To be a Bitcoin miner, you had to have money, you had to be able to find a plot of land, get access to some power, and cut a deal with a hardware vendor. And then you just had to run them as hard as you could. It's a very different game here. It's like going from bush league to major leagues.
All right, Fred. Thank you so much for coming on Mara Foundation TV. How are you?
I'm great. Glad to be here.
Yeah, I'm excited to chat. Last time you were on, which was our sort of inaugural episode, we sort of chatted broadly about the foundation's goals. But today I'm excited to sort of get into Mara Global's goals a little bit and kind of how the company is evolving in the wake of the birth of AI and HPC, the whole world has been turned upside down seemingly. I'm wondering if we can maybe start by you just sort of giving a little overview of, you know, how this shift is happening, why this shift is happening specifically for kind of the big industrial Bitcoin miners, you know, how did what's going on, you know, for Bitcoiners who might not be following this closely?
Well, the AI industry, similar to Bitcoin, needs a lot of energy to operate. If you go back to the late teens in Bitcoin, the challenge was turning on power to sites so you could deploy miners. And that was the big challenge back then. And granted, Bitcoin miners use a lot less energy than AI does. Though, if you go back again, you know, I think it was the WEF said that Bitcoin was going to burn up the planet and use all the available energy that ever existed.
Still hasn't happened. Now it's AI. But essentially, AI needs power. And if you look at the dynamics in the AI marketplace, it's about market share. It's not dissimilar to Bitcoin. It's, you know, how much energy can Anthropic versus OpenAI versus CoreWeave versus whomever, Microsoft, AWS, etc., get so they can plug in their compute, so they can deploy their models, so they can operate inference for their clients, so their clients can get benefit of AI. And you know, to be fair, up until not too long ago, training was the primary use of AI data center. Now clients are doing inference. Inference is, you know, when you're chatting, when you're running agentic frameworks, that's inference. You're getting an answer from AI. That's what generates value out of AI. Training doesn't give you any real value as a customer.
And so as agentic frameworks were launched earlier this year, you saw a huge uptick in token demand. And you just look at Anthropic sales numbers. You know, they have surpassed OpenAI at this point. Huge uptick. And they are all capacity constrained because every model they release requires a lot more capacity to operate as the number of enterprise customers grows and as the growth within each enterprise customer grows, you have an exponential curve on the demand for tokens. And you know, just look at all the news headlines about tokens or companies are spending way too much money, you know. Three months earlier it was about token maxing. Then they got their bills and it was like, oh my goodness. We got to watch out for this. And all that's being driven by a constraint in capacity.
And so if you have constraint in capacity, it means all of a sudden, where is the constraint point? Well, the constraint point is energized capacity. It's locations where you have land, you have water, you have energy, the energy is turned on and you just need to build a data center building. And the AI industry, just like Bitcoin, has always been constrained by what we call the three C's: capital, capacity, and compute. It was no different in 2019 when Mara was busy growing to become the biggest miner in the world. It's the same thing today in the AI industry. You go raise money, you go find capacity, and you go buy a bunch of GPUs and plug them in, compute. And that's the basis of what's going on now.
So, the challenge is if you were to go do greenfield sites, like some of our peers have chosen to do, well, then you get into queue for power. You have to get permits. It can take you 3 to 5 years. That is too far out from what the AI industry is thinking about today. They need capacity to come online in the next 2 years, or they will lose market share. And so, it's a battle, and they're willing to pay fairly high prices.
When you say, you know, 2 to 5 years out is too far away, like what is the race to get these sites AI ready? Like what needs to happen to get all this power online? I mean, is it just finding the sites, or is it also infrastructure as a challenge? Like what other than just finding the electrons is sort of an issue?
From the hyperscaler perspective, the issue is do you have power connected today and available? Can you draw electrons today?
That's it. That's all they care about, basically.
If you can draw electrons today and there's easy access to internet, and you have water for cooling, then all I have to do is build a building, right? And that's an 18- to 24-month process. That's why I'm online in 2 years. If what you're talking about is a site where you have submitted a request to ERCOT or whichever other utility or grid operator you're talking to for power, that may take you 3 to 5 years to get that power turned on.
Is the difference like similar to the difference between owning the power versus a PPA?
No, because I mean, yes, in a way. If you own the power, you have the power available today. If you have a PPA, you have the power available today, too.
Right? So, if you look at the way Mara operates, most of our sites are operating with some form of PPA. But we have the power turned on today. We're drawing on it today. And so, one of the strengths in our portfolio, you know, these over 1 GW of energized power today with capacity running up now, potentially close to 2 GW of power, is that, you know, that power is available today. So, a hyperscaler can come in and work with us and build a site and the energy is available today.
I see. So, the distinction is not between renting and owning. It's just about do you have access, period?
Right. Do you have a right?
Who controls the electrons is very important today.
And just for clarity, so basically there was a giant uptick in demand for this energy once AI kind of got into the inference, you know, wave that we're in now. Prior to that, when it was just training, there was much, much lower demand. There were still AI data service companies at that point, but now it's like, okay, a race to get as much energy as possible because people are actually using the thing.
Yeah, inference has driven an exponential growth in demand for compute. And so, if you look at the ecosystem of the AI industry, it's not just hyperscalers, but it's people like Nvidia, AMD, Broadcom. So, that's all compute. Then you have cooling infrastructure, Vertiv, et cetera. You have power infrastructure, Schneider Electric, and others like that. And what happens is if there's only so much capacity available, all of those companies' revenues are constrained by the lack of available power. Because if you don't have power that's turned on, you can't build a data center. You can't plug in the compute and use it. You can't have a need for cooling, etc.
And so Nvidia, for example, is very interested in ensuring that they get access to power for their customers because they can't sell more compute if there isn't more power to plug that compute into.
I heard they're even giving incentives and all sorts of things. Yeah, they're doing all sorts of things to just make sure basically. They're like, you know, effectively helping small businesses set up these data centers and wildly subsidizing them.
I mean, think about it this way. If you're a producer of something and there's a constraint to your customers buying what you're selling and that constraint isn't money. That constraint is somewhere to plug it in. Then you're going to do whatever you can to remove that constraint. And so you have a lot of actors today, not just hyperscalers, but also compute vendors who are and neo clouds who sell to the hyperscalers, who are all actively looking for available power.
And if you think about Mara's kind of shift more recently towards AI, as we've said publicly, you know, we've always felt inference was going to be the attractive market. We've also said that, you know, most hyperscalers won't want to work with a Bitcoin miner because Bitcoin miners haven't developed tier four data centers. They don't know how to build that level of redundancy, resiliency, cooling, etc. And so we felt that it wasn't right for Mara to go in the market and try and do this ourselves. And so this is why it was so important to find a partner like Starwood because with Starwood, we have a partner who has built and delivered and operated over 70 gigawatts of hyperscale capacity for the tier one operators in the space. And it's somebody who has a captive construction EPC arm who knows how to build these and deliver them on time, within cost, and to the level of service needed by the hyperscaler.
So, is this sort of the biggest challenge for Bitcoin miners getting into AI specifically? Is just the infrastructural layer that needs to change?
So, think of it this way. They do not have a reputation for having ever done this before. So, if I, Mara, go to Google and say, I have a site I want you to come in. I'm going to negotiate a lease with you. Well, I've never negotiated a lease with a hyperscaler. They have never worked with me before. So, they're going to put in every single little protection they can to make sure that they get what they want and that I have to deliver everything.
Working with Starwood, who has delivered already to people like Google, Microsoft, etc., you already have that relationship. There's a certain amount of trust. The whole process goes a lot faster, a lot less friction. When you then look at construction, again, same thing. You have an experienced partner who's delivered before. And so, there is a higher degree of trust from the end tenant that the site's going to be online and available and operated according to spec when it was planned to.
And lastly, with Starwood, the big advantage for us is that they have the ability to credit wrap the project financing. Which makes it more attractive for us in that regard. Versus our peers, you know, you go look, you're reading announcements about delays to sites coming online. Well, that's because a lot of our peers have never actually built these sites before. And so now is when they're discovering that that lack of prior experience is starting to impact their ability to deliver on time. And there are huge penalties in these lease agreements for not delivering on time. In some cases, they're so big that they can remove a lot of the profitability of these deals. So that now you even go and raised all this money, gone and done everything you can to deliver this, and you deliver it 6 months late, and all of a sudden the profitability of that site is severely diminished.
It almost sounds like Starwood is kind of like a business development partner. I mean, is that sort of...
Well, business development is really more, you know, they manage tenant relationships, design and construction, and then operating and stabilizing the asset.
We are really the land and power bank, if you would, in the relationship.
And you know, because we have this portfolio of land and power available today, we were a very attractive partner for them. Because they now all of a sudden had inventory they could go and take to their clients and say, 'Hey, here's available land and power. You know, let's look at getting you in.' And so what we've seen is since we signed the agreement, and you know, we've been working with Starwood for over a year on this, and we had been out talking to tenants already late last year. So we're now seeing a very accelerated pace of conversations with tenants where you're starting to see a lot of activity around the sites, and it's more a question of allocating the right capacity to the right tenant, so that you're not overly dependent on any one tenant, which is risky because, you know, the economics of their business has changed. You know, look at OpenAI, they've gone from we're going to own and operate sites to no we're going to lease sites to whatever their model is today. And so you have to be very careful as to how you do that. But I think we chose a great partner and I think this is going to be a great way of, you know, Mara clearly generating a lot of value out of these sites that we have. And best of all, we get to keep mining Bitcoin the whole time up until these sites turn on and then we pick up our miners and move them to another site.
That was going to be another question that I asked was sort of where does Bitcoin mining kind of fit into this, right? Because, you know, it doesn't sound like you guys are abandoning Bitcoin mining. It's just the role of Bitcoin mining could change over time. How do you see these things working together? You know, what's the sort of strat... How does Bitcoin mining fit into the strategy?
Yeah, so we look at power and land assets through the lens of what we call an economic triad, right? Is the power low enough cost and of a type that you can mine Bitcoin? Is the land and power and access to potentially water and internet sufficient to use it for a hyperscaler or an inference or a NeoCloud site, colocation site? And then is this something that's really attractive for critical IT load potentially? So, critical IT load versus AI HPC is more private cloud, kind of the stuff we do with Exaion, where it is smaller scale sites, but where the primary service you're doing is providing highly resilient, secure private cloud where an enterprise may want to run its own models. They're running their critical applications as well as their AI compute.
And so that's sort of again where Exaion kind of fits into this whole piece of things is that they're really that's like an even higher level service that we're offering essentially. Is that enterprise grade cloud services component of this.
Exactly, and it's not multi-tenant. It's single-tenant. And so, if you think about it this way, you know, Exaion was born inside EDF, which is the largest electrical energy producer in the world. And French company that operates a lot of nuclear capacity, 56 nuclear reactors today with more coming online every year. Their whole reason for existence was to operate the AI infrastructure for these nuclear sites in a way that was highly secure and in no way dependent on any outside third party who could shut off access to some technology, some component for those sites because at the end of the day EDF is a French state-owned company. And so, it's a strategic player.
If you look in Europe today, there is a high degree of concern regarding reliance on US, Chinese, or Russian, or any non-European technologies and technology companies because of the fear that they could shut something off. And now all of a sudden you can't operate. And the perfect example of this is what the US government just did with Anthropic and Fable. If you had been running as a company anywhere in the world, Fable, and the US just directed Anthropic to shut it down, you're now SOL.
So, this is sort of like sovereign AI kind of or like...
It's what we call technically sovereign.
Yeah, so it's technologically sovereign.
Right? If you use the term straight up and down sovereign, the ownership of the company would have to be clearly sovereign, which is very hard because obviously, you know, most of the technology components that come into these systems comes from somewhere outside of Europe. But at the end of the day, what you want to ensure is that no non-European actor can pull a plug on you.
Do you think, I mean this is sort of I think a new idea for folks that, you know, maybe I might want to be effectively like running my own kind of siphoned models that are private essentially to me.
No, it's not a new idea in that simply look at the fact that, you know, look at your bill from Anthropic or OpenAI or Gemini or Amazon and look at what you're spending per token and then say do I really need to use them to do review and edit drafts of emails? Do I really need to use those models to do analysis of a financial spreadsheet? You know, you need to apportion the type of tasks to the model that is best suited to it and has the most appropriate token cost. You can today run your own models and there are lots of different models and I think even Exaion today in their infrastructure operate over 12 different model systems that you can easily switch between. You can easily shift all of that basic bread and butter AI use to your own models and run those in-house at dramatically lower cost.
And people are already doing this or do you see this as like a growth area that's about to take off?
No, they're absolutely starting to do this. I mean we have spoken to enterprise customers who are moving their AI processing if you would and their AI infrastructure out of Amazon AWS and into proprietary private cloud infrastructure because it's a 10x savings.
I mean realize your private cloud infrastructure provider has to have full secure infrastructure, etc. And so you need to ensure that again, another reason for Exaion. That's what they do. If they can do it for the nuclear systems in France, you know, you have to assume that they have a very high level of security in what they're doing. And you know, can you ensure that the data can't be taken by anybody? Can you ensure that the data is always safe, meaning it's encrypted whether it's at rest in flight or in use? Can you ensure that people can't access the sites? Can you ensure that no foreign actor can do something that shuts the site down? You know, and the big issue in Europe is also the Cloud Act, which is a US law that essentially says that, 'Hey, if you're running on Google infrastructure in Europe, guess what? Your data is accessible under the US Cloud Act.' So, all of these things kind of are driving people to really look at running their own infrastructure.
And it's not necessarily efficient for you to build your own data centers, but you can operate with somebody like Exaion who can operate private cloud on your behalf and provide all the services that you need. So, you get all the benefits without necessarily having to build that infrastructure yourself.
Going back to the sort of the relationship or this combo strategy between Bitcoin mining and AI, is that now that optionality kind of the thing that's sort of driving your decision-making when it comes to acquisitions, like energy acquisitions? Is that what you're looking for primarily or like how do you decide, you know, what's an attractive energy property to purchase? And I'm thinking, of course, about Longridge and...
Yeah, so Longridge is unique in one way in that A, we already had a 200-megawatt substation or site with 200 megawatts allocated to it called Hannibal, which was adjacent to Longridge. Longridge had 1,600 acres of land and a 500 more or less megawatt dual cycle gas-fired plant. And we were very interested in the land. Because the land we had at Hannibal wasn't sufficient enough to build a hyperscaler site. And we also wanted to have access to more power. And while Longridge today sells the vast majority of its capacity into the grid and we're not going to change that, there is the ability for us to add more capacity to the site from a generation perspective as well as applying for more load for the Hannibal site. That can get us to 600 megawatts available to build a significant hyperscaler site on the land that we would control as soon as we close the Longridge deal.
Is there like a minimum megawatt size property that you would need in order to service a hyperscaler? Like...
So it depends on the application, right? Hyperscalers today are typically looking at things north of 250 megawatts. But there are lots of neo clouds who want things north of 50 megawatts. And the neo clouds service the hyperscalers predominantly. Core Weaves of the world, Nscales, etc.
What's the distinction between a neo cloud? I've not heard that term.
So a neo cloud is essentially, Core Weave's the perfect example. So they essentially are running GPUs as a service on behalf of a hyperscaler. So, you know, if you're Google, Microsoft, whoever, and you want to grow capacity, you can do it using your own capacity or you can do it asset light, which is you contract with a neo cloud. I'm going to take 100% of your capacity for the next 5 years. You're going to buy the GPUs, you're going to build the site, you're going to operate it, and I'm going to pay you a fee per megawatt.
And that's a fixed amount.
That's a fixed amount. So, if you think about how we grew as a Bitcoin miner, we did the same thing. We grew initially by being asset light. We didn't own and operate sites, we just contracted for capacity so that we could grow as fast as possible. We then vertically integrated to own the sites, which we were able to do at lower than replacement cost, and then we started buying power assets to fully vertically integrate. It's exactly the same thing the hyperscalers do, right? They own and operate a certain amount of capacity themselves. They contract with neo clouds so that the neo clouds have to take the risk of buying the GPUs and doing all of that. This isn't on the hyperscalers' balance sheet, it's on the neo clouds' balance sheet. And so, that gives them extra flex capacity. And then, over time, they will build more of their own sites. And the neo clouds just become middlemen.
Essentially. And so, it's going to be very interesting to see how this business evolves over time, but I think there's going to be more than enough demand for inference capacity that the neo clouds will eventually shift their business from servicing hyperscalers to servicing enterprise customers because it's a much higher margin business.
Enterprise is a higher margin.
Yes, than servicing a hyperscaler. I mean, negotiating with Google, granted, it's great to get a 10-15 year contract with Google for something, but if you can service enterprise customers, there are more services you can bake into what you're doing, which makes it a higher margin business.
Okay, so the higher margins comes...
from bells and whistles, basically, that you're able to offer enterprise high end.
Cuz realize, the hype the enterprise customer is otherwise buying all of those services from Google.
And are the bells and whistles things like security by XaaS? Like, is that an example of like a bell and whistle?
Security is definitely a bell and whistle. Other bells and whistles are the ability to do load shaping, load balancing, the ability to run multiple different models, the ability to automatically route traffic based on the cost of a token, based on the latency that your use case provides, so quality of service.
Quality of service broadly.
Yeah, so I think what you're really going to get to is where essentially it's going to be a little bit like the internet where you are essentially going to just take it for granted and you're going to say, okay, I'm going to run my model and I need to run it within kind of sandboxed, safe architecture and I'm not going to worry about what model's being run or where physically it's running or anything like that. That is all going to be hidden behind a control plane that I operate with and I'm going to work with my trusted vendor who's going to ensure that the data's secure, that the models are all updated continually, that the infrastructure operates, it's fully resilient, etc.
Is there, I mean, I feel like we've kind of talked a lot about kind of the similarities between Bitcoin mining and kind of why Bitcoin miners are suited for this, you know, this shift, this AI HPC kind of game that everyone's obsessed with right now. I'm curious, what are the differences? Like what are the biggest differences between Bitcoin mining and HPC as a business from a business perspective?
Well, the capital required is much more different. You know, in Bitcoin mining you're at about a million dollars a megawatt all in between infrastructure and compute. In the hyperscaler game or in the, you know, large-scale AI, your infrastructure, just the building, power to the building, cooling to the building, is anywhere from 10 to 15 million a megawatt.
So, right there you're at 10 to 15x. And that's even before you consider the compute, which is most probably 40 to 50 million a megawatt. Right? So, now you're talking 50 to 60 million dollars a megawatt, which is 50 to 60 times higher than what it costs to do Bitcoin.
And but what do you get for that on the other side?
Well, you obviously are charging a fee that is giving you a good return on that capital, but you have to have access to the capital to be able to do it.
Now, granted, you know, Mara is not going to be in the business, at least initially, of renting GPU capacity. We're building buildings for data center operators like Google, Microsoft, and others to come in and deploy their own compute in those sites. Almost more like a real estate development play where we control power and assets, and we allocate those accordingly. Over time, we may very well, especially at the lower end of the marketplace, build and operate actual token factories, if you would, as capital is available, as the market continues to develop.
Which basically means acquiring GPUs, essentially.
Essentially. And if you look at what CoreWeave does today, I mean, they are a buyer, they're a partner of Nvidia. They buy GPUs, they deploy GPUs for their enterprise customers. Again, CoreWeave is the high-value, high-margin end of the spectrum, low scale. All right, a typical CoreWeave data center may be 5 to maybe max 25 megawatts. All right, your Neo Cloud site is typically 50 to 200 megawatts, and a hyperscaler site is 250 and up.
And they're basically serving, they're smaller because they're just serving very specific clients who need their specific tech stack essentially.
Absolutely. You know, there are very few companies out there today that require more than one NVIDIA cluster to run their models. Inference is much more efficient from a compute perspective than training is. Training is very GPU intensive. Inference doesn't really require GPUs. You can run it on a mix of ASICs and CPUs. And so, this is why NVIDIA went out and did this deal with Groq. Because it enabled them to release an inference system that costs about 1/30th per token of what it would cost if you just ran it on the latest GPUs.
Are there like benefits also in terms of, you know, okay, so I'm going to pay 50, 60X CAPEX up front, but I'm also going to have, you know, the sort of fixed rate that I'm getting paid over a 10, 15, 20-year period. That's obviously not the case with mining. I mean, like what are sort of like other broader business benefits of this strategy over Bitcoin mining in your opinion?
Well, similar to Bitcoin mining, there's technology obsolescence risk.
So, now think you have a tenant who is leasing GPU capacity from you. GPUs don't last 15 years. Just like Bitcoin miners don't last 15 years.
Every 3 years you're swapping out and upgrading your Bitcoin miners. Every 3 years, think about it. What was NVIDIA selling 3 years ago? What are they selling today? What are they talking about selling in 3 years? Hugely different technology in the sense that it's much more efficient, much more powerful. If you're a hyperscaler and you're renting capacity from a neo cloud, for example, 3 or 4 years down the road, are you going to want to be paying that price for infrastructure that's three or four years old?
I think what you're going to see is, you know, the hyperscalers and the large companies like putting the balance sheet risk of all those GPUs on the neo cloud vendor.
So, I was going to say, is it the neo cloud vendors who are then going to be effectively holding the bag of this obsolescence situation?
If you were to sit down with a financial analyst and look at all these credit deals that have been done by the neo clouds and the credit backstops and who's actually ensuring that these guys can get the credit they need to go buy all these GPUs. The question is what happens when those GPUs reach the end of their economic life? Where in the calculus of all this, where in the pro forma projections of all these neo clouds is the technology obsolescence risk taken into account of?
