Can Data Centers Deliver on The AI Promise?

Hey, amazing job, Luz. Thank you very much. Thanks for being here, and I love hearing that there is front-row energy.

​

We will start with some brief introductions. Thank you very much for joining us. I was raised by a Southern mama, so I can't help it. I'll start with Francesca. Francesca, you were named to Forbes Italia 30 Under 30 and recently received an award as a top entrepreneurial person in Italy. That's amazing. Would you tell us more about your background?

​

Thank you. Yes, I founded ALPS in 2018. We are a Bitcoin mining company. We have operations around the world. Our biggest one is in Oman, where we have 150 megawatts. Then we have two sites in Iowa, where we have 70 megawatts. Then we have South America between Paraguay, Bolivia, and Ecuador. So we are quite geographically diversified around the world.

​

We started from Europe, from Italy, where our headquarters are, with hydropower facilities. We are an infrastructure company. We build mining farms, and we build data centers.

​

That's amazing. Thank you very much. Nir, I don't have any other rhythm, so we'll just go down the line with introductions. I've been a big fan of MARA for many years, and one thing interesting about you is your background as an aerospace engineer, and that you also have paragliding as a hobby. Tell the audience more about who you are and what you do.

​

Sure. Thank you. Glad to be here. I joined MARA about a year ago. For those who don't know MARA, it is quite a big Bitcoin miner and is emerging as a digital infrastructure company. Personally, my background is in robotics, aerospace, and autonomous vehicles, so kind of the intersection of hardware and software. I realized that the emerging need for intelligence and democratizing it is something that requires energy, and that led me to MARA.

​

That's amazing. Aaron, I'll give the floor to you next. I also enjoyed learning a lot more about Hydra Host and the impressive work that you do, as well as a little bit about your background. Would you tell the audience more?

​

I'm CEO and co-founder of Hydra Host. We help data centers essentially make money off of their megawatts. We are a software company that automates the facility to make it possible to access AI customers and resolve all the components needed at the data room level to make it much easier for them to bring their compute capacity to market.

​

You can think of us as a headless NeoCloud. We are the largest NeoCloud that actually doesn't own hardware. The data center owns the hardware and gets the upside. We are a management company and a software company that helps accelerate the adoption of their infrastructure.

​

That's fantastic. Thank you. Xander, all the way down to you. I had the real privilege that you and I were able to share a meal together in Los Angeles two and a half years ago, authentic Sichuan hot pot, one of the hottest meals I've ever had. I love spicy food, and I was sitting there holding my breath trying to stay alive.

​

Xander, I'm a hardcore Bitcoiner, and you are famous in our industry, so it's a real privilege to be on this panel with you. I understand not only have you been building Zillion Network, but you're also launching Tesser. I hope that you'll tell the audience more.

​

It's really my pleasure being here, meeting good old friends. My company collaborates with Hydra, with Aaron's team, quite a lot. I'm Xander, founder and CEO of Zillion Network. Zillion is an AI compute company. We aggregate and feed GPUs and lead that to AI customers, the AI offtakers.

​

Before that, I built data centers, traditional tier data centers. I also took the role at Bitmain as the global mining data center development head. We acted as the designer, but also the offtaker and even backstopper to some big projects. Now we are launching a sister company named Tesser, which is to build and design tessellated AI data centers to speed up the adoption of AI.

​

That's fantastic. I am going to take the first question back to Xander, but then we're not going to go down the line with the same question. I really encouraged this panel to engage with each other. Obviously, you heard from their backgrounds that there is tremendous experience for this exact topic.

​

Xander, would you do two things? Would you start by painting a big picture of what is going on in the AI data center buildout space? What kind of large capacities are really happening? And then, is it actionable? What's actually out there that can be built and developed?

​

I think this industry is amazing. AI demand is booming. Based on McKinsey research, by 2030 there will be $7 trillion invested in global data centers, and $5 trillion of that will go into the AI part, the AI data center. There is also more than 150 gigawatts of demand aggregated, and you also see in the news that $300 billion will be spent by hyperscalers in AI this year. So there is huge demand.

​

But then the problem is, everybody says we have a shortage of chips and power. Actually, in my mind, how to turn the power into actionable AI capacity is the issue. The AI data center buildout conversion rate is not as big as people think. Everybody says gigawatt-level campus, but how many megawatts have actually broken ground?

​

There is always a delay in the market. You see the NeoClouds and the hyperscalers with projects delayed. From our point of view, whether you get the project designed from day one to be contractable and bankable is the key. I can explain that more later, but that is the big picture. The problem is there: how do you turn an AI project into a design that is contractable and financeable?

​

Can you expand more about financeability and contractability?

​

There are a few things. First, if you do an AI data center buildout, you need to know exactly who your customer is and what kind of AI workload they are running. What kind of GPU clusters and models are they running? Today we are doing Blackwell GPUs. Rack density is already 130 kilowatts today, but it will soon ramp up to 230 kilowatts in half a year's time, and it could be an even higher density.

​

If you cannot design your AI data center to fit that need, that will be a disaster. Also, if you have super long lead-time items, if you have a thousand people working outside, but you have a shortage of labor, by the time your data center is ready, the new generation of GPUs will come. So first, you need to know exactly who your top customer is, and you need to know the timing.

​

The second thing is that you need to understand who your lender is and what they need. Lenders definitely need a high SLA, a service-level agreement assurance. They also want to see repeatable and credible design that they can easily do due diligence on. They can easily see the track record from others and whether this design has been built and met the timeline before.

​

The sequence is that you first need to make the correct design, then make your project contractable, and then the lender will see the offtaker and start to finance the project. If you do the sequence wrong, the project will be a nonstarter.

​

That explains it. Thank you. Aaron, I heard something interesting there that I think you could add context to. Xander led with understanding the type of machines that the customer is running as well as their power draw. I learned a lot about Hydra Host. You have actually solved for that so the data center operator doesn't have to understand all those details. Would you add some context?

​

On the question that we're supposed to be here for, can data centers deliver, the answer is that there is no other option. AI is pressing forward, and it presupposes the need to convert power into tokens. There are lots of questions around how you actually move from, will this land work, will this power draw work, and so on. But the overall arc, to give people confidence about what is happening, is why Bitcoin is such a great metaphor for AI.

​

Bitcoin is essentially democratized access to money. AI is democratized access to intelligence. It's not like, oh, I'm moving my web app to GPUs. This is a completely different way of thinking about how energy can be translated into a good. And that good is inelastic. How do you tell someone if they're fully intelligent or not fully intelligent? If you have kids, do you tell them, all right, you're in sixth grade now, enough intelligence for you? It's an inelastic good. There is no real price on it.

​

We create all these abstractions. Harvard says $80,000 a year is smart, or another school says 60 or 30 or 40. We see a degree. We say being a chess champion. We create all these abstractions because the human itself is actually a really horrendous test subject. We have really bad benchmarking because we're all so uniquely different. We're all skilled at different things.

​

This is why I'm so bullish on this space. It's not like anyone said no to intelligence. Look at your phone. Go look at the number of gigabytes of data you actually delete. It's usually less than 5% of your total data stream, but it's actually something you delete intentionally.

​

These are all great things because it allows compute to move down stack to make it work, just like Bitcoin moved money accessibility away from the Federal Reserve to everybody. Now we have the ability with technology to move intelligence away from universities, elites, or whatever benchmark you want to create, to literally a machine.

​

That raises the profile of so many other countries and, specifically, data centers that previously were not accessible and were not interesting to a hyperscaler or to a tech company. Now they are salient. The profile that miners bring on the data center side is the exact type of cost profile that will win: low cost, no bullshit. I'm the Costco of tokens. That's literally what every single AI company wants.

​

They don't want hyperscaler tier-four magical things, like look what this box can do. They're like, does that cost 40% more than MARA does? I don't know if that's worth it, man. Because you just give me Jensen, and MARA is going to give me Jensen, so what's the difference in your Jensen? Nothing.

​

That is really hard for hyperscalers to compete with. Miners are already programmed that way to say, the only way I win is better costs. Ultimately, that is the end state of this entire space. If you can convert power to tokens at the most efficient rate, you will always win, and everything else is just noise.

​

I'll add to that. I love what you said about energy. If you look at the horizon, the cost of intelligence converges to the cost of electricity. Similar to how telecom companies have enabled us all to reliably consume so much media and content online on our phones, now you're going to expect providers to deliver the same reliability on intelligence. It's like your own chief of staff, your own partner, your own thinking apparatus in a way.

​

You're exactly right that it's deflationary. That's why it compresses down to the marginal cost of producing energy. Telecom and any other major infrastructure provider are all deflationary. Hyperscalers are not deflationary. Hyperscalers are extremely expensive, and the rents that we've been paying to three providers that monopolized compute for two decades, Jensen just drove a bus through.

​

The marketing is AI, AI. That's the marketing pitch. But Jensen literally drove a bus through public cloud. Now data centers that previously were just relevant for enterprise colocation and mining have a completely new market they can go into, and there is a very, very lucrative market.

​

You can thank Jensen. Jensen completely changed the infrastructure topology to where you can be just a power provider, provide it directly, and walk away. It's really amazing.

​

Thank you. Francesca, I want to pull you into this conversation. When people say that we need more data centers for AI capacity work, what are they missing?

​

For sure, one thing to consider is that AI runs on a stack, not only on power. That means it is not so easy to convert power into an AI data center. That's what we have been saying from the beginning, because you need reliable power. This is a really important key and completely different from mining.

​

We as miners are used to building data centers from scratch in bad conditions. For example, we have data centers in the Amazon Forest that are off-grid, and they don't have the reliable environment that an AI data center should have. The first thing to consider is whether the source of energy that is going to be used is the right one to run an AI data center.

​

We are going on to develop renewable energies, but renewables are not stable. That is the problem. All the things that we as miners are doing in a grid in order to be a kind of stabilizer of the grid don't work with AI. So if we convert our mining farms into AI data centers, there will be a problem in the grid to be handled.

​

Power is the first thing that is really important. Second is connectivity. Mining data centers can work with Starlink. We have some examples that work with satellite connection. AI data centers need the right fiber connection. This means that not every source of power and not every available site can be switched from mining to AI.

​

I think there can be a synergy that works. For example, if you have a big facility that is purposed for mining, you can repurpose part of that into AI in order to have only a part of the load that is on 100% of the time, with the right connection and redundancy. For mining, that requirement is for sure less.

​

Yes, there are sites that can be 100% converted, but not all. While demand continues growing, is the infrastructure ready for that? We don't know. If we read the data, there are plenty of GPUs available to be installed, both by the hyperscalers and others, but the bottleneck is that you need the infrastructure. You need the data center working, and building that is not easy.

​

We as miners have a lot of know-how on how to convert power into computing power. But the know-how for the AI part is different. We also need to learn that, and it's not easy to switch. Not all sites can be switched easily.

​

Xander, one of the things I heard Francesca say was about reliable power, cooling, and a few other variables. As we look at whether data centers can deliver on this AI promise, is there a variable that we're missing about delivery timelines? These hyperscalers have an immediate need or a future need. Can we transition? Can you talk to us about delivery timelines?

​

We check the box of the technical requirements, but then we also need to understand their immediate demand for now and the demand for the long-term run. What we see from the market is that people will take whatever capacity is available online by this year, by the first half of this year. But when we ask, if there is any data center capacity online by 2028 or 2029, what is your idea? I get different answers.

​

Most people just say, give me whatever you have today. We have no time to think about two or three years later. That's also my question for Francesca and Aaron. Based on your experience, contract terms really matter. I think the answer is yes, but how much do they matter to your decision to turn a facility into an AI data center? And to Aaron, you have a lot of hosting deals. Do you really see seven-year or eight-year long-term GPU leasing, or just three-year to five-year terms?

​

I'll answer quickly on hosting contract terms, or offtaker demand. The way you have to think about making money or returns is that anything long-term means you push risk to the end customer.

​

Everyone here flew here, unless you're from Vegas. How many of you bought your tickets last minute versus farther out? If you reframe the problem statement away from risk management, which is a bank question, not an end customer question, none of you cared what the financial situation of Southwest or United or American Airlines is. You paid based upon the convenience of when you wanted to transact. If you bought nearer term, you paid more. If you bought longer term, you paid less.

​

If the entire market of the airline industry was driven by you paying three years for all of your travel with a deposit of 20%, the airline industry would be very small. Think about homes. When you buy a home, you generally do that every 10 to 15 years. The market scale is directly connected to the size of the transaction and how often that transaction occurs.

​

If a data center is going to make money, they have to realize that if they're just pushing risk to an end customer, they're not going to make very much money. You're going to have to lower your return profile because you're moving risk to him. If you take on more risk, you can make more money. It's exactly how you look at bookings of this hotel or bookings of an airline, or really any other utilization of any heavy capex product.

​

Data centers have to accept that the way they're going to drive returns is revenue mix. If your entire profile of your book of business is three-year contracts, you're really not going to make very much money. If your entire business is completely on demand, you may make some money or you may not. That's the bet you make. It's the dice you roll.

​

Our answer is that it's a mix. Just like this hotel, the Venetian, where we're all staying, has a mix. It has corporate contracts, long-term arrangements, on-demand, and flex capacity. The data center will evolve into that type of strategy. It will just be forced to iterate.

​

Right now, if you had an H100 cluster that you bought, even at the height of the craziness of 2023, you are making so much money right now. It's unbelievable. But the only way you make that is if you are liquid in that capacity. If you locked it up for a five-year long deal, you're not making any of that upside right now.

​

If I can add something, the obsolescence of the GPU is growing a lot over time. So it's difficult to find customers that sign for computing power long term, because the technology will change a lot during that term.

​

These companies are really focused on, okay, I have to deploy, I need GPUs for inference, I need GPUs for training, and that's what I need for a short time because they are growing fast. But they don't have an idea of their needs over time.

​

That's the issue that infrastructure companies are having. You are investing, for example, 10 times more in infrastructure if you are building an AI data center than a mining data center. But the question is, how long can I have contractual computing power? How can I run my business plan and my model? I don't have any certainty, and that's the big issue right now.

​

Wonderful. Nir, we're in our closing minutes, and I want to ask a direct question. We heard Francesca talk about AI spending more, but if enterprise clients are ultimately going to want to control their own capacity, could you talk to this audience, who might be more familiar with Bitcoin mining, about the gap between the capacity being built today versus what the hyperscalers or enterprise commercial clients actually need?

​

Going back to what Aaron was saying, a lot of the data centers are thought of as infrastructure projects. You can think of them essentially like a hotel. You are building the infrastructure and the capability, and then ideally maybe you have commercial contracts for long term on some of the rooms, and then some otherwise.

​

Then came other models, like Airbnb style, where you might want to have a temporary allocation with a specific unique vibe in a specific area that you want. There is another model, like a furnished apartment. That might be long term, or a suite in a hotel for a long-term stay.

​

Everybody right now is building hotels. But I think the emerging need to have your own control of the data, whether it's sovereignty or sovereign-grade, whether it's the modularity of it to be able to handle planned obsolescence, means there is going to be an emergence around the modularity of the build. The builds that are compatible with the crazy cycle of hardware changes, but also enable you to be closer to customers and what they need and how they want it, while putting it where they want.

​

At the end of the day, humans are physical entities, even though all of us are using the hyperscalers through our applications. Enterprises have a need for control. I just heard an anecdote of a Spanish startup that was building its entire capability, its entire stack and product, on Claude. Claude detected that they violated terms, and they shut it off. An entire company shut down because they essentially outsourced their intelligence to a third party and had no control of the model or the inference engine.

​

That's a challenge. Making sure that enterprises have that control means they might need to have their own data center, or mini data center, or a supermodular one. Essentially, every enterprise needs its own digital manifestation, and that's its access to intelligence.

​

That's fantastic. I sat in an audience at a Pomp event in New York a year and a half ago and made a note to myself that these Bitcoin mining public companies are actually pivoting to AI data center infrastructure, and this is just proof of it.

​

I do think we had a good conversation. We're up to our closing seconds. Please thank the panelists, and we appreciate you being here.