AI’s Biggest Bottleneck Isn’t GPUs — It’s Power | Sergii Gerasymovych

Vit Lyoshin (00:01.056)
Welcome Sergii, thank you for your time today.

Sergii (00:03.64)
Thank you for having me.

Vit Lyoshin (00:05.898)
Great. So we're talking about AI infrastructure, data centers and stuff like that today. But I know you started with Bitcoin my mining or crypto mining in general, and now you're trying to shift a little bit to AI data centers, infrastructure and stuff like that. why is that? Why now?

Sergii (00:28.546)
Very good question. I think first it's economics. Economics of crypto mining, Bitcoin mining just became not feasible in the recent couple years. so that's one. Second one is AI data center business is just too good to miss and you definitely have to try. there's a huge change in era of technology where basically

Vit Lyoshin (00:38.838)
Mm-hmm.

Sergii (00:56.854)
All the software is moving to AI. I mean whatever we call AI, is it artificial intelligence in whatever form? But the tools that we've been using before is the software tools, they're becoming the tools with intelligence. And I see that becoming a huge, huge industry and every

every aspect of our life and our work gonna be connected to AI somehow. So that means it will require a lot of compute and compu compute is based on the foundation of energy and power, which is basically what we've been doing for crypto. So it is an easy pivot.

Vit Lyoshin (01:29.749)
Uh-huh.

Vit Lyoshin (01:39.284)
Yeah. Yeah, I was gonna follow up really quick. Some lessons from crypto probably applicable to AI infrastructure, and maybe you can mention a couple of them.

Sergii (01:49.954)
Yeah, absolutely. as a matter of fact, twenty seventeen, when the crypto mining, crypto in general became a big hype, reminds me nowadays. so it's ten years later, almost ten years later, it reminds me when people wanna solve hardware problems very fast. And meaning like they wanna go and build the data center really fast and you have to be online within months, not years.

Vit Lyoshin (02:09.162)
Mm-hmm.

Sergii (02:18.1)
It literally reminds me the crypto world. And that's what we've been doing. So it's easier for us to look at the at the problem of AI data center infrastructure through the lenses of a crypto mining data center infrastructure because they are similar. the cost is a little bit different of compute inside, but in general, they are literally the same. so we are focused very much on

same problems which are going to the market pretty fast, deploying a compute infrastructure really fast, fixing problems on the fly, and inventing on the fly. So they have these are the similarities.

Vit Lyoshin (03:02.663)
I see, okay. And what exactly that you're doing right now? Is it like actual data centers or computers, chips? What is it exactly?

Sergii (03:10.562)
Yeah, well, Easy Blockchain started in twenty seventeen as a collocation company. We provided collocation for people who wanted to have exposure to Bitcoin mining through buying equipment. right now that's still most of our business. we're still a data center provider where we own land power and modular data centers. Those are in a form of like forty foot shipping container.

Vit Lyoshin (03:22.644)
Mm-hmm.

Sergii (03:38.718)
AI data center business that we recently started focusing on is more in a concept of la of layers, right? The first layer in AI data center is is power, second is land, the third one is infrastructure, and the fourth one would be the compute. And then there are other layers, software layers. But we are focused right now on the first two layers, which is power and land. So our job is to make sure to we are

Capturing as much power as possible, as much land as possible with the power, because we believe that these layer layers are finite and all others are not. if we speak about the compute layer, Nvidia are gonna keep printing chips. They're they're good at it. And they're gonna be printing more. Qualcomm gonna be doing that, Google gonna be doing it. You can print more land and you can't print power that fast as they can print compute.

Vit Lyoshin (04:27.029)
Mm-hmm.

Sergii (04:37.29)
so we are focused on those two layers. Our next step is developing the data centers themselves. we most likely gonna listen to the market what market wants, and the market wants right now speed. that's why they focus on modular deployments, where you build AI data centers somewhere else and you deliver it to the site which is already prepped, and you bring the computers online.

Vit Lyoshin (05:04.071)
Mm-hmm. That that's an interesting approach. I I never heard this b not many people talk about this that that land and power is really the bottleneck here, is really a challenge. because everybody's talking about GPUs and CPUs shortage and just building actual computers that compute stuff, but not everything around them that keeps them cool and keeps them powered and things like that. And why do you think that

Land and power is actually a bigger bottleneck than Nvidia and Qualcomm printing chips.

Sergii (05:38.178)
Yeah, well, the reason I think that this is a a big bottleneck because the traditional world of business is not moving at the speed of Silicon Valley and innovation. power is run by mostly cooperatives, electrical companies, and they got used to do things slowly. you know, y like think about this like well, if you need to build a new power line.

Vit Lyoshin (06:01.365)
Mm-hmm.

Sergii (06:06.058)
It's gonna take you years to do that. well if you wanna do a software problem in in you know in in one of your products, you're gonna try to focus on the problem as fast as fast as possible. So data centers rely on power as the key because the demand for compute has skyrocketed, but the supply of power, a foundate foundational part of the compute, has not skyrocketed

Just because the utilities, well, first of all, they need to bridge the gap of understanding who is their customer, for example. A utility gonna listen to to me, for example, you some tech bro, and gonna say, Well, what are you gonna do there? Like you're gonna do compute? What AI data center? What what does that mean, right? well, you have to they have to understand that the computer.

Vit Lyoshin (06:55.401)
Mm-hmm.

Sergii (06:59.402)
the compute the data centers is a new type of a customer that's going to be on the grid for next years. Well then they have to go and force the board to finance this project. So they go and order transformers. When then they order transformers, electrical equipment, they get into a problem where then everyone is ordering transformers. Well so if you want to build a transformer, you have to order aluminum and then you have to order copper. And then it just becomes all this cascade effects of many, many parties.

from real physical world that participate, participate in this process of getting that electron flowing into a data center where the computer can be installed. And it's usually a very slow process. It's usually the process that takes literally years. And well, today's demand, today's AI race is not designed for it. Google wants sites online tomorrow. Meta wants to do the same.

Vit Lyoshin (07:48.361)
Mm-hmm.

Sergii (07:56.854)
Microsoft lost the game to Google in search. They don't want to lose the game in in compute in AI, right? So this this is what's happening and now these big companies they're throwing trillions of dollars literally downstream and say, Well, you guys go figure out the problem. Well, it creates a huge cascade effect effect of of things that happen around this demand, how to capture this demand, and everyone is running around like chickens without the

Vit Lyoshin (08:05.833)
Mm-hmm.

Vit Lyoshin (08:25.493)
I see. Okay. And yeah, okay, makes sense. And then also we have solutions for like clean energy, right? People talking about nuclear energy, building huge, huge data centers that will be like self sustaining and things like that. Is that not gonna be helping at all here? Or what's the like idea for this type solutions?

Sergii (08:51.682)
Yeah, I think I think there are a few a few parts to this. There's short term, long term, and medium term, right? Well, short term, that's when we talk to our customers and they say we want to deploy our compute tomorrow. That's not this compute compute deployment of tomorrow is not gonna work with with nuclear power plant, right? Well then the l the midterm would be next two, three years. That's when, well, I need fifty to hundred megawatts of compute.

Vit Lyoshin (08:58.196)
Mm-hmm.

Sergii (09:20.866)
To be deployed in my campus. that's also not gonna work with the nuclear power plants. Then there's a long term, and where we talk probably 10 plus years. I think that's where this game is gonna start literally learned and shown to the market that nuclear power plants could be built. Mu we're probably mostly talking about modular reactors. But I have this experience, right? In my experience, this one of our sites

in Georgia is actually co-located through power contracts with the only nuclear power plant that has been built in America in the 21st century. The power plant is called Woggle 2. Woggle 2 is over budget of billions of dollars and took 10 years longer to build. Right? So I literally have an experience when I was told that we're gonna have it next year, next year, next year. It took them like a few years to get it online.

Vit Lyoshin (10:01.171)
Mm-hmm.

Vit Lyoshin (10:09.928)
Yeah.

Sergii (10:18.246)
after they said it's built, by the way. So so we we we have to kind of break down the problem in this timelines. but in reality if you if you're if you question or you interest my opinion more about will there be a different innovation in power generation? Absolutely. Because with this demand not only on AI but no in general the demand in power is changing I think there will be new

Power businesses coming online. Like we've heard about fuel cells. We are hearing about natural gas power plants. We're t we're talking about, you know, new startups that even come up that they say, hey, we're gonna generate power differently. That's an innovation curve. And that's very good because well, if you're gonna break lots of things, it's one thing will work up. And I'm sure that nuclear is the path to go because it's well, it's a long term power, but it's scalable.

Vit Lyoshin (11:16.369)
Mm-hmm. Yeah, understand. Okay. And then going back to chips, not a problem with chips because these companies are gonna be keeping optimizing them, keeping or keep printing them more, from what I understand. Is that right? Or are there any other kind of solutions to that that it's not creating a really big issue that all people are talking about? It's probably more like a visible issue issue to people.

Sergii (11:44.3)
Yeah, well look, chips, chips is not a problem. The price to pay for chips is a problem. there the equipment is always available. It's always available. It's just how much is it gonna cost? Well, Nvidia's job, Nvidia's Qualcomm, I mean right now Google is getting to this game. The job is to be printing as many new new innovative technology chips with new attack in them that's expedites this speed of.

And they're gonna do it as long as the market's gonna buy it, buy it. And but my market is buying it because there are use cases, like in tropical open AI, G GPT product is is one of the use cases there. And people are like thinking it's like, my god, this is great. We're gonna just right now compute costs so much because the chips are expensive. But when we're gonna get to a level where the the curve is gonna flatter, our compute not gonna be that expensive and the use case is gonna be even more. So I think.

The supply of chips is not a huge problem whatsoever. We have had only Nvidia as a big kid on the on the block, the you know, multi-trillion dollar business. well now people see companies, big big companies see that there is a there is a a huge market and they want a piece of that to business, so it creates competition. Well, look, Google is printing chips right now, they're making TPUs, Amazon Web Services creating their own chip.

Vit Lyoshin (12:56.276)
Mm-hmm.

Sergii (13:13.25)
Qualcomm is creating their chip. AMD is in the game. So now we're just seeing more and more companies. Well, Cerebrus, for example, is is a company that's now building also a competition. Grok a company, not the Elon Musk Grok, a different Grok. It's a company that builds chips, ASICs actually, was acquired by Nvidia. So now we like the big demand creates supply, and then more and more companies gonna be

participating in this market as long as demand persists and demand gonna persist as long as there are use cases.

Vit Lyoshin (13:49.072)
Mm-hmm. Yeah, that makes sense. Okay. And then let's shift a little bit to actual AI data centers. Like can you break down like what it means and what it looks like and what's the infrastructure look like for a an ID AI data center?

Sergii (14:06.7)
Yeah, it's it's a it's a good question. You know, the answer to this question is rather simple. It's a building. And and and look well the building with some systems and why I bring this why I kinda say it this way, 'cause like when you open the news and there is so much negative information about the AI data centers, some people don't even know what they look like. And they think that this is some matrix style of you know

Vit Lyoshin (14:15.579)
Right.

Sergii (14:35.702)
Flying in the sky and they're gonna suck out energy from you. But anyway, w the reality is data traditional data center, AI data center, that's that's literally a building. It's a big 150,000 square foot building, for example, which has which has backup generators, big backup generators next to it. those generators are usually part of a utility room. Utility room is like you this is in your house, it's there is a generator.

I mean some houses don't have a generator, but for example, there would be a generator, backup generator. Then you're gonna have a pumping station, the station that just keeps your AC. It's pretty simple. Like you you have an AC, a chiller, AC unit per se. but that's a that's a chiller. In data center is gonna be a chiller, a big chiller. You can see those chillers while for example, when you're shopping and they are next to Walmart, targets, you can see them all all everywhere. those are chillers, they're usually mounted on on the top. And

Vit Lyoshin (15:08.853)
Uh-huh.

Vit Lyoshin (15:33.023)
Uh-huh.

Sergii (15:33.462)
Then you have electrical infrastructure such such as transformers and a re electrical substation where power is interconnected. Well, inside it's a clean building with racks where the computer's sitting, consuming power and being cooled. That's honestly it. what is important about it, all of this has to work as a single system. is what's important is an uptime. it's important if there is no power in the rack.

the the there should be a backup. That's why we have generators. But in reality, data centers and AI data centers have existed for decades. They're not new. it's just the demand of power that supposed to be in this data center is new.

Vit Lyoshin (16:19.091)
Yeah, it it just becoming more mainstream and news making a big problem out of it, I guess. Or maybe not a problem, but a lot of people complaining. Like I heard in the news the other day, somebody's trying to build buy a bunch of farms and build data centers. People are not really like that, or some other like deals like that that are really controversial in the media. And but

Sergii (16:44.27)
And I'm glad I'm glad you said you said there's some people trying to buy. I think the media's message is like that someone is taking advantage of someone. Well, you know, in the United States is an op open, open business country, right? Like I come in, I want to buy the land, I want to build something there, I should be allowed to do that. All right, like it's I'm not stealing that land. And that's the issue we can see everywhere. Is like, well, we we're against the data centers. Well

Vit Lyoshin (17:03.861)
Yeah.

Sergii (17:11.628)
Someone bought the land and if as long as the data center or that building is in w is being built in accordance with the local rules, there shouldn't be any controversial part about it. You you have to play any when there is a Walmart being built, right? Or something like that.

Vit Lyoshin (17:24.179)
Yeah, I agree.

Yeah, I guess they're more concerned about like, you know, some people think that electricity is creating some issues in the area, like those huge power lines that are nobody wants to live next to, plus the data centers will suck out all the energy from the neighborhoods and the grid is gonna break, or what whatever the excuses are there. So it's just not enough confidence that the infrastructure, the grid will support all that demand.

for the energy.

Sergii (18:00.172)
Yeah, well well, correct. But that's the biggest myth, right? That the power gonna be sucked out from their homes. That's not how electrical grid works. Electrical grid works that there is a generator somewhere, and this generate generation is get to getting to the transmission lines, and then there are end users. And if there is no use, well this power is being transferred somewhere else. And at the end of the day, it's not that this power is specifically all allocated, it's not a zero sum game, right? It's not allocating only to one home and if some some other home comes as

Vit Lyoshin (18:26.559)
Mm-hmm.

Sergii (18:30.196)
online that would steal power from the other home. It's like you you have a house or a condo, where you live is like if there is another neighbor, it doesn't mean that he or she takes your power. You still have the power. It's just more power gonna be built. And and that's that's just the craziest argument to me that the grid gonna grid gonna be out of power. That's the whole point about r this this new trend where the

Vit Lyoshin (18:44.82)
Yeah.

Sergii (18:58.542)
Power companies are doing lots of deep studies to be able to estimate if a data center can go online in this area. And those studies are pretty expensive to do. and we are doing them, and other companies are doing them for that particular reason. So we are actually very certain that the data center that will go online is not gonna cause outages where the utility has to run blackouts or brownouts in that area.

Vit Lyoshin (19:08.927)
Mm-hmm.

Vit Lyoshin (19:26.505)
Yeah, yeah. And just something a little bit off topic. I had a while ago I had sustainability scientist on the podcast and we were arguing a little bit about like some things people can do to save electricity, right? And like I see all the time lights going during the day all around me, everywhere I go, in on the streets, in other houses, in other buildings and everywhere.

And if people would be a little bit more cautious about that and a little bit more mindful of that, we could pro potentially like offload a lot of pressure from the grid and save a lot of power and redirect it to where it actually needs to go, rather than just burning all these light bulbs all day long, twenty four seven for no reason, other than just keeping them on. So I mean that's crazy. So

Sergii (20:18.2)
No, absolutely. Well, that's grid management, right? That's a big thing. Power management is a huge thing. while our company is participating in demand response programs across the country. We operate across multiple states, and we are that a big load user which turns off when the utility sees the overload on the grid. So as a matter of fact, the data centers

Vit Lyoshin (20:22.772)
Yeah.

Vit Lyoshin (20:43.465)
Mm-hmm.

Sergii (20:46.914)
Be it crypto data center or A data centers, they are good for the grid because instead of shutting down thousands of people during the overload of the grid, like for example, today it's pretty hot, you their AC is running all over the place. Instead of shutting down the people's loc people's houses electricity, you can shut down one user that is flexible to participate in that demand response. And that user would get

would get reward for that. Well that's why we have been in this space for quite so long are actually so much surprised when when there is a big pushback saying that util utility customers like our ourselves, Bitcoin miners or data centers are bad for the grid, which is a completely in reverse, completely opposite from that, because we are good users and we are flexible load users, as well as we are base load users.

Which allows the grid to plan better their load across their whole system during the day and during the week.

Vit Lyoshin (21:54.632)
Mm-hmm. Okay. Yeah. And also the idea of modular data centers. Can you explain this a little bit? And like I I hear that the demand is growing for those and what they look like basically.

Sergii (22:09.57)
Yeah, modular data centers is a new thing in AI, but it has not been a new thing in Bitcoin mining. those are imagine a forty foot container and ha that has electrical infrastructure in it. modular data centers for for AI would be a little bit different. They would be bigger type of containers, but the whole idea is that the elect the whole infrastructure of a data center is built somewhere else.

Vit Lyoshin (22:16.873)
Mm-hmm.

Sergii (22:37.75)
And then it's delivered to the site in a form of a module. So you you have like Lego blocks. So you develop the site from blocks that are already built versus you start from the scratch at every site at the same time. So that's the modular aspect of it. They could be different sizes. So for example, like you can have a generator, it will be in a in a in a 40-foot container. the chiller can be in a 40-foot container. The whole white space for data center could be in a 56-foot.

Vit Lyoshin (22:44.394)
Mm-hmm.

Sergii (23:07.61)
container so it's a little bit bigger. But all of this data centers will well all these modules for data centers they're gonna come in from different different parts of cu from a country and then they're gonna be installed together in one location. So what that does is that allows the developer to develop a site faster and have some some modules in stock that are fungible. Like for example

You can have hundred or twenty whatever number generators sitting in stock which are the same for any site anywhere you're gonna deploy. If it makes

Vit Lyoshin (23:46.45)
Mm-hmm. I see, yeah, like a spare leg of blocks you can buy and plug in. So who are the customers of those? Is it like smaller companies, bigger companies, or individuals?

Sergii (24:00.3)
Yeah, well the customers of those gonna be smaller neo cloud companies, the data center users, such I mean any neo cloud, new new type of cloud service providers who are trying to hit the market, who want to get into faster, faster to the market, they would be develop well, they will be working and signing the deals with with someone like us w where we can deliver

the white space faster versus for example, for who could be a big developer, like QTS, for example. QTS, digital reality, they will not be building modular data centers. They would be building big buildings. It takes time. It takes years. But their customers would be, well, that would be Googles and Microsoft and Amazon Web Services of this world. Well, that's kinda one niche of the business. Modular could be other startups who are willing to get into

this market fast, who have GPUs, who have raised capital and they wanna be online as soon as possible.

Vit Lyoshin (25:04.255)
So actual developers of the models basically, right? yeah.

Sergii (25:08.012)
Yeah, there could be developers of the models, those AI labs, those could be neo clouds, though could be those could be even enterprise customers. Like, for example, I don't know, like for example, HubSpot. HubSpot is a CRM or Salesforce, for example. They are enterprise customers, they they are getting into AI, they are using AI as a technology, as a part of their business. It's not their primary business, but it's a part of their business. So they probably need another 10 to 20 megawatt data center somewhere.

Vit Lyoshin (25:23.754)
Mm-hmm.

Sergii (25:37.162)
And they don't want to sign with QTS or digital d digital realty, they could sign with someone like ourselves and we can develop them a a data center faster than other firms.

Vit Lyoshin (25:47.828)
Yeah, yeah. Okay, makes sense. All right. so will data centers for AI look the same ten years from now, for example? Will what do you think?

Sergii (26:00.246)
No, they would look different. I think the if we still look at the traditional data centers, AI data center factory as a big building, I think the data centers from ten years from now they're gonna be all modular and smaller. I think they're gonna be ten to twenty megawatt data centers scattered across the world. that's when when the fiber, the internet, the the technology itself gonna be more robust and you would not need to build mega giga campuses.

And you will be focused on how to build a network of sites and create a virtual campus, for example. A virtual campus that would be, let's say, a data center in Atlanta, another one in Chicago, the other one in some let's say Birmingham, Alabama, or in some remote area in Kansas. And then the internet speed and technology technology development will allow the data cent all those data centers that I mentioned to talk to each other.

at the speed that is enough to to provide compute to the customer and to user like ourselves. So I think the virtual campuses, that's the future, and that's where data center is gonna be. We've heard this a concept of data centers on the edge. that means like a data center somewhere close to a user. I that idea has not picked up. I've heard about that idea for honestly, like eight years. that idea has not picked up, but now it is picking up because people need to build.

facilities faster. Everyone is realizing that building a gigawatt data center is not feasible. It takes a very, very long time, it takes a lot of capital, and power is just hard to get for in that scale. However, to get 10 megawatt site, 20 megawatt sites is way easier than get a a a a one gigawatt site. It's way easier to get it across the whole country

than get one gigawatt in one location. So I think that's people are realizing that and they starting hunting those smaller sites. So I think the data center of next decade gonna be a smaller, modular site scattered across the the United States and the

Vit Lyoshin (28:13.159)
It looks like it's a trend that everything starts huge and big and then gets smaller and smaller, just like with personal computers, right? Now we have them in our pockets and we used to have rooms, we used to have buildings before that. that's very interesting. Yeah, it's gonna get so much more efficient and faster and and better than it it becomes smaller. Everybody will have data center in the pocket.

Sergii (28:19.95)
It's good point.

Sergii (28:29.568)
It's a good point. It's a good point. You have a computer in your pocket. It's a exactly.

Sergii (28:42.828)
Yeah, kind of thing. Well at least access to data center, right? Well, in order to have access to data center in the pocket, you do have to have a data centers closer to where you are. So that's why I think the data center is gonna be closer to popul populations cities, of course. but they will also be remote. And I think that's the next phase. 'cause like we've heard about like Ashbourne, Virginia. Everyone wants to be in Ashborne, Virginia. Well, Ashbourne, Virginia is all packed. You can't build more data centers anymore.

Vit Lyoshin (28:42.877)
That's it.

Yeah.

Vit Lyoshin (28:57.748)
Mm-hmm.

Sergii (29:09.91)
Well, now people say, well, it's Dallas, Dallas Ford Fourth area. Well, it's gonna get packed very soon. Chicago has been a big data center market, Atlanta. Well, but then what's next? Well, then new hubs are gonna start coming up, right? But people are gonna go after locations where there is power, not where there is data center. The power gonna be a driver. The way I I s the way I whole build the whole business is you bring data center towards the power, not power towards data center.

Vit Lyoshin (29:21.332)
Yeah.

Vit Lyoshin (29:38.025)
Mm, okay. Yeah, that's a good idea. Okay. What about this idea that some people talk, like Elon Musk, for example, they talk about space data centers. What do you think about that idea?

Sergii (29:51.042)
Well, with all due respect to Mr. Mosk, I think that's that's a idea that has no legs whatsoever. and I don't think it's gonna have any w legs at all. not now, not in the future. Well it may be used as a something I'm sure like the international space station does have a data center, maybe well what do we call the data center, right? Like a a rack, a small computer. I don't think that idea has any lags at all.

I don't even know why we're discussing well I kinda do know because we need to have a big Elon Musk needs to big build the pick big picture so his stock goes up. And he has been doing that successfully for two decades and things have been working out for him as even though I mean Tesla hasn't been a very successful car company, but now Tesla apparently is not a s in the car company, so who cares? SpaceX is not only a space company anymore, it's a data center company. but the idea of building AI data centers.

Or any data centers in the space is just just not making any feasibility studies like work, right? it's just because it's very expensive to move things to the space. you would not be able to move this to in the space. Well then who gonna operate it? You need to operate a data center. Well, it needs like like think about this. When I sell a data center, when I'm trying to sell a data center somewhere in Kansas, people first they say,

It's too far. And I'm like, well, if Kansas is too far, then space is definitely further. Right? That's the whole thing. It's like, well, first of all. Second of all, why do we need to do that? You do that. Well, like, you don't like it's like you do something that is worth doing economically. You don't do something if there's no mo motivation. You can build more power generation.

Vit Lyoshin (31:24.359)
Right.

Yeah.

Sergii (31:46.358)
here on earth than anywhere else. And I mean on earth, if we talk about this as an earth, not as a country, there are so many different countries that are on the tap to build a generation. You know, you you talk about any oil gas oil and gas rich country. You can build so much generation. Like Africa. In Africa, there has been a trend building big dams to energize Africa. And I'm talking about like many countries, not just the whole continent. And

there is so much more potential to build power generation. and now even if we we touched it briefly with the nuclear power plant, well you can build serious nuclear power plant capabilities in the United States, in the world in general, that will be able to generate very cheap emission free energy. And that's where the world will be going. it will not be going into solar in the space. I mean you can use solar in the space to help

space missions of course, but not data centers. I will die on this hill of this argument.

Vit Lyoshin (32:49.801)
Yeah, may maybe something else. Okay. No, it's a good argument. I mean, I I puzzle myself with this argument. Sometimes I see it, sometimes I don't. And I think, like, exactly what's the reason? Is it cooling? Is it closer to the sun for more power? What is it exactly? And then the argument goes, Well, those solar panels are not that much that much efficient yet. So the technology needs to improve. And I'm kind of on the fence and thinking like

Well, we'll see. We'll have to give it maybe twin.

Sergii (33:21.016)
Yeah, it's it's a cost per kilowatt. It's cost per kilowatt of development, I'll be honest with you. That's what it is. That's the math, right? Like the way I I look at this is like you go to a grocery store, right? And you buy apples. You you buy an apple and it costs a d two dollars per per pound, for example, right? Well, if the apple gonna cost fifty dollars per pound, you're gonna like, damn, I'll better e I'll better eat peaches, you know? Like it's like apples is too expensive for me, right? So it's the same here. It's like how much does it cost

Vit Lyoshin (33:24.181)
Mm-hmm.

Vit Lyoshin (33:42.558)
Yeah.

Ha ha ha.

Sergii (33:48.526)
Create one megawatt of a data center. In Kansas, for example, it's gonna cost ten million dollars. And let's say in Chicago it's gonna cost twelve. In in the orbit is gonna cost a billion or whatever that number is right now, right? But it's just well, it's I'll put it this way, that number is out of this world.

Vit Lyoshin (34:04.073)
Yeah.

Vit Lyoshin (34:10.421)
Yeah, okay. so we talked a little bit about this idea of centralized big data centers versus small ones. Your argument is we're moving towards probably a little bit more distributed and things like that. what about AI in general, do you think the actual compute power of it or compute capabilities of it?

Are they going to get also a little bit more distributed? Like, I maybe I'm not asking this clearly, but I basically read about a few startups, a few companies right now trying to invest in something similar to blockchain and decentralized kind of like compute on almost on your personal computers, if you will. And I'm not sure if you how much you're familiar with that idea.

Sergii (35:02.143)
Yeah.

Vit Lyoshin (35:06.537)
that people can have like really like an almost like a nano data center so nano compute nodes and whether it's have any legs and if feasible for the future.

Sergii (35:11.97)
Yeah, yeah.

Sergii (35:17.314)
Yeah, it's a very good point. I do know this company called Gonka AI. they have been doing something like this. I think this idea doesn't have any legs. It's it's great in concept. it's who gonna finance it. I think this is the and the reason I say that just because I've spent the last decade in cryptocurrency world. And that's where all this distribution has been around and the idea thesis of distribu distributed, democratic, all that. It's all BS. It's bullshit.

Vit Lyoshin (35:36.319)
Mm-hmm.

Sergii (35:45.846)
No one wants any distributed. Everyone wants to get just get rich. This is the the all ultimate goal of everyone all the time. Be it a Vitalik Butarian, be it the founder of Solana, it doesn't matter. What matters is how is all make makes money. And I be vel I may be very kind of blunt here, but that's how it is. there has been an idea of a distributed money. That's probably the best idea out there that exists, right? The government can't control my money. It's failed. It failed.

The only big use case of cryptocurrency we got is a stable coin. An industry for a decade that's been around is a stable coin. Any idea of a distributed storage, for example, file coin, there was a distributed storage idea, failed. Now we can see super not supermicro, excuse me, micron printing billions printing like $360 a day a million dollars a day in profits just because they issue they built the the chips for for storage, right?

Filecoin is down. Why is that? Because no one cares. Everyone cares about about micro micron stock because they actually make physical product. I think the idea itself is is is glorious, it's great, right? But at the end of the day, where's the the use case? Where's the market going? Right? Who gonna fund it? I'll tell you who's gonna f who funds AI today. That's entropic, that's

Vit Lyoshin (36:57.907)
Mm-hmm.

Vit Lyoshin (37:09.663)
Mm-hmm.

Sergii (37:14.572)
Chat GPT OpenAI, that's Google, that's Microsoft, that's Matter. That's who funds it. That's about it. And it's just that huge chunk of cash is going down from that mountain. And there's like as a snow that's melting, and the rivers are going down, down, down, and they are kind of distributed across the world, and some people get a piece of it. Right? Well, the moment the mountain is going to start stop melting and there will be no water down, there will be no money. It will dry up. The rivers will dry up. So the whole idea of

of distributed something, it's great, it's novel. But I don't think it's feasible. Because in reality it's all about the money. Who gonna finance it? Is this is this an altruistic project? Then probably yeah it's a great idea it's gonna work but if it's a project that needs to make money it's not gonna work. And I'm gonna be a party pooper here unfortunately for my critical listeners and blockchain listeners. blockchain has failed as a use case, as a product.

Vit Lyoshin (37:52.615)
Mm-hmm.

Vit Lyoshin (38:06.153)
Ha ha ha.

Sergii (38:12.706)
And I'm I am the guy who spent decades in this space. I've tried it looking at from different sp different angles. It's just failed. Bitcoin even failed, unfortunately. A great novel idea. Story is just amazing, but it did fail. So I don't think it's gonna be implemented in the AI. Because in the AI, all what I want, I want to sign a contract with someone like Meta or someone like Google, so I know that I'm gonna get paid for the next decade. That's about

Vit Lyoshin (38:39.163)
Mm-hmm. Yeah. Okay. Yeah. And maybe initially it works with really enthusiastic people and it kicks off and then it dies out when when the hype goes away and people can't really put it to use. Interesting. Okay.

Sergii (38:54.604)
Yeah. No, ideas themselves, they they are not ideas are great, right? Like but in this world today, like look look look what look what's happening with with with ChPT with with Claude with other models. I mean, you can test any idea within seconds. Right? This is crazy. Before you need to get put together in a room, five people, call your friends, find some smart people, or a beer and discuss this.

Vit Lyoshin (39:05.066)
Mm-hmm.

Vit Lyoshin (39:12.287)
Right.

Sergii (39:21.568)
Now you can just test this idea and just say, well, nope, it's not gonna work. You continue, move forward. So that's the whole thing, right? But that's actually the value of this technology, right? That it can test those ideas so fast and help us out to fail, not to get on the path that is by design is going to fail. That's how I see it. distribution, distributed computing, the only distributed computing model I see that's physical distribution, that's

Vit Lyoshin (39:27.155)
Ha ha.

Sergii (39:49.816)
I'll repeat this. That's a data center in Kansas, that's a data center in Mexico, that's a data center in Canada, that's a data center in Ireland, in France, and all in like you know, in Africa somewhere, in in Nigeria, for example. That's a distribution of AI that I believe in. But not something on a blockchain. You don't have to, you know, you don't have to complicate things.

Vit Lyoshin (40:03.957)
Mm-hmm.

Vit Lyoshin (40:09.705)
Yeah, yeah, okay, yeah. Usually simple as asked answer is the best one, right? Yeah, I kinda agree with that. Yep. Okay. So what what do you think are the biggest, biggest misconceptions about AI infrastructure today that people have that maybe we haven't talked about it yet?

Sergii (40:28.236)
Well, it's a very good question. There there are a few. First, the I data centers using a lot of power. It's not really true. Not all the I data centers using a lot of power. Excuse yeah, a lot of power. the the they are using a lot of power, but they're using the power that is allocated to them. That's no one else is using. And first of all, I think the word using is bad. I went through this in Bitcoin mining world. We're not using power, we are buying power. So I'm buying power.

Vit Lyoshin (40:35.165)
Right.

Vit Lyoshin (40:48.178)
Mm-hmm.

Vit Lyoshin (40:56.925)
Right.

Sergii (40:57.934)
Like, right? Because when you say use the word using, it's like it's like a common thing, right? It's like, well, I use the road, right? What's so it's a common drive road, right? So it's a common road. And if too many people drive through there, you get upset because you stay in traffic, right? Well, here we're not using someone else's road, we're using our road that we built and we paid for it, right? So that's kind of the thing I think it's important to understand. Second, the big mis misconception, in my opinion, is water.

Heard it all over the place. AI data centers are using a lot of water, using a lot of water. Same story. First of all, they're not using lots of water because of the technology. You have direct to chip cooling technology where water is cooling the chips, it's a closed loop system. it's really fascinating technology. We're getting too early into this, but I think the next tech is gonna be only direct-to-chip cooling. The reason to that is because the compute is distributed, and we don't want to have

limited factor where we deploy. We want to be able to deploy it in Arizona, in Kansas, and in Wyoming, and let's say in in in Sahara, for example, in in the desert. So you want to be able to deploy it everywhere where there is power. Well secondly, on the power on excuse me, excuse me, power and where there is not dependency on water. And then the water the water fact is is like well if you're using water that means there is water for sale.

Vit Lyoshin (42:08.073)
Uh-huh.

Sergii (42:26.828)
You buying water as well. So that's kind of I think the big misconception of this, of this, that we're using someone, someone else's thing. I had I had conversations with local utilities in different states where they say, Can you please buy our water? And I'm like, that's that's a fascinating thing to hear because most people say they want to use their water. It's like, well, listen, we have raised taxes in this county.

Vit Lyoshin (42:36.436)
Mm-hmm.

Sergii (42:57.624)
To raise money to build a water treatment plant. So now we've built the water treatment plant, and we're using only 10% of capacity of this water treatment plant, and 90% of this capacity is sitting idle. But we already took the money from the from the people, and we cannot pay them technically back in a form of goods and services, in a form of better development of the community. So if you

Vit Lyoshin (43:21.129)
Yeah.

Sergii (43:25.912)
Could buy us water from us, another 50% of all water that we already have. We will not need to raise taxes to build a new school, to build a new road, to do an upgrade, to create a park. And that's kind of the thinking should we have. Because when we're kind of thinking about this thing, we're using someone else's thing. Of course. What are we buying for it? Right? So that's a misconception number two, I think, a big one. And third one is well, that AI, AI,

Vit Lyoshin (43:40.381)
Yeah.

Sergii (43:56.408)
data center is is bad for our community because it has some you know unhealthy waves and some type of aura there, whatever. That's the craziest part, right? Well, there's no scientific proof of that. Right? We have to talk about this. Can we say it was like, well, it probably causes because that's what we've hearing. It's like it's gonna cause cancer. It's like, well, I mean

Where's the proof of that, right? Like at the end of the day, I mean we can call and say whatever, but we need to have a proof of it. So those are three big things. And I I think they I think the data center world, well, the tech world in general, kind of didn't pay attention to that. And that those misconceptions, those myths became so much widespread that they became the truth. You know, it's like the legend is as good as we believe in it already. You know, it's like, well, if

Vit Lyoshin (44:48.809)
Mm-hmm.

Vit Lyoshin (44:52.52)
Yeah.

Sergii (44:53.698)
Well, it's probably it's true, right? It's like, well, if everyone talks about it, probably it's true, right? But it's just like common bias. Common bias is a dangerous thing because if we all agree on something, even though it's a wrong thing, we may think that that's the truth. and we see crazy stuff. I think people have to give benefit of a doubt. They have to understand that.

Vit Lyoshin (44:55.995)
Everybody's talking about

Vit Lyoshin (45:10.558)
Yeah.

Sergii (45:22.368)
AI data centers or any data center or any development is a good thing, especially in the areas where there is power and there's not a lot of going on. I think those communities have to take advantage of it. Because, in my opinion, this is the only time in history where money from Wall Street, from Silicon Valley can go towards very small undeveloped communities in America. They would never go otherwise.

Because usually that money stays in Wall Street and in Silicon Valley. But here, this this data center spent can go to rural c economy in Kansas. It will go into rural economy in Alabama. There's no other other industry that would do that. well there are other industries, but in general, this type of spend from those in from from like Wall Street and Silicon Valley could go anywhere else. So I think the local communities have to embrace that and take advantage of it, to be honest.

Vit Lyoshin (46:19.721)
Yeah, I think it all comes down to like educating the public because we heard the same things with nuclear powers, stations, right? Or whatever they call. we heard the same with AI. In fact, last September I was in San Francisco at the AI conference and it was really interesting contrast because everybody is so enthusiastic at the conference, but then you see these protesters outside screaming

on the top of their lungs that AI is all gonna kill us and Doom's machine is coming, the Terminator and whatever. And it's it's like it's crazy. I'm looking at these people and these people and I'm like, my God, what's going on? They're gonna start fighting with each other or something. And right. And and it's the same thing here with AI data centers now. Somebody's like miscommunicating this thing and people getting scared.

Sergii (47:01.625)
yeah. it's real. It's happening. It's really they are fighting. really fighting.

Vit Lyoshin (47:15.229)
like somebody who's non technical, never dealt with electric like me, I never dealt with electricity or physics of it and stuff like that. And I'm curious, like what's going on and trying to get to the bottom of it, like how bad it is, how good it is, or whatnot. Interesting.

Sergii (47:31.856)
yeah, no, absolutely. And the idea of a data center is actually like killing us all, I think, I think there is some well, there is some risk to that, right? But we have to think about this not from a perspective prov from a perspective as an isolated incident. We we can say, well, it's like we're gonna stop doing innovation here in America. Well, let's say we're gonna stop a data center in that well known Kansas that we already discussed.

Well, it doesn't mean the data center development gonna stop somewhere in Shenzheng, China. Why? It doesn't mean that it's gonna stop there. It's gonna be still being developed, right? Well, in that we're gonna have a situation, well, it's basically who's AI gonna win? Is it China China's AI or is it American AI? That's the reality. So I mean I'll take American AI all day long, every day of the week. Right? And that's how I look into this and that's how people have to think about it. They have to think it not as an isolated thing.

Vit Lyoshin (48:04.479)
Right.

Sergii (48:30.038)
Because if it's not done here, it's done somewhere else. And because this is the world it's going. The tech cannot be ignored. It's like we cannot go back. I'll give you a quick example. I was working with Claude last week. And I was working with Claude on a presentation. I was working and was great. And then at the end of the week, I started working on this presentation again and Claude was not doing the that good. I was like, what is this guy? Why is not he's doing so good? It was so good. Well now, just think about it.

Vit Lyoshin (48:44.265)
Mm-hmm.

Sergii (49:00.064)
My thinking is not just like, well, if Claude is not doing well, let me go and call the designer that I knew and let him do it better. Now I'm just trying to figure out why the Claude is not working. So like I've tested it, I've used it, I'm not going back. And it's just as a matter of fact, when this gonna get to to a point when everyone did it. Well I'll tell you this, my mom, she's a teacher in a small village in Ukraine, she uses AI daily, she uses Chi GPT to create

Vit Lyoshin (49:08.221)
Mm-hmm.

Sergii (49:29.438)
study materials for the kids. I mean, if she's doing it there and I'm doing it here, I think the technology is adapted so much that you can't just get rid of it anymore. For that we need AI data centers.

Vit Lyoshin (49:38.635)
Yeah. No. No, I totally agree with that. You cannot stop the progress. It's always gonna push forward no matter what. More people will adopt it than people who not, and people who are not are gonna just be left behind. That's always happens. Yeah. Yeah. So something I wanted also to ask you there's a lot of people, startups, entrepreneurs, developers building something with AI.

Sergii (49:55.97)
Fortunately it does.

Vit Lyoshin (50:06.766)
And just advice from you being on the infrastructure side, what should be they might be mindful of when they build in their products that maybe new model or maybe utilizing existing model and things like that? Anything any advice for them to like how to be more maybe saving the power mindset type thing?

Sergii (50:26.188)
Yeah. Well, it's a good question. I don't think I I can recommend on saving the power power, but what I think the what they should be mindful for of is planning ahead. I I mean it may sound crazy, but software moves faster than hardware. That's just fact. And unfortunately, people who build these models, they want the hardware, the physical world, to move at the same speed they're building those models. And it's very hard.

Vit Lyoshin (50:45.718)
Mm-hmm.

Sergii (50:54.354)
And that that's what we facing this huge gap when people come out come over to us and they say, Well, we need this data center to be live, like in three months. And we're Well, we just can't do it physically. It cannot be done. The speed the speed of cars driving is just not at the lighter speed. That's what it is. It's physics laws, right? So so so that's I think the my mindfulness in in terms of kind of connecting the gaps here.

well are probably narrowing the gap of understanding how fast, right? Because there's of course AI arms race right now. And it's like we need to roll out the model, we need to boot tomorrow, it needs to be online tomorrow. We need more compute, more space, more this, more that. But it would be way easier if we really could manage expectations and understand what is that compute needed. If the customer came in to me and showed me the proofs that they're gonna have need a fifty megawatt facility next year.

with the proof of course, because everyone is hopeful and everyone is excited, then we on our end would be able to kind of start putting together a pen to paper and figure out how it to we can provide the the supply of for that demand. But if we receive a customer already figure everything out and they say, Well now now I have the GPUs, I need it to be online. Well it's very hard to meet. And that's when you're very fast, you're making mistakes, you're making

projects not to the fullest quality you would want and that's when the projects are becoming a risk and they can fail.

Vit Lyoshin (52:33.512)
Mm-hmm. Okay. Makes sense. Okay, Sergei, thank you very much for your perspective and answers. It's been very insightful. I learned a few things certainly. at the end I usually have my innovation QA for every guest, three questions, and it's just my personal library of answers. I I'd like to hear your perspective. So can you define innovation in a few words?

Sergii (52:57.1)
Yeah, sure. Thank you.

Sergii (53:03.138)
Wow, it's this is a very good one. Innovation in a few words. I think I I I wish I had ChGPT open here. but I think innovation it's it's hanger for the new it's a hunger for the new things.

Vit Lyoshin (53:12.276)
Ha ha ha ha.

Vit Lyoshin (53:25.13)
Hmm, okay, interesting. the second question is which innovation in the human history you think changed the world the most?

Sergii (53:35.316)
It's a very good one.

boy. I honestly, honestly, like I if you asked me this question six months ago, I would say that would be energy and electricity. I would say it like no question about it. Without electricity, the world dies. Stops. Okay. I think AI data centers, or excuse me, AI as a tech is on par with that right now. I think in in a five, ten years we're gonna look at this as they say, holy cow, this was Pivotal.

Vit Lyoshin (53:52.381)
Vit Lyoshin (54:09.29)
Yeah, yeah. Okay. And the last question is which technology or software or tool or something that we use today we will be laughing at ten years from now?

Sergii (54:22.168)
Very good question. that's a very good question. well on a personal on a personal use, I think the calendar. Well we'll probably never gonna use it. it's the calendar, the meeting invites and all of that stuff. I think AI gonna be just handling all of it. Ch GPT already created the task manager like an assistant. So I think anything assistant assistant wise, reading an email, like I think that's gonna go away. And that's my dream.

Vit Lyoshin (54:42.294)
Mm-hmm.

Sergii (54:50.836)
If if AI cannot fix me, not re fix the email, then we failed. I think Yeah, i i I think reading an email, calendar, all this stuff, like scheduling, 'cause that's where the where a lot's time is spent, right? Like communication, like like, what about this? Can I follow up with this? You know? So I think that's where we're gonna literally be laughing. Remember those times when we used email ha ha ha.

Vit Lyoshin (54:58.134)
That that's your benchmark, okay.

Vit Lyoshin (55:08.832)
yeah.

Vit Lyoshin (55:17.566)
Yeah, yeah. All right. Okay, great. Thank you very much for your time. Hope we can stay in touch. So

Sergii (55:23.904)
Absolutely. Absolutely. Thank you very much for having me. It w it's been it's been a pleasure having this conversation.

Vit Lyoshin (55:29.524)
Yeah, absolutely. Bye.

Sergii (55:32.386)
By the way.