just showing the solar names. They had a really rough 2024, also a challenging

06:512023, despite a lot of optimism from the U.S. president and the prior administration.

06:59We think they are back to growth in 2025 and future years as well, even with all the discussion about

07:07tax policy that's going on in the U.S. So look for solar names and storage names to be geared to

07:17the idea that data centers are rapidly propelling power demand.

07:23We think that the hyperscalers, despite their ambitions to be carbon neutral and all of their ESG

07:31goals are a very practical lot. They are focused on rapidly scaling AI and we don't see how you get there

07:41without a significant surge in gas use as well. This is a U.S.-focused chart. But by 2030, we think U.S. gas

07:51demand could increase somewhere by 3 to 10 BCF per day by 2030. And this is not the only pillar of the U.S.

08:02gas story. LNG demand also growing quickly, as you know, here in this country. And we think that,

08:13you know, the sort of playbook is that you see a combination of gas, solar and storage.

08:22Maybe a quick comment on nuclear, because as we saw with the video just leading up to me,

08:28CEO of NVIDIA talked about nuclear. Nuclear, great technology. It's certainly carbon-free.

08:36And what I'm showing on the chart here is nuclear output from the U.S., very stable over the last 25

08:43years. And we see this stunning rise of wind plus solar. At the end of the day, nuclear power plants

08:57they've all been, you know, behind schedule, over budget. Maybe there's some new technologies, SMRs,

09:05that will change that narrative. But good luck getting them built before 2030. I think you have to have

09:12a long vision for nuclear. Maybe it can play a very significant role in this question as you look

09:19out to 2035 and beyond. And then quickly, this is a chart showing data centers in Saudi Arabia. I think

09:29we're going to see build out in regions like this. We saw, while President Trump was over last week,

09:37the Humane announcement. Humane is a 500 megawatt data center. For reference, Saudi Arabia currently

09:45has about 165 megawatts installed. So this is a great leap from their current AI data center capacity

09:53with continued momentum to build much more. Now, if you want to go deeper on this topic, I'm going to

10:00quickly throw up a QR code. Feel free to scan that. You can go read our thoughts on the matter, connect with

10:09me on LinkedIn, say hello, and maybe just a parting thought. OpenAI has recently said that when we say please

10:17and thank you to the LLMs, it costs the company tens of millions of dollars. That's because of all the compute and

10:26energy necessary. So if you want to help alleviate some of the near-term energy constraints, perhaps

10:32don't be as polite with the AI. Thank you.

Category

Transcript

Be the first to comment

Recommended