Good morning, everyone, welcome to our 54th Annual TD Cowen CMT Conference. My name is Michael Elias, and I'm the Communications Infrastructure Analyst here. For this session, we're joined by MARA Holdings, and from MARA, we have their Chairman and Chief Executive Officer, Fred Thiel. This session is structured as a fireside chat. We have about 30 minutes for it. I will try and open up for questions, but I always get carried away and I ask too many questions. With that, Fred, thank you so much for being here. Really appreciate it. Great to be here. Awesome. All right. Let's kick things off by walking through MARA's pivot to AI and HPC infrastructure. I'm curious, how did the strategic shift come together in the last few months, and what brought us to this moment? If you think about the history of kind of Bitcoin miners or great aggregators of low-cost power and land, we operate in a market where we're totally tied to a commodity, we are very good at being efficient low-cost operators. Over the years, MARA started originally as an asset-light company and by meaning over the years, go back to 2021, which in Bitcoin terms is a long time. Essentially an asset-light company where we brought the compute, but we didn't own the data center. We didn't own the power. We started deploying in 2019, really accelerated in 2021, and by the end of 2023, we were one of the largest in the world. Why that strategy? Well, at the time it made sense because of $100 million of CapEx, you used to have to spend $30 million on infrastructure and then $70 million on compute. We said, we'll just go with hosted. We can spend 100% on compute and grow faster. There are three constraints in the Bitcoin mining world which actually exist also just the same in the AI world, and that is capital, compute, and capacity. Capacity being data center capacity. In 2024, if you think about 2023, 2022, there had been a big drop in the price of Bitcoin. Over the course of 2023, things started recovering, a lot of data center operators in the Bitcoin world were hurting because they had had issues with loans, debt, and so we were able to buy 70% of all the capacity where we operated at less than replacement cost. Essentially bought out the owners of the data centers where we operated. That gave us then ownership of land and power. Then, in 2025, we started adding actual power assets to that. We bought a wind farm. We started mining off of flare gas and oil fields. We are also one of the very first Bitcoin miners of the U.S. publicly traded ones to expand internationally. We did a 250-MW deal with the sovereign in UAE where we developed two data centers in Abu Dhabi that today still operate and are some of the most efficient in the world because it's all liquid-cooled immersion, no air conditioning. Oh, wow. All the types of technologies you need to use for AI in the future. We started looking at AI back as early as 2023, really more 2024. having been in the tech industry for 40 years at this point, and having dealt with large hyperscalers, when you're going to engage with a Google, a Microsoft, an AWS, and try and convince them to use you as a partner where they're going to host their systems, Bitcoin miners are Tier 1. Yeah From a complexity perspective and redundancy. Hyperscalers are Tier 4. They have their recipe. Microsoft literally has a, here is my design. You have to build it like this. Yep. We felt we wouldn't be a credible counterparty to a hyperscaler because we would have to go hire a team specifically to do tenant leasing. We'd have to go hire a team that was expert in design and build. We have to find an EPC partner, and then you have to raise a huge amount of money to go do this. We were more believers in the world of inference AI, which is think more where Anthropic operates, think more where what the neoclouds today are really servicing, which is not training, but it's inference AI. Because inference is where the operator really makes money because you're getting paid for token consumption. Token factories, and especially token factories at the edge, was an area that we were interested in. We started deploying some test installations in one of our Texas sites where we were able to use modular inference infrastructure right next to our Bitcoin mining. Container of inference right next to a container of traditional Bitcoin mining. This market started getting hotter and hotter and hotter. We realized that we were sitting on a huge portfolio of power assets, over 1.1 GW of power that was already running and energized. We hired a third-party advisor to come and evaluate our sites and tell us what of this a hyperscaler would actually be interested in, and what do we have to do to these sites to make them actually attractive. Because there's a certain amount of upfitting, if you would that you have to do, substations, things like that. What came back surprised us. It was, well, the vast majority of your sites actually are attractive. There are 10 counterparties to 15 counterparties we think you should be engaging in on this. Then the question came down to, okay, are we going to go hire a whole team to do this or are we going to find a partner? So we had been talking to Starwood about us providing services to them regarding load balancing for their data centers. As we started talking with their team, it became very clear that there was a lot of synergy here. So over the course of a year, we put together a framework for an agreement, and we actually started testing this late last year, going out, talking to prospective tenants. We went to the PTC conference together with Starwood. They ran us out in front of people and did the dog and pony show, and we got a huge amount of interest. At that point, we finalized the agreement in February, and that was when we announced it to the world. This has been a 2+ year process, so not quite a pivot that we decided in our sleep one morning. Yeah. This clearly building up to this moment. Yeah. As I think about what you just outlined, 1.1 GW of power, you diligence it, what sites would folks be interested in? Can you just help frame up for my own understanding, how many sites is this across? What's the average megawattage per site that you have? Yeah. We have a mix of site capacities. Some sites are multi-hundred megawatt sites, and then we have sites that are the sub-100-MW sites. From a hyperscaler perspective, the interest is in the 100 MW and up sites, which makes up about a third of the sites in number, but the vast majority of the capacity. The interest in what we call subscale sites is predominantly neocloud- type interest. What we're seeing now is there's a shift in this market. If you think about the market today, it is totally constrained by power. NVIDIA can't sell more GPUs if there isn't more power for their customers to plug into. Now you have AMD, Microsoft's got their own silicon, Meta's got their own silicon, et cetera. There is a market share war going on at the silicon level to own access to power. Yep. They want to be the gatekeeper so they can say, I want to control this amount of power for my customers. What's happening now is on the inference side, inference, you can have multiple inference sites, and you can distribute your inference workloads across multiple sites. Now the subscale sites become really interesting for the inference providers, the Anthropic of the world, people like that, where it allows them to grow capacity, albeit in smaller chunks. It allows them to grow capacity where otherwise it's hard for them to grow. Again, realize everybody needs to grow. Yep. You start trying to get creative about how you can grow, and there's a whole food chain here. It's not just the neocloud or the hyperscaler that wants to grow, it's NVIDIA, AMD want to grow. They're in the hunt for this. You saw, I think earlier this year, there was a deal where NVIDIA put down a $120 million deposit on a site so they could lock up the capacity there for one of their customers. Where you have seen the silicon vendors be the credit backstop to some of the neoclouds, what you're now seeing is the silicon vendors are instead saying, I'm just going to take this capacity, and I'll be the owner/operator. Yep. I think you've had this huge acceleration in revenues of the silicon guys, and what's going to happen is at one point, that acceleration will catch up with limited capacity. Yeah. Say that. This capacity isn't going to fix itself overnight. At the same time, when you take a site, like our Granbury site, for example, which is we operate 200 MW. We have 300 MW available there. It sits right next to a power plant that could up to a gigawatt of power to the site. Even though our portfolio is 1.1 GW now, it'll grow significantly with the Long Ridge deal we just announced when we close that. The capacity that's available at the sites to grow gets us over 2 GW already. Okay. Got it. I want to take the comment that you made about, all right, we have sub-100 MW sites, and then we have above 100 MW sites. There's a decision tree of outcomes where we could go. On the 100 MW plus scale, I presume you're doing with Room or with Starwood, you're doing the direct leases with the hyperscalers. When you're talking about the sub-100 MW level, we're talking neocloud, is that regular way leases to the neoclouds, or are you yourself, given your background in compute, looking to serve as, let's call it a GPU-as-a-service provider or maybe a TPU-as-a-service provider? Do you want to be the owner of the infrastructure, or do you also want to be a provider of compute? It's interesting. If you do an analysis of the EBITDA multiple on GPU-as-a-service versus doing real estate development. Yeah Real estate development is a much better multiple, and it's a lot less CapEx intensive. Give you an example. Take a 200 MW site, or let's just say 100 MW site. 100 MW site, it costs you roughly between $10 million-$15 million a megawatt to build that site from an infrastructure perspective. You're going to put $30 million of compute in it, right? Yep. Per megawatt. Right. From a capital perspective, with Starwood, for example, I can take 100 MW site. I contribute that to the partnership. I, just for indicative purposes, say $1 million a megawatt. I just contributed $100 million to the partnership. For them, they have to catch up to that before we have to put dollar one in. Yeah Beyond that, right? If we're going to go the neocloud with GPU-as-a-service route, I've got to take my site, I've got to build $100 million of infrastructure, and then I've got to add $300 million of compute to it. I'm now $400 million of cash out of pocket, and I'm not going to get necessarily a better return. The way we look at it is, there are phases to all things. You can't be all things to everybody, right? Oh, yeah. We think that right now we have a portfolio that will do very well in a leased environment. As it converts, we continue to acquire more land and power assets. We think that in the longer term, domestically in the U.S., different market than internationally, but domestically in the U.S., what is going to become attractive are some of these smaller sites. There will be opportunities to partner with the silicon vendors to basically say, now you're going to be my partner on these sites. You're going to bring the silicon, we'll operate it for you. We'll run it for you. But you'll be the partner. I think that's an attractive model as phase 2, then phase 3 is large data center campuses. Think of it as like an industrial campus. You'll have a couple of anchor tenants that are hyperscalers, you'll have some neoclouds. You'll have some owned and operated stuff that you run yourself. We've covered the journey to here. How we're thinking about the leasing up the portfolio. We talked about Starwood. Now, as we think over the next, let's call it six to 12 months. Yep. What are the strategic priorities for you? Where do we go from here? Signing tenant leases on the existing portfolio, closing Long Ridge, and getting that leased. I think if we do all of that this year, it will be a very great year for our shareholders. Yeah. Great. Let's talk about the demand environment. You're making this pivot, we've seen a big acceleration in the broader industry. One, I'm curious, given the partnership, are you engaging with the clients directly? If so, what are those conversations looking like? What's the sense you get for the urgency of demand? Where does that leave you in terms of, let's call it near term execution or even medium term execution on potential data centers? Yeah. lots of conversations going on. Obviously, if you look at the model providers, OpenAI, Anthropic, Google, Microsoft, et cetera, they are all struggling for market share. Market share is totally driven by how much compute you have. The amount of compute says how many tokens you can make available to your customers. To give you an example, in our operation in France, they complain because, when they run Claude, until 5:00 P.M. East Coast U.S. time, the tokens are hugely expensive. At 5:00 P.M., the price drops. What does that tell you? 80% of all AI compute is in the U.S. to service the global demand. That's very interesting. Right? We can get to international in a second. Right. They are constrained. What's Anthropic having to do now? They're raising prices, right? It's a yield game. It's like you're running an airline. If you only have so many seats and you're selling out your seats, just keep raising your prices until you get to a point. That's right. Right. You have to be wary of market share. OpenAI, Anthropic, everybody is now focused on this capacity game. You have the silicon guys worried about capacity, like I talked about earlier. Yeah. You have the model guys worried about capacity. The conversations you're having are with silicon providers, with model providers, with the hyperscalers, with the neoclouds, and with enterprises. Oh. There are large financial, I have to word this carefully to not step on an NDA. Large financial institutions, traders, if you would, who run multi-hundred megawatt data centers f or model provisioning and operating their quant systems. They are disintermediating the neoclouds. They're coming direct, and they're saying. We want X hundreds of megawatt in a data center. Yeah. You have a lot of other enterprises that are now starting to get into the game, who are wanting 30 MW, 40 MW, 50 MW of capacity. We're having conversations up and down the tree of potential customers. For us, it's about financeability of the tenant. Okay. Right? We'll come to that in a moment. Right. Who's creditworthy and who's not. The advantage we have with the Starwood partnership is they credit wrap these deals for us. We have a benefit there. You want somebody who's very creditworthy because that lower cost of capital is really important. Granted, they negotiate on the lease price, too. Yep. Right now, demand grew 80% year-over-year, I think you're going to see in the industry. I think you're going to see continued growth. There was a big step up this year in February when the agentic world kind of turned on. That is now only going to increase even more. At the same time, the way some enterprises are responding to this cost issue, if you would, is that they are now starting to look at private cloud as an alternative. Essentially operating models themselves in their own infrastructure or leased infrastructure, which is why Exaion, which is the investment we made in France, becomes very interesting because that's exactly their business. It's interesting and better what you said now, we're talking about hyperscale. I would put AI lab in a separate category from hyperscale. We talk neocloud, then we talk enterprise. When I think about enterprise, that's a bit of a different motion than going after. I appreciate the deals are getting larger. Is enterprise a vertical that you see yourself pursuing, and is it exclusively on the large footprint side when we're talking about a financial institution needing 100 MW, let's say? Is that the type of opportunity? Are we looking at 10-MW opportunities with enterprise? Yeah. I think enterprise is sub 50 MW, definitely. 5 MW to 25 MW is how I would estimate it. For example, we have a prospective tenant or customer in France for our Exaion company. They have a very specific latency requirement that you have to meet. You have a very specific token cost target you have to hit, and you have an amount of storage, amount of network bandwidth. It's very much like traditional data center business, which is something I know very well from many years ago, critical IT loads. You'll see in our materials, we talk about critical IT loads. Critical IT load is when you all of a sudden have agentic systems running in parallel with your operating technology, the applications and things that run your factories. If you think about it, what is the real value equation for AI in the enterprise? It's take my data and make it more valuable and actionable, right? This is the threat to Salesforce, who just opened up an API so that you can now run my AI on my own data, but you own my data. The threat to SAP, the threat to all these companies of AI is it disintermediates the user interface. Enterprises are going to start bringing this home. I'll give you an example. If you go to AWS and you say, I would need dedicated infrastructure for my AI, it costs you 10x more than running it yourself. That's interesting. You don't need B200s for the vast majority of AI. If you start looking at the workload of AI, what is the AI actually doing? All the stuff that is review drafts, prepare an email, search for this, do an analysis of this confidential investment memorandum. Look at public filings, all that stuff can be run in an open source model that is free, that you run on your own infrastructure using RTX technology from NVIDIA. We're talking N- four generations. Yeah. You don't have to run it on H100. That costs you pennies on the dollar. What these larger enterprises that we're talking to are waking up to, it's like, we want to bring this stuff into a private cloud. We want you to operate the private cloud. Think private cloud is essentially a dedicated cloud infrastructure that is near- prem, that has application server storage, networking, GPUs, if you would, where you're running fully secure, technically sovereign, right? The data never leaves. Yep The data's always under your control, et cetera. Security's the key thing. Exaion, the company that we invested in in France, they were part of EDF. EDF is still a 33% owner of the company. They are the operator of all their nuclear data centers, which are all private cloud. They now are broadening their offering to French banks and large financial institutions and others. We're now broadening that across Western Europe. In Europe and Canada, there is a desire to not buy American. Canada, for example, just chose a European fighter over U.S. fighters for their defense. This whole issue of sovereignty, data sovereignty, and the fear of the U.S. CLOUD Act, which for those of you who aren't familiar, if you are running on Google infrastructure in Europe, that is not sovereign, because of the CLOUD Act allows U.S. law enforcement agencies and regulators to reach in and access your data. Right? It's a U.S. law that has put the fear of God into the Europeans. There's a huge moat, essentially, that blocks off Microsoft, Google, and other U.S. providers from Europe because of this. Yeah For defense, healthcare, finance, et cetera. That's why we view this Exaion opportunity as so unique. It's sort of a Trojan horse. That's very interesting. I don't want to put words in your mouth, what I'm taking from this is it sounds like you're incredibly bullish on the enterprise long term. You're talking about them essentially disintermediating the neoclouds. If right now an avenue of growth for the broader data center industry is the neoclouds, it sounds like that's a place that you want to be positioned long term on the enterprise side. Yeah. Again, been in the technology industry 40 years, cycles tend to repeat themselves. Yeah. My first things that I worked in were large bank enterprise software systems. Over the years, banks still run the same old COBOL systems because it hasn't been worth the while for them to replace them, right? Yeah. What's happening today is AI is a bigger and bigger cost factor. They need to lower the cost of AI. What AWS gives you is flexibility. It gives you elasticity, but at a cost. What always happens in the tech industry is that who is the end person who's paying all the bills of everybody in the food chain? It's the enterprise, right? Eventually, the enterprise is going to want to go somewhat direct. Right? There was a whole outsourcing phase, which is what created the cloud industry. Yeah. Look at what was really outsourced, email, calendar, documents. Not the most- 70% of corporate data still sits behind the firewall. It is not in the cloud. You go talk to any manufacturing company, you go talk to any healthcare company, all that core, most critical data is not sitting in a Google cloud somewhere, or a Microsoft cloud. That's why I believe that as they want to adapt and adopt AI, they're going to want to do it in a private cloud framework. Long term, I'm very bullish on the enterprise. Meanwhile, you have to pick and choose your battles, and you have to look at the assets that you have and how do you maximize the shareholder value of those assets. Yeah. One philosophical question. We talked about financeability. You've done a great job of assembling this portfolio of power across different sites, and you want to make sure you get the highest return on it. How do you think about recognizing that there's a ton of value coming from or being created by the Anthropics of the world who are not investment grade and are not, or not yet investment grade? Right? There's probably more issues financing them at the moment. How do you think about working directly with those kind of folks? Is there, to the extent that you were going to, is there some sort of backstop structure, or would it have to come indirect through a hyperscaler? In the case of, for example, OpenAI, the problem is that it depends on what drugs they're on that day, because one day they want to own data centers, one day they don't want to own data centers. Right? It's like Oracle, right? The risk is, on the one hand, credit. The bigger risk is, are they going to change their minds? Yeah. I mean, think about this, right? When you build a data center for somebody, you're building it to their needs. I can't take a Google data center that I've half built and give it to Microsoft. Yeah. They'll say, tear it down. Right? When you get married in this business, Talk about a prenup. You are stuck. We look at it as a company like Anthropic, if we're going to work with them, then yeah, their silicon provider has to backstop them more than likely. Yeah. It needs to be iron- Ironclad Clad. Okay. That's very helpful. I do want to pivot talk. We've talked about the Starwood partnership. From what you said earlier, it sounds like you contribute an asset, and then there's a decision of what's the appropriate stake that you're going to retain and the stake that they'll have. How are you thinking about the calculus of what the appropriate stake to retain in a site is? It just reminded me, is that your discretion, their discretion? It's our discretion. Okay. Basically, we're allowed to invest up to 50%. The way it works is we first contribute the site. They have to catch up to that value. There's a predetermined value per site in the agreement. Any asset we contribute later will have a predetermined value to it. If you use $1 million as an indicative value, then a 200 MW site, that's $200 million of value we're contributing. We've most probably done some pre-leasing enhancements to the site, upgrading power station, substation, stuff like that. Say another $80 million we've put into it. We have $280 million of equity credited to us in the JV. Starwood has to catch up to that $280 before we have to write another check. Yep. 200 MW data center at, say, $12 million a megawatt, which is the midpoint between $10 million and $15 million, more or less, right? That's $240 million. Sorry, $200 would be $480 million. Half of that would be $240, but we're already at $280. Yeah. Now Starwood has to go out of pocket until we're equal on that. If you get 80% project finance, it works out pretty nicely. At what point does it get contributed? Is it once it's leased? Yes. Okay. That answers it. Yeah, it doesn't actually get contributed until it's leased. A couple of other bennies. They have to deal with the cost overruns. Okay. That's not on our balance sheet. Okay. They credit wrap it, obviously. We get to keep mining at the site until the very last minute when the tenant says leave, and then we get to pick up our Bitcoin mining infrastructure and move it because it's all containerized. Over the long term, there are some data center operators that it's like build to sell, right? Now you lease it, you contribute it, and then from there you'll stabilize it. Is the plan to then monetize that asset, recycle that into the next set of projects and acquire more land and power and keep the flywheel going? Or do you want to continue to keep your share of those data centers over the long term? Yeah, I think A, it's too early to tell, but it's the financial calculus is what makes the most sense for the shareholders. If the demand for data center capacity is such that a prospective investor or owner of the data center would view those cash flows as very attractive, then it's like commercial real estate at the end of the day. Yeah. Right? It is a commercial real estate game. It may make sense to stabilize it over two years and then, okay, we'll flip this and we'll go do it again. Okay. You continue to operate it and collect cash flows until you retire. Both options exist. What I will say is the attractiveness for us of the model is really designed ideally for us to just keep aggregating land and power and throwing it into these deals. Okay. If we exit, that's great. If not, we're just getting these really nice cash flows from it. The other thing is we're not consolidating debt on our balance sheet. It's a very attractive structure. Over time, most probably we'll do smaller deals ourselves and we'll see how it all develops. We're really bullish on it, and I think we have a really good path and vision for what we're going to do over the next 5 to 10 years. Perfect. With that, we're out of time. Great. Thank you so much. Thank you. I really appreciate it.