Q3 2025 Earnings Call Transcript

Greetings. Welcome to MARA's Q3 2025 Earnings Conference Call. [Operator Instructions] Please note, this conference is being recorded. I will now turn the conference over to Robert Samuels, VP of Investor Relations. Thank you. You may begin.

Thank you, operator. Good morning, and welcome to MARA's Third Quarter 2025 Earnings Call. Thank you for joining us today. With me on today's call are our Chairman and Chief Executive Officer, Fred Thiel; and our Chief Financial Officer, Salman Khan.

Today's call includes forward-looking statements, including those about our growth plans, liquidity and financial performance. These involve risks and uncertainties, and actual results may differ materially. We disclaim any obligation to update these statements, except as required by law. For more details, see the Risk Factors section of our latest 10-K and other SEC filings.

We'll also reference non-GAAP financial measures like adjusted EBITDA and return on capital employed, which we believe are important indicators of MARA's operating performance because they exclude certain items that we do not believe directly reflect our core operations. Please see our earnings release for reconciliations to the most comparable GAAP measures. We hope you've had the chance to read our shareholder letter and look forward to your feedback. We'll begin with some brief prepared remarks from Fred and Salman.

After their comments, we are going to be conducting an analyst interview with management. Today's session will be conducted by Reggie Smith, analyst at JPMorgan. And with that out of the way, I'm going to turn the call over to Fred to kick things off. Fred?

Thanks, Rob, and thank you all for joining us. This quarter, we continued to evolve MARA from a pure-play Bitcoin miner into a vertically integrated digital infrastructure company, one that converts energy into both value and intelligence. At the heart of our strategy is a simple belief, electrons are the new oil. Energy is becoming the defining resource of the digital economy, powering everything from Bitcoin mining to artificial intelligence.

And we believe those who control abundant, low-cost energy will shape the future of both finance and intelligence. Bitcoin has now entered its institutional phase. We're seeing financial leaders such as BlackRock, Citicorp, and now even JPMorgan, integrating Bitcoin into traditional frameworks. And we're seeing the establishment of strategic Bitcoin reserves by corporations and governments alike and even the Secretary of Treasury has posted positive notes about Bitcoin on X.

What miners have always understood is now being recognized by global markets. Bitcoin is digital energy, a mechanism for storing and transmitting value. As one of the largest Bitcoin miners in the world, MARA sits at the center of this shift. Our energy to value infrastructure allows us to convert raw power directly into Bitcoin that we hold on our balance sheet, a distinct advantage that grounds our broader mission, transforming energy into intelligence.

Every electron has potential value and artificial intelligence represents the next frontier of this transformation of energy into even higher value. We believe that inference AI where the value of AI is actually created and derived, and not training in foundational models is where the AI industry will create the greatest amount of value over time. Every insight produced by an AI model has a cost per token, driven by the cost to build and operate the data center, of which the energy cost makes up a major component. Over time, compute and the cost to build the data center will drop as technology advances such as low-cost ASICs, open-source models and the ability to operate in less sophisticated and less costly data centers drive efficiencies resulting in rapidly declining drops in cost per token, making the AI data centers of today unable to compete on cost per token over time without significant technology refreshes, requiring even more and higher capital injections.

We believe energy, not compute, really becomes the primary constraint on AI growth. We are already seeing the alternatives to GPUs enter the market and open-source AI is making it far easier and much less expensive for companies to deploy advanced AI systems directly in their own private cloud environments. In the past, most models were only available through public cloud APIs. That meant enterprises had to send data off-site and pay high per token fees to access AI capabilities.

But today, many of the world's most capable models like Llama, Mistral, and others are available in open-source form, giving companies full control to run AI more cost efficiently and fine-tune their models privately. This is a major inflection point for enterprise computing and a shift that plays directly to our strengths as we build out low-cost, high-efficiency compute powered by our own energy infrastructure. We believe we're positioned to provide the kind of private, scalable environments enterprises need to deploy these open models securely. MARA is positioning itself at the nexus of these two AI trends.

Open-source AI is expanding the addressable market for private cloud compute. We believe that the future infrastructure will be built to serve that demand efficiently and profitably. This is where MARA's expertise in securing and operating low-cost power gives us a distinct advantage. Just as we optimize for the lowest cost per petahash in mining, we're now optimizing for the lowest cost per token in AI inference.

Our long-term vision is to integrate these two energy pathways, Bitcoin and AI into a single platform. Bitcoin mining monetizes underutilized energy and stabilizes grids, while AI inference transforms that same energy into intelligence and productivity. By bringing Bitcoin and AI together, we seek to maximize the value of every megawatt hour we manage. We've already begun executing on this strategy.

This quarter, we installed our first AI inference racks at our Granbury site within a modular non-water cooled containerized data center. This site currently has 300 megawatts of nameplate capacity with potential opportunities to expand our growing AI inference business in combination with our Bitcoin mining operations at the site. This milestone marks a significant step forward in proving out our AI infrastructure and next-generation inference hypothesis. It also demonstrates the versatility of our platform, underscoring the potential flexibility of our mining sites to support AI workloads along with Bitcoin mining.

Two major initiatives this quarter are propelling our strategy going forward. First, our pending acquisition of Exaion, a subsidiary of EDF in France. Once regulatory approvals are completed and closing conditions have been met, Exaion will expand our capabilities into enterprise-grade AI-optimized private cloud and HPC infrastructure. We believe this will position MARA as a credible partner for enterprises seeking secure localized inference capacity.

Second, today, we announced an initiative with MPLX, a separately traded public company formed by Marathon Petroleum Corporation, the largest petroleum refinery operator in the United States, to develop and operate multiple integrated power generation facilities and state-of-the-art data center campuses in West Texas. Under this initiative, MPLX will provide long-term access to lower-cost natural gas at scale, where MARA will develop and operate on-site power generation and compute infrastructure. The initial capacity is expected to reach 400 megawatts with the option to expand to up to 1.5 gigawatts across three plant sites.

We are also evaluating additional prospective sites to support modular AI and HPC data centers alongside mining operations, creating optionality for future AI inference workloads. MARA's approach is to deploy smaller, modular facilities directly at lower-cost power sites instead of building hyperscaler campuses. We believe this distributed model will enable us to capture value at the inference layer while continuing to monetize mining and grid sales. This modular structure also gives MARA the optionality to shift capacity towards HPC over time as and if economics and infrastructure maturity support greater AI utilization.

We believe MARA is positioned to capitalize on a key structural advantage as power becomes the primary constraint in AI growth. Together, Exaion and MPLX connect the two sides of our AI and data center business, energy and compute, and strengthen our ability to control both cost and performance from power to inference. Internationally, we're deepening relationships across Europe and the Middle East, where we see significant opportunity to deploy our integrated energy and compute model. Our pending Exaion acquisition exemplifies this, and we're honored to welcome Gérard Mestrallet, President Macron’'s special energy onboard as an advisor tomorrow.

His expertise strengthens our global strategy as we pursue our goal of driving 50% of revenue from international operations by 2028. On the financial front, we continue to operate with discipline and transparency. We ended the quarter with 52,850 Bitcoin, having mined over 2,100 BTC during Q3. We remain focused on improving free cash flow through ongoing cost optimization, site level efficiency gains and disciplined capital allocation. We have begun opportunistically monetizing Bitcoin from production to fund operating expenses and aim to limit reliance on our ATM to support growth initiatives, helping to mitigate shareholder dilution.

As I spoke about last quarter, Bitcoin prices have consolidated within a range since Q2. With intermittent volatility, we view this as a healthy period of equilibrium characterized by institutional inflows into ETF balanced by long-term holder liquidation activity. Using Jordi Visser's IPO analogy, Bitcoin is going through an IPO where early investors in VCs are exiting and institutional investors are coming in, forming a new base and foundation for growth. Meanwhile, broader macro trends, including rate cuts and expanding liquidity suggest improving condition for risk assets. Regardless of short-term volatility, our long-term trajectory remains unchanged, building enduring value through energy ownership, operational excellence, and strategic execution.

Finally, I want to provide an update on 2PIC. While we continue to recognize the long-term potential of 2-phase immersion, its practical broad application is still a few years out, and direct-to-chip cooling remains the preferred cooling methodology of data center operators and compute OEMs. We have exited near-term investment in 2-phase immersion to focus resources on opportunities with more immediate and higher return potential. In closing, MARA is evolving from a Bitcoin miner into a digital infrastructure leader, combining energy generation, Bitcoin mining, and AI compute under one scalable platform.

Our guiding metric is simple, profit per megawatt hour. It measures how effectively we convert energy into value, whether in Bitcoin, AI inference, or grid stability. As we continue to execute, we believe the market will increasingly recognize the strength of this diversified model and the strategic importance of energy ownership in the digital economy. I want to thank our employees for their exceptional work this quarter and our shareholders for their continued support as we build MARA into the world's leading digital energy and infrastructure company.

With that, I'll turn it over to Salman to review the financials.

Thank you, Fred. During the quarter, global hashrate grew by roughly 20%, with the hashrate and network difficulty both hitting new all-time highs by end of the quarter. Bitcoin's price remained relatively range-bound, trading between $104,000 to $124,000, closing the quarter with a modest $7,000 gain. It was one of the most competitive mining environments in recent times and a difficult backdrop for our performance this quarter.

Despite this, Q3 was the highest revenue and exahash quarter in the company's history. Our focus on operational and financial discipline over the past year is reflected in the substantial growth of our compute capacity and Bitcoin holdings. Between Q3 2024 and 2025, our Bitcoin holdings expanded by over 98%, growing from approximately 27,000 to nearly 53,000 Bitcoin. Our energized hashrate also expanded, increasing 64% from 36.9 to 60.4 exahash per second.

Bitcoin price appreciation resulted in approximately $4.3 billion or 256% increase year-over-year. While Fred spoke to our vision and strategy, our vertical integration and capital allocation strategy is reflected on our financial results. That balanced execution allowed us to expand our holdings and take advantage of favorable market conditions while maintaining liquidity and flexibility. We mined 2,144 Bitcoin and purchased an additional 2,257. The impact on our financials is evident in the results we achieved.

Let's dig in. Revenues increased 92% to $252.4 million from $131.6 million in the third quarter of 2024. Bitcoin's average price increased 88% over that time period, which contributed $113.3 million. We mined an average of 23.3 BTC a day throughout Q3 compared to 22.5 BTC in Q3 of 2024, which resulted in 74 more Bitcoin mined this quarter. Our strategy to deploy exahash responsibly resulted in growth of our BTC mine despite a significant growth in global hashrate and the network difficulty level.

We reported a net income of $123.1 million or $0.27 per diluted share last quarter compared to net loss of $124.8 million or negative $0.42 per diluted share in the third quarter of last year. We also booked a $343.1 million gain on digital assets, including Bitcoin receivable during the third quarter of 2025 reflecting the positive impact of the Bitcoin holdings on our balance sheet. Now let's talk about our cost structure. Our purchased energy cost of Bitcoin for the quarter was $39,235 and our daily cost per petahash per day improved 15% year-over-year, which we believe at scale is one of the lowest in the sector.

This improvement is directly tied to our growing inventory of owned and operated sites, which now account for approximately 70% of our nameplate megawatt capacity. That transition supports our vertical integration strategy, but also pays dividends both financially and operationally. Since we do not control the price of Bitcoin we mine, minimizing the cost of inputs like energy are critical to the financial resilience and long-term success of the company. Next, I'll provide some insights into our Bitcoin holdings and digital asset management strategy.

MARA is the second largest corporate public holder of Bitcoin, and we seek to generate returns on our holdings as Bitcoin price appreciates. Our approach combines the potential for long-term Bitcoin appreciation with disciplined efforts to generate return while managing risk. Additionally, we have also used Bitcoin as a collateral to borrow under lines of credit. As of September 30, 2025, we held a total of 52,850 Bitcoin, including 17,357 Bitcoin that were loaned, actively managed, and pledged as collateral.

As such, approximately 1/3 of our total holdings were activated through our digital asset management strategy. In Q3, we issued $1.025 billion of zero-coupon convertible notes due 2032, extending our maturity profile and increasing balance sheet optionality. With additional liquidity, MARA gains strategic flexibility to act on opportunities, whether that's acquiring more Bitcoin, funding acquisitions, balance sheet management, or general corporate purposes. We have positioned MARA to act in response to market conditions in order to maximize long-term shareholder value.

As of September 30, 2025, we held over $7 billion in liquid assets, giving us the flexibility to fund domestic growth and pursue international expansion. To streamline our communications starting in Q4, we will share our production on a quarterly basis. Investors can continue to monitor our monthly MARA Pool production in real time on the mempool. As we have stated previously, electrons are the new oil, and we are laying the groundwork for 2026 and beyond.

We're executing on a pipeline of energy infrastructure projects, both in the U.S. and internationally, and we expect these investments to expand our capabilities while keeping costs low. With that, I'll turn it over to Reggie Smith from JPMorgan to begin our management interview. Reggie?

I appreciate you selecting me for this call here. I have a very big announcement this morning. I guess kind of help me interpret this morning's announcements versus, I guess, kind of your prior strategy. Like what's being emphasized, deemphasized?

Maybe talk about that from the -- like what's the most emphasis you're placing on the business and maybe the least, because there's a lot going on here, certainly relative to the other Bitcoin miners. I know the other guys, it's either Bitcoin mining or kind of colocation. You guys seem to have a lot more balls in the air. Maybe talk through those differences. [Technical Difficulty]

Hey, guys. Is everything okay?

Yes. I didn't hear anything. Did you catch my question?

Sorry about that. Sorry. I had a comms issue here. I'm in the U.K., so it was a little bit of a problem.

I'm back on now. So yes, I heard your question, Reggie. Sorry. So if you think about the deal we announced today, it's about getting access to low-cost energy that is reliable, available 24/7, where because we are the generator, it provides us with a very low cost.

If you look into the details of the announcement, you'll see that the pricing on the gas is amongst the lowest in the market. The other thing is that it gives us now the capacity to add potentially up to 1.5 gigawatts of data center capacity if we want to, which gives us lots of flexibility. A lot of our Bitcoin mining sites are very attractive to use for inference AI, as we discussed earlier. We talked about what we're doing at Granbury and what we'll be able to do at some of our other sites in a similar fashion where we can blend inference AI and Bitcoin mining.

But the relationship with MPLX and the opportunities it provides give us a much broader canvas that we can paint on, whether that is traditional HPC like some of our peers have done or whether we want to build it out as hybrid AI, inference AI, and Bitcoin mining sites. So it gives us a lot of flexibility. And we believe controlling and owning power is a core part of any company that operates in the digital infrastructure space. When you look at the spending that's going on, and I think Sachin Ardell said this, in a recent podcast that was quoted, where he was quoted as saying that compute isn't the constraint, energy is the constraint.

And so access to energy, we believe, is critical. We think inference over the long term is where all the value is going to get created in this space. But we believe that Bitcoin mining has a very important role to play in not just balancing grids, but providing a flexible load when mixed with AI, such that AI can begin to operate in more places than it does today. And the last thing I'd say is that we believe that the technology curve is going to move so quickly in this space because you have to realize that just like in Bitcoin mining, where cost per petahash is the most important metric that drives profitability in the AI business, unless you are in the application layer.

In other words, running is the -- owning the data and the application that is generating value for the enterprise. In healthcare, owning the healthcare data, running the actual AI analysis. The only thing hosting providers and model operators provide are tokens in the sense of we need lowest cost per token if we're going to use that service. And using the APIs from the cloud providers is a very expensive way of running AI.

And most enterprises today are being confronted with the fact that the cost per token is too high using existing systems, and they want to move to lower-cost systems. And we're going to start seeing, and we already are seeing ASIC-based solutions coming, open-source models, all of which will allow enterprises to build their own and operate their own private cloud or use those services from third parties, allowing them to drive value from AI. So I think for a lot of the big guys, the challenge is they are doing deals with colocation partners where they are not taking on the debt. The debt is being laid on the joint venture or the SPV related to that colocation facility.

And that colocation partner is having to deploy a lot of capital to build those sites and equip those sites, and you have technology obsolescence. Over the course of a 10-year lease, you will have to upgrade the hardware in that location. And you have to estimate that in the cost of what it's going to be to build and operate. And I think there's a risk potentially that $1.4 trillion of data center contracts signed by OpenAI over the -- that will have to be operating in the next 5 years according to what was recently reported in the press, that some of that may not actually be able to come online and generate revenue.

So I think our approach is much better, more prudent, certainly much more capital efficient. And by being at the end of the spectrum where we're vertically integrated and able to operate at lowest cost per token and deliver lowest cost per token, we will have a significant advantage in the marketplace.

And Reggie, just a reminder, we -- today, we control approximately 2 gigawatts of capacity. And this added capacity is incremental to that, that takes us to close to 3.5 gigawatts over a period of time.

Got it. Understood. I'd like -- Fred, I appreciate the color there. And I was doing some kind of light math this morning.

And I think about, I guess, kind of AI and HPC, you made a comment in your shareholder letter about the price of power and the price of compute. You made some parallels between Bitcoin mining and HPC. And I was looking at the numbers, and I think they may be a little bit off, but directionally, this is, I think, a fair statement. When you look at Bitcoin mining, the price of power and the price of the actual ASICs, if you think about depreciating per hour, are about the same, like a 1:1 ratio there.

For GPUs, that ratio is more like 1 power 10 GPU. So like depreciation charge, and depreciation is super high. So you talked about ASICs and somehow, I guess, driving the cost of the hardware down there. Am I thinking about that right?

Like what are you seeing? And kind of where do you see the role going there?

Listen, just think about it this way. When Bitcoin mining started, we were running CPUs, right? Then we went to GPUs, then we went to FPGAs, then we went to ASICs. And when you look at the amount of compute power for -- think of it as the number of terahash we could produce for a jewel of energy, it has dramatically changed.

So you are now processing many more calculations at much lower cost of energy. And in our business, we depreciate the compute over 3 years. So if you're a hyperscaler and you're signing a deal for 10 years, some of these are 15 years and the depreciation schedule is 5 years for the machines. Does that mean they're going to have to replace those machines 3x in that cycle, right?

And to your point, GPUs to power is most probably a 10:1 ratio. And as you get to ASICs, that starts dropping and power starts becoming an even more important component. And when you start looking at the end cost per token, at that point, the model cost also comes into play. And so if you have open source models, if you have low-cost hardware that's energy efficient, you're operating in data centers that don't cost you $10 million a megawatt to build, you start getting to economics that start resembling Bitcoin mining over time.

Okay. Understood. Now, help me understand this. I wanted to understand or make sure I'm hearing you correctly.

When you think about kind of the investment risk and the CapEx risk within this chain, obviously, you've got guys that are building data centers, you've got people that are buying like GPUs and hardware. And then you say, obviously, you got the model guys as well. But I guess your comments on kind of where the CapEx risk is greatest, are you suggesting that the people that are buying the machines are taking on the most risk? Or do you think there's still substantial risk in building big data centers?

And I ask you that in the context of, I mean, you guys just, I guess, bought a few GPUs yourself. And so like help me square all of this together to understand kind of what your view is there.

Right. So part of the question is, are you in the business of being a bare metal shop, right? You're providing essentially hosting and GPUs. Look at what Iron is doing, right, bare metal.

Somebody has to load their software on it, but they're renting capacity on GPUs effectively. And that's what GPU cloud is called. In that case, the operator is funding the GPU purchases, right? In the case of a colocation, there are some deals that have been done where the operator is funding the GPUs.

And there are other deals where the lessor of the space, if you would, is bringing the GPUs, and they are the buyer and operator. So if Microsoft comes and is going to contract with you to just buy capacity from you, they're going to bring the GPUs, hopefully. You would hope at least. And they're taking that risk.

But there are lots of different models out there being operated by people. What we're doing with inference at the edge is much more around providing inference AI, which is not running on GPUs. We're running on ASICs, ASIC type solutions. And so it's a very -- it's a different model from a hardware cost perspective.

It's air-cooled. It's not liquid-cooled, for example, which means your infrastructure is much less expensive. You're not having to spend many millions of dollars per megawatt on building infrastructure, specialized cooling infrastructure. And all of that adds up to the economics of what you can do.

But inference is also done at smaller volumes, right? You don't have to do 100-megawatt sites yet. Most of the needs for inference still are quite young. It's early in the market.

But if you believe what Gartner and the analysts say, over the next 3 to 5 years, inference will be the primary generator of revenues and value creation within the AI space. So that's where we're swimming.

Understood. I'm going to skip around a little bit here. I wanted to talk about Exaion. And kind of loop it back into the broader discussion. But obviously, you guys announced that acquisition. Help me understand what they do today? And maybe talk about the scale of their operations. Like are they running data centers today? And if so, what's the size of those? Like what do they do exactly?

Yes. Exaion is today, until we close, a fully owned subsidiary of EDF that operates EDF data centers where all of the data for the nuclear fleet operates in this process. So they run EDFs, AI and traditional data centers across the EDF enterprise. They have about four data centers today, three in France, one in Canada.

They also operate quantum technology in the Canadian data center, which is made available for research purposes. And they have built a whole set of software solutions that allow you to operate the data center, store data in full private mode, meaning the users' data is fully encrypted. Exaion doesn't have the keys to that data. And so, were that data center to be broken into, if you think of -- if somebody were to steal data, the data in the data center is encrypted.

So it's the customer who holds the keys to that data. And so it's a way to build private cloud solutions that are fully secure. And so the whole reason for making the investment in Exaion is it gives us access to a team and an existing set of data centers that are Tier 3 and Tier 4 already. They know how to operate the most sensitive data.

They know how to protect it. They have existing customers, so they have experience, and we are going to leverage their knowledge, their experience, their technology, and their platforms to expand what they do on a global basis.

Got it. So they're asset like. They're more of a service layer, their engineers, their software, things like that. Like they don't actually own any data centers. It's really running that data center, securing data. Is that the right way?

Yes.

Understood. Okay. Is there a way to frame it, maybe early, their revenue run rate? And interestingly about that transaction, I think the first 64% of the transaction you bought them for $168 million. The next 11% will be at a much higher rate. Like what was the thinking there? Any opinions you can provide there.

I mean, I think you can think of how many times deals like this are structured. You're paying for a portion of the business based on where it is today. And then the growth opportunity for the existing investors is in executing on a plan to help grow the business. And therefore, you're going to pay a higher multiple for that. That's how you should think of it.

Understood. Okay. Now, I want to tie all this back. So the MPLX transaction, real quick on that. Does it require any like ERCOT approval? Let's say these guys have the natural gas. You guys would make the power plant or the generation assets and then the data center. But is anything needed from ERCOT? Any roadblocks there? And like how quickly could you have a data center up and kind of running?

Yes. I think you have to think of it more as the first thing we're doing is building a power generating station, which will be gas-fired power plant. So you have regulatory requirements around air permits, for example, which in the current political environment should not be exceedingly difficult to acquire. We feel fairly confident that we'll be able to get those without much problem.

So once you built the power plant, then because you are not directly grid attached yet, you then have to apply to attach to the grid and be a provider to the grid. So there's a regulatory process for that. Meanwhile, you can be producing energy and operating data center fully behind the meter. And ERCOT gets involved when you connect to the grid or the utility does once you connect to the grid.

So -- and the goal here, what's really important to remember about this MPLX relationship is it gives us the ability to own and operate gas-fired power plants with very low-cost gas with the ability to colocate large-scale data centers with reliable 24/7 power in a very attractive part of the marketplace. And so it gives us a lot of control to really drive our growth in a very cost-effective way. And I think it positions us very well, come what may in this HPC AI market and give us a lot of opportunities to really operate and continue to generate a lot of value for our shareholders.

I agree. I've been thinking about this idea of like vertical integration, and I didn't know if it was going to be a power company acquiring data center capabilities or the other way around. So this is very interesting. If I could dig in a little bit more.

So I think you talked about 400 megawatts of capacity to start. How should we think about like the minimum effective dose to kind of get started. So I don't know if you want to commit to 400 megawatts right off the bat. Is it 20 megawatts?

And how quickly can something like this come together? And then I know it's early days, but we've heard estimates of up to $10 million per megawatt to build out a data center. Like what are you thinking about from that perspective as well?

So you don't build a power plant in 20-megawatt increments. You build it right to a certain size at each site. So there are three sites. We'll likely think of it in 100-megawatt increments initially, but you have the ability to scale these plants much larger.

As it relates to the data centers, we have the optionality. We can build these as traditional Bitcoin mining data centers that are fully containerized at somewhere around $1 million a megawatt, including hardware costs for compute. If you then want to look at going the AI route, if we're doing it similar to how we're running the inference AI we're running today, the actual infrastructure cost is very similar. It may run a little bit more expensive depending on the cooling technology, if we use direct-to-chip cooling or we continue to use air cooled.

And if you use direct-to-chip cooling, your cost of infrastructure will end up somewhat higher. But the key is we're not building buildings that take 3 years to build. We're doing these as modular containerized solutions, which gives us full flexibility to reconfigure a site depending on whatever we want to do at it. And I'm a big believer that you will see very high-performing HPC capable modular solutions on the marketplace within the next 2 to 3 years, where you will be able to deploy the same sophisticated solutions you're building in these very sophisticated data centers where people can run some of the most sophisticated AI they need to.

Remember, there are not many customers in the world who need data centers of the scale that OpenAI needs it, right? OpenAI needs a lot of compute capacity because of the breadth of data and the breadth of the solutions their models operate. If you remember what DeepSeek did and how DeepSeek created the stir in the market, it's because instead of operating with a broad foundational model, they only load into memory specifically the model segments that they need and the data to solve the query, which means you now don't need all of that scale. So what I think will happen in the marketplace is that you're going to have efficiencies in models going to open-source, clients developing their own models and training their own models because the clients don't want to give the data to OpenAI.

If I'm -- and I'll give you an example. I was at FII last week in Saudi Arabia, and I was sitting with the Head of Strategy for Aramco on a panel. And they don't put their seismic data in the cloud. They're not going to do that.

What do they do? They build their own models. Other companies do the same thing. Look at what Lockheed just did the deal they just did with Google, right?

It's an on-prem solution. You are not -- I'm not going to put my data up into your cloud Google. You're going to build a cloud instance on-prem, on my site that is air gapped from your systems. That's what corporations want.

They want data sovereignty. They want private cloud. They don't want to run up in Meta's cloud, Amazon's cloud, or OpenAI's systems. 70% of corporate data today is still not in the cloud. There's a reason for that.

And I think when you look at inference, inference is driving insights from the data that runs your company, right? It's -- if you're in the healthcare business doing drug discovery, it's all the patient data, the lab samples, et cetera, all that data, you're driving insights from it, right? And if you are doing -- you're building airplanes, it's all the design data and the manufacturing data. If you're running a factory, it's the operations data of the factory, right?

If you're running a power plant, it's the operations data of that power plant. You don't want to run that off-site. You want to actually run it on site because as those systems become mission-critical and actually operate the resources and operate parts of the business, you can't take the risk that you have a system failure that brings your whole business down just because you lose a link to a cloud or Amazon goes offline like it did the other day. So I think it's -- people really have to understand that there is a limit to what data and how much risk people want to do in putting their core critical assets into a cloud operated by a third party.

And if they can solve the model issue and do it at lower cost, near-prem or on-prem in a private environment, they will do it. And I have been speaking with the heads of AI for major corporations in the financial market today who tell me that they are relocating AI systems out of the cloud back to near-prem, on-prem private solutions because it is significantly less expensive to operate than doing it in an Amazon Cloud or other places like that. And I think that the analyst community really needs to do a much better job of talking to the enterprises who are the users. These are the people who are actually going to pay the money that will allow OpenAI to be successful or not, that will allow Microsoft to be successful or not.

You can talk till you're blue in the face with the people building these things, but it's like building railways. If there isn't passenger traffic and there isn't cargo, the rail lines fail. So I hate to be a downer on this, but this is an important thing that a lot of people aren't doing. You need to talk to the customers.

Who's going to pay for this stuff?

And I want to make sure I'm hearing this right and connect these dots. I think you mentioned kind of a smaller kind of, I think, a 1 megawatt, what do you call it, I guess, kind of like a sample or a small micro data center, pilot site. If I'm hearing this right, are you saying that like that could become like the prototype for enterprises having their own on-premise like AI capabilities? Is that what you want to say?

Yes. So think of it this way, right? I'll give you an oil drilling example, right? So you have an exploration drill that's drilling, you have seismic data.

Today, you have to plan exactly the drill profile and what some -- what the drill operator is going to do. And so the oil companies have built these very sophisticated AI models that run in a modular container typically out on the drilling site that are collecting real-time data from the drill and then feeding back instructions into the drill master. That's an existing example. You can go to a trading -- a financial trading company.

And their whole thing is speed and latency. They want their systems operating on their local network, not on a wide area connection where there's latency because 25 millisecond delay in a response means they lose the profit on a trade. And so there are -- whether you're looking at defense, which is going to be a huge growing sector when it comes to AI, just look at the amount of AI that's needed to operate in any theater of war today, look at healthcare, look at manufacturing and production, look at the movie television industry. The single largest consumer of tokens in AI are video illustrations and audio generation.

Those are the systems that consume that these diffusion models are the single largest consumer of tokens. And so cost per token is very critical to them because if you're going to generate a 5-, 10-, 15-minute clip of video, it takes multiple factors of magnitude more tokens than asking OpenAI where you should eat lunch today. And so I think, again, the marketplace gets all hyped up about these big contracts, but they really need to look at who's actually going to use this stuff. What are they going to use it for?

What can they afford to pay for it? What will the pricing trends be over time? And to use the worn-out Wayne Gretzky technology. If you're in our business, you want to be skating to where the puck is going to be, right?

You don't want to be chasing the puck. And I think there are a lot of people announcing deals out there, getting on the bandwagon to pump their stock when they need to look at what's this industry going to look like in 5 years.

That actually leads to my next question, Fred. So thinking about announcements and catalysts, like what should we look for from MARA to know that like this strategy is taking form and we can start to frame an economic story or a creation story around some of these initiatives. Like what are the milestones and announcements we should be looking for from you guys?

So here's what I think you should look for. 4 years ago, I made a presentation at a conference where I said that Bitcoin miners are either going to become energy companies or be owned by energy companies. I think what you should look for is when large energy companies start signing partnership agreements with companies like us to monetize their energy assets at large scale. That will tell you that if that happens to be us, that they have chosen us to do it with because they feel we are the best option for them to maximize the value of the electrons that they produce. That's one step.

The next step is as you start seeing customers using more and more inference AI and you see us reporting a greater and greater mix of inference AI in our data centers. And the real metric you should look for is what is our profit per megawatt hour that we talk about. It's not a GAAP measure, so it's not going to be reported that way. But you can think of it as an operational KPI where the profit we can generate from every megawatt hour of energy that we consume or produce is a data point that our investors will be able to see and that will directly correlate to our profitability and ability to have a cash-generating business.

Okay. And just to make sure -- and I apologize, this is a silly question. Are you looking to sign colocation clients or deals for this site in West Texas? Or is this something that you're thinking about putting your own machines in?

It gives us -- I'm not going to answer the question directly because I think our competitors spend more than enough time listening to what I say and then emulating it. So I'm just going to say it this way. It gives us maximum optionality to decide what we want to do with whom.

Got it. Okay. Because I want to bring it up because you mentioned signing a colocation as like a milestone and...

No, you see if I can operate inference AI and make money on it without signing a colocation facility that will give you a little bit more insight into what the business model might actually be. Because think about it, the best thing about our Bitcoin mining business is we don't have a customer. What's the hardest thing all these colocation deals have is they have to go find a customer.

Yes. Okay. Okay. People say, I change my opinion when the facts change.

And this is a pretty -- this seems like a pretty major shift for MARA. Like I said, you guys bought GPUs, I guess, in the last 3 months and you start to run them. Like what in your mind has changed -- that has changed your opinion or has your opinion changed? Because strategically, it seems like the company is kind of pivoting.

Talk to me about that. Like what have you learned or gleaned in the last couple of months or quarters that has driven this shift?

Yes. Reggie, I wish I could tell you that I had a lightning bolt strike me and I came to an epiphany, but this is -- we're executing the strategy we decided to execute over a year ago. It's just we have decided not to go totally open with the market and tell people what we're doing because it just gives our competitors insight into what we do and they can emulate it. And we prefer to control the timing on how we talk about what we're doing.

But I've been talking for the longest time about inference at the edge, and that's where we would make our mark in the marketplace, and we are. We've talked a long time about owning power and the desire to run our business based on controlling energy assets so we're fully vertically integrated. And we're doing that. There's no change in strategy.

There's no pivot. It's just we have been purposely operating more like a start-up in the sense that we have really wanted to make sure that we had everything in place so that as the market becomes aware of what we're doing, they just start seeing kind of announcement after announcement after announcement that just gives them more and more confidence in that we're executing on the vision that we set out a year ago.

Yes. No, I'd say from where I sit and I think about all the pieces you guys have. There are a lot of pieces, and I'm not smart enough to figure out how to put it all together, but it seems like you guys have a lot of ways to kind of win here. I guess we just have to kind of sit back and wait for those announcements as they kind of come through.

I know we've kind of spent a lot of time on this. I hope it wasn't a wasted time for people. Maybe we could shift gears a little bit and talk about your like sovereign and foreign government initiatives and things that are going on there. Like one of the questions I have, as you think about this is like what do you think gives you guys a right to win in the sovereign compute kind of load management space versus competitors?

Who's even competing with you there?

I think -- so here are a couple of ways to look at it. Most of our competitors enter a marketplace by partnering with somebody or contracting for power. They don't bother talking to the government because they're afraid that if they do, they may not be allowed to do what they want to do. And that's the case in a lot of places in the Middle East.

We, on the other hand, chose to do it the other way. So in UAE, where we've been operating now for a couple of years, we chose to directly go and work with the sovereign. So we partnered with ADQ and IHC and operate a joint venture together with them where we balance the grid in UAE. It's one of the most advanced liquid immersion technology sites in the Middle East.

The only one that's bigger than that is the liquid immersion site we operate in Granbury. And so that has given us a reputation of being somebody who works well with government entities, follows the rules, and is focused on being a good grid citizen and balancing the grid. So when we talk to people in other countries, such as in France, such as in the U.K., such as in Kenya, in Saudi Arabia, in other places, we are welcomed with open arms because we are focused on how can we make your grid more efficient and more effective. How can we make sure that every electron your generators generate -- sorry, generate maximum value.

And we are here to be a good grid citizen, and we are here to operate such that your grid becomes more stable and it becomes easier for you to bring on new types of loads, be them AI data centers or whatever. And the challenge, the way most people see it is that takes time. I have been crossing the Atlantic very frequently, but I have been having meetings in the top levels of government, and we have a lot of support. We certainly have seen a lot of support on the European side because there are certain dynamics in Europe that create very large opportunities for us.

And so same thing exists in Saudi Arabia, for example, and other places. And we think that it's worth our effort to spend the time and take the time to do this carefully and prudently and well thought out so that we're able to execute successfully and have long-term success in the countries. Because if we're friends with the government, then we have the advantage that as they look to expand what they're doing, if we are a good partner, they will come to us and say, "Hey, we want to do more with you." And that's the type of relationship we want to have with our partners across industry, be it governments, vendors, or end customers.

And it's funny, we haven't talked about Bitcoin mining at all. I know we're running short on time. Just an update there. Love to hear about the stuff that's happening at the wind farm and some of your flared gas initiatives.

Maybe talk a bit about Auradine. And then I guess, your plans for growing hashrate here and how you think about that in the context of kind of where hashprice is and why it makes sense to continue to grow your hashrate at these levels?

Yes. So maybe I'll look at this in kind of a somewhat reverse order. So there is more hashrate coming online every day from lots of players. There are very well-capitalized companies who are not public, who are -- have access to huge amounts of capital, who have a stated goal of becoming the largest Bitcoin miner in the world.

And if we don't grow our hashrate, we will have an ever-decreasing amount of the global hashrate and produce ever decreasing amounts of Bitcoin. And we think that it's our duty to continue to grow hashrate, not just in the United States, but globally to support the security and diversity of the Bitcoin blockchain and the Bitcoin network because we don't want it to be dominated by any small handful of players. And so we believe it's our duty to continue to grow hashrate. So how do we do that economically?

We do it with low-cost power, which we can control, which ties to the MPLX deal. It ties to what we're doing with our wind farm, Texas. It ties to what we're doing with flare gas. We have doubled -- by the end of this year, we'll have doubled our flare gas capacity, and we're going to continue to grow that.

The wind farm is fully built out from a data center perspective, and that's running. And we're going to continue to look at opportunities to acquire more energy that is low cost that we can then allocate between Bitcoin mining or AI. You have to kind of think of us as we are going to own lots of electrons, and we're going to put those electrons to best possible use. In regards to Auradine, Auradine's more recent hydro model, which competes very well with the Bitmain and other vendors' models is doing well.

We're deploying Auradine in our fleet. We're not deploying exclusively Auradine at this point. There are still different machines have different characteristics that are really good for different environments, and we have a lot of different environments. And so we're continuing to deploy a mix of systems.

But over time, it would be logical to feel that we're going to add more and more Auradine to our fleets. Their systems offer some very unique capabilities, especially around load balancing that in a model such as the one that is beginning to gain steam in Texas, where the utility wants to regulate your curtailment and shut you off and turn you on, that requires special capabilities in the miners, and that's something that exists in the Auradine systems. And so as more and more utilities start looking for those capabilities amongst miners who are on grid, I think they will continue to gain some market share there.

Other than that, they have spun out some very interesting AI-related businesses, One or Escape, which is around securing large language models -- sorry, which recently had a lot of positive reviews at the RSA show earlier this year and then also ScaleUp, which is a start-up around ultra high-speed cluster interconnect switch technology. So that has been a great investment for us, and we continue to look for investments like that where we can acquire or build technologies that can become part of our solutions over time.

I guess last one for me. You kind of talked about it earlier, but obviously, a lot of market cap, a lot of value has been created in the Bitcoin mining space amongst the publicly traded guys. I'd argue that you guys haven't received or gotten your share of that. Like what do you think the market is missing and hopefully will come to appreciate in the near term or medium term?

I think.

About that specifically.

Yes. I mean I think the key for us is the floor on the valuation of our stock is essentially the value of our Bitcoin holdings. And people don't put a lot of value on the Bitcoin mining infrastructure or the Bitcoin mining business per se. And I think as our business continues to evolve, especially with the energy generation story and as AI becomes a bigger piece of this and we generate more profit per megawatt hour consumed, we'll start getting more attention from people.

And I think you'll start seeing people realizing really the benefit of what we're doing in our model, and we'll get more credit for that.

Reggie, just to add to that, the power capacity that we have secured through these transactions that puts us at the forefront. And here's the actual value flows with Bitcoin mining option value between AI-ready assets, our operational flexibility with integrated power, that's what's going to drive value for our stockholders from a long-term perspective.

Perfect. Congrats on the quarter.

Thanks, Reggie. We appreciate it. Most of the questions that we received from our retail shareholders have been answered. We're obviously running short on time. But thanks, everyone, for joining us today. If you have any questions that were not answered during today's call, please feel free to contact our Investor Relations team at ir@mara.com. Thanks very much, and enjoy the rest of the day.

Ladies and gentlemen, thank you for your participation. This does conclude today's teleconference. Please disconnect your lines, and have a wonderful day.