WEBVTT - Bloomberg Talks: Applied Digital CEO Wes Cummins Talks Data Centers

0:00:00.040 --> 0:00:01.920
<v Speaker 1>All right, it's time now for our daily Wall Street

0:00:01.920 --> 0:00:05.680
<v Speaker 1>Week conversation, and the rise of AI is posing challenges

0:00:05.720 --> 0:00:09.440
<v Speaker 1>to data centers. Applied Digital, which designs, develops, and operates

0:00:09.520 --> 0:00:12.480
<v Speaker 1>data centers, is looking at how to adapt its infrastructure.

0:00:12.800 --> 0:00:15.400
<v Speaker 1>Wes coom and See is the company's co founder, chairman,

0:00:15.440 --> 0:00:18.400
<v Speaker 1>and CEO. He joins us now alongside Wall Street Week's

0:00:18.520 --> 0:00:21.160
<v Speaker 1>David Weston and data centers. I mean, it's one of

0:00:21.200 --> 0:00:22.639
<v Speaker 1>the hottest toppings out there, it really is.

0:00:22.680 --> 0:00:25.160
<v Speaker 2>Every day we're talking about data centers somehow, some way.

0:00:25.200 --> 0:00:27.440
<v Speaker 2>But we thanks so much for joining us because normally

0:00:27.440 --> 0:00:29.280
<v Speaker 2>when Katie and I talk with people who at data centers,

0:00:29.280 --> 0:00:31.159
<v Speaker 2>it tends to be the commercial real estate side of

0:00:31.160 --> 0:00:33.639
<v Speaker 2>it and investig in it. But you actually developed these

0:00:33.680 --> 0:00:36.720
<v Speaker 2>things and actually install and operate them. So give us

0:00:36.720 --> 0:00:39.199
<v Speaker 2>a sense of what's the big deal with data centers

0:00:39.240 --> 0:00:42.200
<v Speaker 2>now that didn't exist before the rush to AI. We've

0:00:42.200 --> 0:00:43.800
<v Speaker 2>had them around for some time, right.

0:00:44.479 --> 0:00:49.720
<v Speaker 3>Yeah, David, I think that the key to understand, the

0:00:49.760 --> 0:00:53.519
<v Speaker 3>primary thing to understand is the difference in the compute

0:00:53.520 --> 0:00:55.440
<v Speaker 3>workload that's going into data.

0:00:55.160 --> 0:00:56.960
<v Speaker 4>Centers now versus what it was in the past.

0:00:56.960 --> 0:00:59.360
<v Speaker 3>So in the past, you know, we were building data

0:00:59.400 --> 0:01:02.640
<v Speaker 3>centers that were in population centers that we're providing ultralo

0:01:02.840 --> 0:01:05.240
<v Speaker 3>latency communications so that.

0:01:05.200 --> 0:01:07.800
<v Speaker 4>We could get YouTube, so that we could get Netflix.

0:01:07.360 --> 0:01:09.080
<v Speaker 3>So we could get TikTok, right, so we can get

0:01:09.520 --> 0:01:11.560
<v Speaker 3>all of our things that have been driven mostly by

0:01:11.600 --> 0:01:12.600
<v Speaker 3>video applications.

0:01:13.000 --> 0:01:14.240
<v Speaker 4>Now, as we're moving.

0:01:14.200 --> 0:01:18.759
<v Speaker 3>From applications that need ultralo latency communications outside the data

0:01:18.800 --> 0:01:23.200
<v Speaker 3>center to applications that need massive amounts of compute. So

0:01:23.480 --> 0:01:26.000
<v Speaker 3>think of, you know, doing large math problems algorithms. This

0:01:26.120 --> 0:01:29.240
<v Speaker 3>is what AI is doing inside the data center. Compute

0:01:29.280 --> 0:01:32.640
<v Speaker 3>equals power, and then inside the data centers you need

0:01:32.720 --> 0:01:36.200
<v Speaker 3>a lot of compute close together, so you need really

0:01:36.280 --> 0:01:38.240
<v Speaker 3>high power density. And so these are kind of some

0:01:38.280 --> 0:01:41.600
<v Speaker 3>weird numbers, but you know, historically data centers were let's

0:01:41.600 --> 0:01:44.959
<v Speaker 3>call it seven kilowatts for an entire rack of computing,

0:01:45.319 --> 0:01:48.360
<v Speaker 3>and Nvidia a one hundred server one by itself takes

0:01:48.400 --> 0:01:51.120
<v Speaker 3>ten ten kilo wants. So and then you need to

0:01:51.120 --> 0:01:53.280
<v Speaker 3>stack all of these together, so you have a massive

0:01:53.400 --> 0:01:55.880
<v Speaker 3>amount of power consumption that needs to be close together.

0:01:56.000 --> 0:01:59.200
<v Speaker 3>The latency inside the data center now versus outside, needs

0:01:59.200 --> 0:02:01.440
<v Speaker 3>to be ultralove, and so you have to build a

0:02:01.480 --> 0:02:04.560
<v Speaker 3>completely different style of data center, you need large amounts

0:02:04.600 --> 0:02:06.920
<v Speaker 3>of power in the same location, and so that's what

0:02:06.960 --> 0:02:10.520
<v Speaker 3>we're doing. We find large amounts of power, typically different

0:02:10.600 --> 0:02:12.080
<v Speaker 3>locations than you would do city center.

0:02:12.120 --> 0:02:12.840
<v Speaker 4>So it's no.

0:02:12.840 --> 0:02:15.960
<v Speaker 3>Longer New York or Virginia or Dallas or Los Angeles.

0:02:16.200 --> 0:02:18.760
<v Speaker 3>We're building in different locations where we use a lot

0:02:18.800 --> 0:02:22.120
<v Speaker 3>of renewable energy that have these hundreds of megawatts, if

0:02:22.160 --> 0:02:25.840
<v Speaker 3>not gigawat of power so that you can run these

0:02:25.880 --> 0:02:29.000
<v Speaker 3>workloads because they're just very different style workloads versus what

0:02:29.040 --> 0:02:29.480
<v Speaker 3>we're used to.

0:02:29.760 --> 0:02:32.720
<v Speaker 1>Right and when it comes to that adaption process really

0:02:32.840 --> 0:02:36.000
<v Speaker 1>sort of converting those existing data centers to be able

0:02:36.040 --> 0:02:40.120
<v Speaker 1>to handle the AI workload that people are trying to do.

0:02:40.200 --> 0:02:43.000
<v Speaker 1>I mean, how long the process is that? How long

0:02:43.080 --> 0:02:44.200
<v Speaker 1>is that timeline?

0:02:44.760 --> 0:02:48.680
<v Speaker 3>So I'm a big believer that you're better doing green

0:02:48.760 --> 0:02:51.720
<v Speaker 3>Field in that you just build an entirely new style

0:02:51.760 --> 0:02:53.959
<v Speaker 3>of data center versus trying to retrofit.

0:02:55.160 --> 0:02:56.160
<v Speaker 4>Cooling is different.

0:02:56.240 --> 0:03:00.560
<v Speaker 3>Again, the power density is significantly different, so there's there's

0:03:00.600 --> 0:03:01.280
<v Speaker 3>a big issue.

0:03:01.600 --> 0:03:04.519
<v Speaker 4>So the number one issue right now is finding power.

0:03:04.600 --> 0:03:07.960
<v Speaker 4>Where's near term power? We have we have a very.

0:03:07.960 --> 0:03:10.600
<v Speaker 3>Unique way of finding power, which is we find things

0:03:10.639 --> 0:03:13.040
<v Speaker 3>that are called stranded power, and we typically use a

0:03:13.080 --> 0:03:15.680
<v Speaker 3>lot of renewable energy that's more in the middle of

0:03:15.720 --> 0:03:16.360
<v Speaker 3>the country.

0:03:16.960 --> 0:03:18.920
<v Speaker 4>So number one is power, and then number two is

0:03:18.960 --> 0:03:19.640
<v Speaker 4>supply chain.

0:03:19.760 --> 0:03:23.560
<v Speaker 3>So this is electrical components, it's just you know, it's transformers,

0:03:23.639 --> 0:03:27.000
<v Speaker 3>it's chillers, it's high voltage switch here. So the supply

0:03:27.080 --> 0:03:30.400
<v Speaker 3>chain because of the demand for this has become really

0:03:30.440 --> 0:03:31.160
<v Speaker 3>stretched as well.

0:03:31.200 --> 0:03:33.239
<v Speaker 4>So number one is power. Number two supply chain.

0:03:33.400 --> 0:03:35.400
<v Speaker 2>So wes are you building your own power? I mean

0:03:35.400 --> 0:03:38.280
<v Speaker 2>you're building those wind farms or solar event or even

0:03:38.400 --> 0:03:40.520
<v Speaker 2>nuclear for that matter, or are you tapping into some

0:03:40.560 --> 0:03:43.120
<v Speaker 2>other part of the grid. And similarly on the supply chains,

0:03:43.440 --> 0:03:45.200
<v Speaker 2>how long is the lag time on all this? It

0:03:45.200 --> 0:03:46.480
<v Speaker 2>sounds like it's going to take an awful lot of

0:03:46.480 --> 0:03:48.840
<v Speaker 2>time to build that power and also come up with

0:03:49.040 --> 0:03:50.040
<v Speaker 2>all that supply chain.

0:03:50.840 --> 0:03:53.440
<v Speaker 3>So right now we're not building the power that's something

0:03:53.480 --> 0:03:56.120
<v Speaker 3>for the future for us. So there's there's a unique

0:03:56.120 --> 0:03:59.520
<v Speaker 3>setup in North America with that we take advantage of

0:03:59.760 --> 0:04:01.360
<v Speaker 3>what I call it is stranded power.

0:04:01.760 --> 0:04:04.680
<v Speaker 4>So there is power that is generated and not used.

0:04:04.720 --> 0:04:06.280
<v Speaker 3>And I know a lot of people would think that

0:04:06.320 --> 0:04:08.560
<v Speaker 3>if you can generate an electron anywhere in the US,

0:04:08.560 --> 0:04:09.720
<v Speaker 3>it can be used anywhere else.

0:04:09.560 --> 0:04:11.000
<v Speaker 4>In the US, and that's just not the case.

0:04:11.360 --> 0:04:14.720
<v Speaker 3>You need the transmission infrastructure to actually be able to

0:04:14.720 --> 0:04:18.400
<v Speaker 3>do that. So we find places where, you know, for Ellendale,

0:04:18.440 --> 0:04:22.240
<v Speaker 3>North Dakota, for example, there's a massive amount of wind

0:04:22.320 --> 0:04:25.760
<v Speaker 3>power that feeds into a substation, and oftentimes they have

0:04:25.800 --> 0:04:28.920
<v Speaker 3>to curtail the wind farms because there's not enough demand,

0:04:28.960 --> 0:04:32.039
<v Speaker 3>there's not enough load to use the power that's generated.

0:04:32.320 --> 0:04:33.799
<v Speaker 4>So we go to that location.

0:04:34.279 --> 0:04:36.320
<v Speaker 3>And we have several of these locations where we use

0:04:36.360 --> 0:04:39.880
<v Speaker 3>power that's currently not being used now to go to

0:04:39.920 --> 0:04:43.000
<v Speaker 3>the other part of your question is for new power,

0:04:43.080 --> 0:04:46.640
<v Speaker 3>so you can build power gen on renewables over you know,

0:04:46.720 --> 0:04:49.000
<v Speaker 3>let's call it a twenty four maybe thirty six month

0:04:49.000 --> 0:04:49.440
<v Speaker 3>time frame.

0:04:49.520 --> 0:04:52.039
<v Speaker 4>The bigger issue is transmission.

0:04:52.160 --> 0:04:55.520
<v Speaker 3>So transmission to transmit that power that typically is built

0:04:55.720 --> 0:04:58.520
<v Speaker 3>somewhere in the middle of the country to end users.

0:04:58.720 --> 0:05:01.080
<v Speaker 3>You know, that's more of a tend thirteen year process,

0:05:01.120 --> 0:05:03.960
<v Speaker 3>and so we're kind of shortcutting that by taking these

0:05:04.000 --> 0:05:06.160
<v Speaker 3>workloads directly to the point of generation.

0:05:06.800 --> 0:05:09.080
<v Speaker 1>I want to talk about the competitive landscape here because

0:05:09.120 --> 0:05:12.200
<v Speaker 1>you think about AI and it's just really made its

0:05:12.200 --> 0:05:14.919
<v Speaker 1>way into every part of the market. It seems like

0:05:14.960 --> 0:05:17.960
<v Speaker 1>every investor in every asset class is trying to figure

0:05:17.960 --> 0:05:20.920
<v Speaker 1>out what's the AI narrative in their industry. When you

0:05:20.920 --> 0:05:23.559
<v Speaker 1>think about your own industry and data centers right now,

0:05:23.760 --> 0:05:25.120
<v Speaker 1>how fierce is the competition.

0:05:27.279 --> 0:05:29.320
<v Speaker 4>So it's an interesting time for sure.

0:05:29.400 --> 0:05:32.120
<v Speaker 3>So if you go back a year ago, everyone is

0:05:32.120 --> 0:05:34.640
<v Speaker 3>trying to figure out on the data center side, what's

0:05:34.680 --> 0:05:37.679
<v Speaker 3>happening in this market right Why is all of this demand?

0:05:37.800 --> 0:05:39.320
<v Speaker 4>Why for the high power density?

0:05:39.680 --> 0:05:43.200
<v Speaker 3>Now in this industry figured it out, and now it's

0:05:43.520 --> 0:05:47.000
<v Speaker 3>trying to make this solution because again, getting the amount

0:05:47.000 --> 0:05:50.160
<v Speaker 3>of power that's needed and operating the style of data

0:05:50.200 --> 0:05:54.840
<v Speaker 3>centers is extremely difficult. And so think of on the

0:05:54.920 --> 0:05:57.640
<v Speaker 3>data center side, we're the picks and shovels. This is

0:05:57.720 --> 0:06:01.240
<v Speaker 3>digital infrastructure. This is the digital infrastructure was built out

0:06:01.240 --> 0:06:03.840
<v Speaker 3>to run everything that we're you know, everyone's used to

0:06:03.920 --> 0:06:05.719
<v Speaker 3>running on their phone or on their TV at home,

0:06:05.960 --> 0:06:09.000
<v Speaker 3>and now it has to be rebuilt. It's not a replacement,

0:06:09.000 --> 0:06:11.599
<v Speaker 3>by the way, this is an add on, but rebuilt

0:06:11.720 --> 0:06:15.320
<v Speaker 3>for compute and AI workloads. So we're just we're at

0:06:15.360 --> 0:06:18.640
<v Speaker 3>the building blocks. We're the infrastructure that runs all the

0:06:18.680 --> 0:06:20.400
<v Speaker 3>AI that you know, everyone wants to do and we

0:06:20.520 --> 0:06:22.880
<v Speaker 3>talked about you know, everyone knows chat GBT, but if

0:06:23.080 --> 0:06:26.360
<v Speaker 3>you're doing anything in in uh, you know, in video

0:06:26.640 --> 0:06:30.919
<v Speaker 3>and visual anything generative AI runs on this infrastructure and

0:06:30.960 --> 0:06:33.159
<v Speaker 3>a massive amount of this infrastructure used to be built,

0:06:33.480 --> 0:06:35.520
<v Speaker 3>you know, and when Jensen and in Video talks about

0:06:35.520 --> 0:06:38.880
<v Speaker 3>AI factories. So so what we've done we actually designed

0:06:38.880 --> 0:06:42.159
<v Speaker 3>our data center and you know, in partnership with in Video,

0:06:42.160 --> 0:06:44.560
<v Speaker 3>we work with their team to build the style of

0:06:44.640 --> 0:06:47.400
<v Speaker 3>data center that we're doing because it's this is the

0:06:47.400 --> 0:06:49.520
<v Speaker 3>issue in the data center industry is it's just a

0:06:49.760 --> 0:06:52.960
<v Speaker 3>very different style from data centers that were done in

0:06:52.960 --> 0:06:56.200
<v Speaker 3>the past. But we're running all AI workloads and you

0:06:56.200 --> 0:06:59.360
<v Speaker 3>know these are really split between training and inference. So

0:06:59.440 --> 0:07:03.000
<v Speaker 3>the terms that are used in AI, but the entire

0:07:03.000 --> 0:07:06.599
<v Speaker 3>industry is i would call it scrambling to build the

0:07:06.640 --> 0:07:09.720
<v Speaker 3>amount of infrastructure that's needed to house the compute to

0:07:09.840 --> 0:07:11.280
<v Speaker 3>run these applications.

0:07:11.960 --> 0:07:13.640
<v Speaker 2>Well, it wasn't that long that we hadn't even heard

0:07:13.680 --> 0:07:16.280
<v Speaker 2>about all these data centers. What's going to prevent absolutcence?

0:07:17.800 --> 0:07:19.440
<v Speaker 4>So that's a great question.

0:07:19.520 --> 0:07:22.160
<v Speaker 3>So when we build our facilities, we think about what

0:07:22.200 --> 0:07:25.280
<v Speaker 3>we call future proofing. Right, so we think the power

0:07:25.280 --> 0:07:29.480
<v Speaker 3>density goes higher cooling, but I think you know, the

0:07:29.680 --> 0:07:33.880
<v Speaker 3>last twenty thirty years were again really driven by applications

0:07:33.920 --> 0:07:37.440
<v Speaker 3>that required call comms, high speed communications, and as we

0:07:37.480 --> 0:07:39.800
<v Speaker 3>go for the next twenty or thirty years, applications are

0:07:39.840 --> 0:07:42.120
<v Speaker 3>really going to be driven by compute and that's what

0:07:42.200 --> 0:07:45.080
<v Speaker 3>AI is, and so we're working to future proof that.

0:07:45.160 --> 0:07:49.600
<v Speaker 3>And by the way, the older style data centers, those

0:07:49.640 --> 0:07:53.200
<v Speaker 3>are not obsolete. We still need all of that capacity

0:07:53.280 --> 0:07:55.720
<v Speaker 3>for the applications that we continue to use. This is

0:07:55.960 --> 0:07:59.480
<v Speaker 3>absolutely a complete add on to what's been done in

0:07:59.520 --> 0:08:02.720
<v Speaker 3>the past and so but trying to not be obsolete

0:08:02.800 --> 0:08:05.600
<v Speaker 3>is making sure that your future proofing your facilities as

0:08:05.680 --> 0:08:08.320
<v Speaker 3>much as possible. With power density, with cooling, we're trying

0:08:08.320 --> 0:08:11.440
<v Speaker 3>to run an extremely high efficiency so that we lower

0:08:11.440 --> 0:08:15.080
<v Speaker 3>our carbon footprint. You know, we build where there are

0:08:15.400 --> 0:08:18.680
<v Speaker 3>renewable we mostly use renewable resources. But those are the

0:08:18.720 --> 0:08:21.000
<v Speaker 3>things you try to do obviously, you know, not being

0:08:21.040 --> 0:08:23.280
<v Speaker 3>obsolete in the world is one of the keys here.

0:08:23.400 --> 0:08:25.600
<v Speaker 1>All right, future proofing, that's a good place to leave it.

0:08:25.680 --> 0:08:28.560
<v Speaker 1>Really enjoyed this conversation our thanks to wes Coomins. He

0:08:28.720 --> 0:08:31.720
<v Speaker 1>is Applied Digital co founder, chairman and CEO