Earnings Call Transcript - KEEL Q3 2025

Operator, Operator

Thank you for standing by, and welcome to Bitfarms' Third Quarter 2025 Earnings Conference Call. I would now like to hand the call over to Jennifer Drew-Bear from Bitfarms' Investor Relations. Please go ahead.

Jennifer Drew-Bear, Investor Relations

Thank you, and welcome to Bitfarms' Third Quarter 2025 Conference Call. With me on the call today are Ben Gagnon, Chief Executive Officer and Director; and Jonathan Mir, Chief Financial Officer. Before we begin, please note, this call is being webcast with an accompanying slide presentation. Today's press release and our presentation can be accessed on our website, bitfarms.com, under the Investors section. Turning to Slide 2. I'd like to remind everyone that certain forward-looking statements will be made during the call, and that future results could differ from those implied in this statement. The forward-looking information is based on certain assumptions and is subject to risks and uncertainties, and I invite you to consult Bitfarms' MD&A for a complete list. Please note that references will be made to certain measures not recognized under IFRS and therefore, may not be comparable to similar measures presented by other companies. We invite listeners to refer to today's press release and our MD&A for definition of the aforementioned non-IFRS measures and their reconciliations to IFRS measures. Please note that all financial references are denominated in U.S. dollars, unless otherwise noted. And now turning to Slide 3. It is my pleasure to turn the call over to Ben Gagnon, Chief Executive Officer and Director. Ben, please go ahead.

Ben Gagnon, Chief Executive Officer & Director

Good morning, everyone, and welcome to Bitfarms' Third Quarter 2025 Earnings Call. We made strong, steady progress in Q3, building on the momentum from the first half of the year as we advance our transformation into a leading North American HPC and AI infrastructure company. Today, I'll walk you through our investment thesis, value proposition and key developments, including updates on our energy portfolio and site-specific advancements, all of which gives Bitfarms a competitive advantage to capitalize on the surge in demand for HPC and AI infrastructure. Turning to Slide 4. I would like to kick off today's call by outlining our market thesis, one that we believe differentiates us from our peers and best aligns Bitfarms with long-term investors in our transition to HPC and AI. Infrastructure is not a bubble. Since the invention of modern compute, the supply of compute has increased exponentially. As compute grows, so too does the data center industry that powers it. This is a trend that has a trajectory of over 20 years of exponential growth and an annualized growth rate of 8.8% behind it. This isn't a bubble. It's a reflection of a new paradigm that showed no signs of slowing down before AI and now as AI rewrites the rules of how humans interact with computers, the demand for data center capacity is accelerating. But the demand for compute and infrastructure has reached an impasse. Data centers that used to be measured in kilowatts are now being measured in megawatts and gigawatts. Racks that used to support 10 kilowatts are now being designed to support 370 kilowatts. The exponential increase in demand for power can no longer be met at the pace the market demands. And as a result, the lease rates for data center infrastructure, which have grown at an average rate of 3% over the last 20 years, are now growing at an average rate of 12% since 2022, and we expect this trend to continue. Turning to Slide 5. Infrastructure is a bottleneck. As manufacturers continually introduce newer, more efficient chips and increase production every year, this trend continues to accelerate. Next year, NVIDIA alone is expected to be shipping somewhere between 10 and 15 gigawatts of GPUs. And that doesn't include, of course, AMD, Intel, Qualcomm and others who are also producing their own hardware with over 100 gigawatts of chips expected to be produced by 2030. While the supply of compute chips continues to increase, the growth in data center infrastructure is happening at a much slower pace. It is not silicon nor capital that will be the real bottleneck for continued growth in HPC and AI, but power and infrastructure. Over the next few years, the gap between the amount of chips that are being produced and the megawatts and the racks available to plug them in and operate them will continue to widen significantly. We strongly believe that as this dynamic continues to play out, the value and the economics will continue to move in favor of those who own the energy and data center infrastructure. We've watched this play out in the market with the contracts that have been announced in the industry to date. When Core Scientific and CoreWeave announced their landmark transaction in April of last year, the rates were contracted around $120 per kilowatt per month. As we've moved further along this curve that's shown on the slide, those rates have continued to trend upward. Most of the contracts over the past few months have been around $150 per kilowatt per month. As time goes on, this trend is expected to continue with analysts predicting a massive shortfall of nearly 45 gigawatts of power for data centers by 2030. Just within the last two weeks, Satya Nadella, the CEO of Microsoft, confirmed the shortfall when he publicly stated on a recent podcast that they have GPUs they cannot deploy. We believe that over time, the companies who've allocated and will continue to allocate billions of dollars into compute will be increasingly economically incentivized to pay rising prices in order to deploy their compute faster and with greater certainty, because every day they do not deploy is a day of revenue they will never recover and because their customers will simply move on to a competitor. With direct operating margins for new GPUs typically in the 80s or 90% range, this infrastructure expense is a modest cost driver for those who own the compute, equivalent to a low single-digit percentage of OpEx. If this cost were to double, it would not impact direct OpEx for the customer by more than a low single-digit percentage. These rates, which are largely inconsequential for the customer, are very significant for Bitfarms as the developer. With OpEx costs that are largely fixed, every additional dollar earned in a lease goes to the bottom line. This is what Bitfarms is aiming to optimize for, not the fastest contract, but the highest value per megawatt and the greatest margins for the longest period of time with great customers. We believe this will be the primary driver of our multiple expansion and what drives shareholder value creation long term. Our investment thesis is clear and backed by decades of data. Our conviction is high, backed by consistent incoming demand. We don't want to cap our upside by signing leases prematurely. Instead, Bitfarms plans to optimize and achieve higher lease rates and margins through the following three strategic actions: one, prioritize infrastructure development first by minimizing the time between signing a lease and generating revenue for a customer, we will minimize the discount that would otherwise be applied to the lease rates and locked into multiyear contracts; two, take advantage of the increasing gap between supply of data center infrastructure and data center demand to lock in higher rates and greater margins under multiyear agreements; and three, while the industry is focused on NVIDIA GB200 and GB300, Bitfarms plans to leapfrog NVIDIA's Blackwell architecture and lead the industry in developing infrastructure for NVIDIA's next-generation Vera Rubin GPUs across 99% of our 2026 and 2027 development portfolio. With Vera Rubin GPUs expected to begin shipping in Q4 of 2026, and the infrastructure requirements to support them largely incompatible with facilities designed for Blackwell GPUs, we believe Vera Rubin infrastructure will be in the greatest demand and shortest supply in 2027 and will command significantly greater economics. Turning to Slide 6. We are able to take this approach because we have a robust balance sheet to fund development and know the value of what we own. While we don't have the largest portfolio of power among the public miners who are transitioning to HPC and AI, we do have the largest portfolios of power in each of the regions in which we operate, none of which are in Texas and all of which are either existing or emerging data center hubs. With consistent inbound demand for our sites, we have high conviction in the value of our unique energy portfolio, the demand for our power and our ability to develop next-generation HPC and AI infrastructure. We believe that not all megawatts are created equal. Our megawatts are strategically located in high-value areas that have multiyear waitlists to secure the power we have today. Our campuses are close to major metros and existing data center clusters, have ample access to major fiber trunk lines and undersea fiber optic cables and benefit from a temperate climate compared to places like Texas. While Texas is undisputably a great energy market and arguably the easiest market to grow and develop megawatts in the U.S., there are, of course, trade-offs. The trade-off to short-term development efficiencies is long-term operating inefficiencies. It is no secret that besides power, the primary challenge with data centers is cooling and cooling is becoming an increasingly more difficult problem to solve as energy density continues to increase with every generation of new hardware. Building and operating data centers in a hot, arid desert climate like Texas as opposed to cooler northern climates like Pennsylvania, Washington and Quebec means more CapEx and OpEx for cooling. This isn't an opinion. It's math and engineering. If we built our exact same data center for Panther Creek with the same design, equipment and materials in Texas, it would have a PuE of about 1.4 to about 1.5. Whereas in Pennsylvania, Quebec or Washington, it would be about 1.2 to 1.3. That means for every megawatt we are converting, more of those electrons are going to compute, which is the revenue-generating activity for customers as opposed to supporting revenue generation through cooling. Simply put, our megawatts are harder to get, in higher demand areas, produce more value for customers and are worth more per megawatt. In Pennsylvania, we have the strategic foresight to acquire our three campuses and submit our energy applications in 2024 before the HPC and AI demand really came into play in the state earlier this year. This has positioned us with secured power at Panther Creek and Sharon and at the front of the queue with very well-advanced power applications at Scrubgrass. In Quebec, new power allocations are almost impossible to get with numerous data center applications denied by the province in the past year. Bitfarms has 170 megawatts operating with some of the cheapest power rates for data centers in North America and 100% renewable. One hundred percent of these megawatts are currently being utilized for Bitcoin mining. And just in the last month, we confirmed that we will be able to convert our Bitcoin megawatts for HPC and AI. This means our Quebec portfolio represents a unique and strategic opportunity to increase total data center megawatts in the province by 25% from about 700 megawatts today, while fulfilling two strategic national and provincial objectives: scaling back Bitcoin mining megawatts while increasing HPC and AI infrastructure and data sovereignty. In Washington, we have 18 megawatts of secured power in the largest data center cluster on the West Coast with the cheapest power in the U.S. for data centers and 100% renewable. Because of this, the area has a 10-year waitlist for power. Everybody is looking to grow here, and it is nearly impossible to do so outside of secured megawatts like ours. This means that despite the relatively smaller scale of Washington, sites in the area are in high demand by both enterprise and hyperscalers alike. I'd now like to spend a few minutes discussing Washington and the news we issued this morning in more detail. Turning to Slide 7. Earlier this morning, we announced plans for the conversion of our 18-megawatt Washington site to HPC and AI workloads. We signed a fully funded binding agreement for $128 million for all the critical IT infrastructure and building materials to develop the full 18 megawatts of gross capacity with anticipated industry-leading energy efficiency between 1.2 and 1.3 PuE. The state-of-the-art facility will feature: one, validated reference designs, ensuring compatibility and performance with NVIDIA GB300s; two, modular infrastructure, enabling phased deployment and scalability, reducing the downtime of Bitcoin mining revenues and ramping up our time to HPC and AI revenues; and three, proven thermal and power management systems critical for HPC and AI operations. The construction team is in Washington today with the general contractor and are kicking off the conversion of the Washington site, which is targeted for completion in December 2026. Turning to Slide 8. I would now like to discuss monetization strategy at Washington. With decade-long wait times for new power and the cheapest power in the U.S. for data centers, we are actively pursuing colocation for both hyperscaler and enterprise, where we can capitalize on the long wait times as previously discussed. This morning, for the first time, we announced we are also pursuing GPU as a service or cloud. While our focus is on developing next-generation Vera Rubin infrastructure across most of our portfolio, we believe there are some compelling reasons to potentially go with cloud as a monetization strategy at Moses Lake specifically. One, GPU as a service would enable us to capture the benefit of the lowest cost power for data centers in the U.S. for ourselves and generate what we expect to be above-market margins and returns for cloud. Two, the relatively smaller scale makes cloud at this site easier to execute and finance. We have more than enough liquidity to consider the site and strategy fully funded today and are in active discussions with leading GPU manufacturers on GPU sourcing and financing, which we believe could be done on very attractive terms. GPU financing could materially reduce CapEx requirements and enhance expected returns. Three, we expect that by demonstrating our ability to execute across the entire stack, we will also be able to better understand customer needs, provide better quality service and negotiate better leases at our other facilities. Lastly, but most importantly, despite being less than 1% of our total development portfolio, we believe that the conversion of just our Moses Lake site to GPU as a service could produce more net operating income per year than we have ever generated with Bitcoin mining, providing the company with a strong cash flow foundation that would fund OpEx, G&A, debt service and contribute to CapEx as we wind down our Bitcoin mining business. I will now walk through the rest of our sites in a bit more detail, starting with Panther Creek. Turning to Slide 9. Panther Creek is our flagship HPC and AI campus in Eastern Pennsylvania. As we've discussed previously, we have 350 megawatts of secured power with PPL. This power is contractually obligated to be delivered with 50 megawatts at the end of 2026 and 300 megawatts at the end of 2027. The site has sufficient acreage for the development of the entire 350 megawatts with capacity to go beyond that. Additionally, we have $200 million remaining on our project facility with Macquarie that is intended to finance Phase 1 of the project as well as a few long lead time expenses for Phase 2. We also have some exciting news around potential further capacity expansion at Panther Creek. Lately, there have been a number of developments, including the recent 403 letter from the Department of Energy and commitments to deploy more natural gas energy generation in Pennsylvania that have given us line of sight to expand beyond the existing 350 megawatts of secured power capacity. We have received positive indication on converting our existing interconnection service agreement, or ISA, 60 megawatts to a firm energy service agreement, or ESA, of 60 megawatts to expand power to 410 megawatts and on a recent load study to expand power capacity to over 500 megawatts of growth capacity. With these positive developments that could meaningfully expand capacity at this campus and in line with our investment thesis, we are modifying our original Phase 1 designed for Blackwell GPUs and planning a new Phase 3 and Phase 4. The entire campus will now be developed for NVIDIA's Vera Rubin GPUs and their greater energy density to accommodate our new expectations on future expanded power capacity. This is expected to delay the energization of Phase 1 marginally from December 2026 into the first half of 2027, with no anticipated impacts to Phase 2 timelines. We believe this will enable the company to achieve significantly higher economics in line with our long-term thesis and strategy. Turning to Slide 10. Moving on to Sharon, where we have 110 megawatts of power secured by an ESA with FirstEnergy and PJM under development. We are currently operating 30 megawatts of Bitcoin mining on site, but have started development on an additional 80-megawatt substation, bringing the total available for HPC and AI uses to 110 megawatts. We expect to have the full 110-megawatt substation online by year-end 2026. We recently closed on the purchase of the land for the site, effectively ending our lease and enabling us to move forward with our planned development of HPC and AI infrastructure. Similarly to Panther Creek, we will be working to develop the campus for Vera Rubin GPUs, targeting site completion and revenue in the first half of 2027 for the full 110 megawatts of gross capacity. Turning to Slide 11. In Quebec, we have 170 megawatts of low-cost hydropower currently operating across multiple Bitcoin mining sites, almost all of which are within a roughly 90-minute drive from Montreal. This is an incredibly attractive opportunity for hyperscalers who are following what's called a regional campus strategy. This is something that was pioneered by Amazon, where smaller sites can be directly connected with direct fiber infrastructure in order to reduce the latency between sites below 2 milliseconds, enabling many sites to be connected together to function as one larger site. As I mentioned, it's almost impossible to grow organically in the province. And in October, we confirmed the ability to convert over our Bitcoin mining infrastructure to HPC and AI with regulators and utilities in the region. With that pathway clear, we are accelerating our plans in Quebec. We will focus our development efforts on the city of Sherbrooke, where we have 96 megawatts, robust fiber connectivity, a strong and developed local labor force and ample support from the local energy utility and municipality. We will be applying some of the standardized engineering and design plans completed for our Washington site to Sherbrooke in order to convert these facilities from Bitcoin mining into next-generation HPC and AI infrastructure adapted for Vera Rubin GPUs. Similar to Washington, Quebec has a cool climate and some of the lowest cost energy in North America for data centers. With strong unmet demand for GPU cloud in Montreal, Sherbrooke also represents a potential opportunity to scale up a cloud business in 2027 with VR200s, a strategy that we will evaluate as we work through the engineering and development plans for Sherbrooke. The remaining 74 megawatts of Bitcoin mining in the province are earmarked for potential expansion in 2028, and we look forward to providing more detailed plans for Quebec in 2026. Turning to Slide 12. Last, but certainly not least, we have our Scrubgrass campus in Pennsylvania. This is about 30 minutes away from our Sharon, Pennsylvania campus on the western side of the state. With the exception of the new Panther Creek Phase 3 and Phase 4, which I spoke to a minute ago, this is the only power in our portfolio that is not 100% fully secured today. This is a very, very exciting development opportunity for Bitfarms. We believe this is the only campus outside of Texas for public miners converting to HPC and AI that has over 1 gigawatt of potential capacity. And while we have made great progress on developing the power story for this giga campus, there are still quite a few steps to be taken in order to contractually secure the power, which falls into two buckets. First, we have completed three conceptual load studies with FirstEnergy, starting with 250 megawatts, 500 and then 750 megawatts, thus moving over to what's called a detailed load study with FirstEnergy, which would eventually be converted over to firm service in an ESA. Second, we have made substantial progress on evaluating the potential to add additional generating capacity on site. This could be accomplished by building a three- to four-mile pipeline from our campus to the second largest natural gas pipeline in the U.S., the Tennessee Natural Gas Pipeline, which we have confirmed could supply up to 550 megawatts of natural gas, multiplying our generation capacity on site. We're still in the early stages of evaluating how we would expand the generating capacity, and we'll provide more details as we progress. Combined, the two buckets could potentially provide 1.3 gigawatts of gross capacity. And additionally, there is very good fiber infrastructure in the area with our eight fiber infrastructure networks nearby and is in close proximity to Pittsburgh and Cleveland as well as the other data centers, which are starting to pop up throughout the state. The earliest time that we anticipate we could have additional power at this kind of scale implemented at Scrubgrass is around 2028. Though this is a longer lead time campus for us, we believe that with the forecast on power and demand for HPC and AI infrastructure, the timing for our giga campus will play-in well with the cycle, our investment thesis and our other development plans. Turning to Slide 13. To sum up, we believe that we are incredibly well positioned to execute against our investment thesis in 2026 and 2027 and maximize long-term shareholder value. One, we have a very unique portfolio of energy assets that we aim to fully convert to HPC and AI infrastructure. Two, we have announced our plans to convert our Washington site to HPC and AI workloads and lead the industry in the development of next-generation data centers for NVIDIA's Vera Rubin GPUs. Three, we are actively evaluating a potential cloud monetization strategy for our Washington site, which we believe would be a meaningful driver of cash flows and could eclipse any Bitcoin mining cash flows we have ever generated. Four, we are well capitalized to make our currently planned investments with a financial flexibility that exceeds $1 billion across cash, Bitcoin and our Panther Creek project facility with Macquarie, all of which are going to fund CapEx. As we continue to produce strong free cash flows from our Bitcoin mining operations that fund OpEx, G&A, debt service and contribute to CapEx with no further planned minor CapEx. And lastly, we continue to execute on our U.S. pivot with the anticipated sale of our Paso Pe facility and our full LATAM exit. Our transition to U.S. GAAP for Q4, the establishment of our New York City office and working towards a U.S. redomicile in 2026. We believe this would give us significantly greater index inclusion and meaningfully improve the institutional composition of our cap table. I now have the pleasure to hand the call over to our new CFO, Jonathan Mir. Turning to Slide 14. Jonathan, over to you.

Jonathan Mir, Chief Financial Officer

Thank you, Ben, for the warm introduction. I'm excited to join Bitfarms at this pivotal moment in the company's transformation. My principal objectives as the new CFO are centered around capital allocation, capital sourcing and capital structure. I'm working hand-in-hand with the operations and development teams on the ground to ensure we implement financing plans that are appropriate for the company and its assets, efficient and support long-term shareholder value creation and that we are also allocating capital to its best possible risk-adjusted returns. With an extensive background in energy infrastructure strategy and financing, I believe there's an extraordinary opportunity to use our strong balance sheet, unique assets and the talents of our people to create value in the high-growth HPC/AI space. I look forward to working closely with the team to deliver on our strategy and capture the exceptional long-term shareholder value that would accompany our successful execution. Turning to Slide 15. Today, Bitfarms has the strongest balance sheet and most available capital in the company's history. In Q3, we were able to execute across several initiatives. First and foremost, we recently completed a very successful convertible note offering, where we were able to upsize the offering to $588 million while improving on pricing, preserving upside and minimizing potential equity dilution through a 125% capped call. Bitfarms chose to issue convertible notes because they allow us to access capital at a lower coupon than straight debt and with less dilution than straight equity. The cash settled capped calls we purchased allow us to offset economic dilution up until $11.88 per share, representing a significant premium to the share price today. It is also important to highlight that investor commitment to Bitfarms is strong. One hundred percent of institutional investors that management met with during the marketing process participated in the transaction and invested their capital in Bitfarms. We're thrilled with the outcome of this raise, and it will allow us to advance our pipeline in tangible ways. Second, we converted our previously announced $300 million debt facility with Macquarie to a project-specific financing facility dedicated to the development of our Panther Creek data center. Moving the debt facility from a corporate level to the asset level materially enhances financial flexibility for the entire company. In October, we drew an additional $50 million from the facility in order to accelerate development of the site for a total of $100 million drawn to date. Finally, we maintained steady and efficient mining operations throughout the quarter, achieving approximately $8 million in monthly free cash flow after G&A. We expect to use this cash flow to support our HPC/AI development projects. Looking ahead, we anticipate continuing to use a mix of both corporate level and project level debt and equity financing as we advance our project milestones. On an ongoing basis, we will evaluate a wide range of opportunities and choose those that we believe support both a strong, stable balance sheet and realize the full potential shareholder value creation that would accompany the successful execution of our plans and fund milestone objectives. Turning to Slide 16. Let's focus now on our third quarter financial performance. In Q3, we achieved a total revenue of $84 million from continuing and discontinued operations. With the intention to sell the Paso Pe site in order to complete our Latin American exit, all revenue from that asset is classified as discontinuing operations. From continuing operations, we earned 520 Bitcoin and achieved revenue of $69 million, representing a year-over-year increase of 156% in revenue. For our continuing operations, our gross mining profit was $21 million, representing a gross mining margin of 35% and an average direct cost of $48,200 per Bitcoin mined. During the third quarter, we introduced a new program for digital asset management, Bitcoin 2.1, which is designed to offset Bitcoin production costs and achieve higher value per Bitcoin sold as a low-cost and low-risk funding mechanism for the energy infrastructure investments that define Bitfarms going forward. It is important to highlight that we are not a Bitcoin treasury company. The goal of this program is not to accumulate Bitcoin, but rather to offset the production cost of Bitcoin and by doing so, contribute to cost effectively funding our HPC/AI initiatives. This is a multi-strategy program that primarily sells both short and long-dated out-of-the-money covered calls on the Bitcoin and treasury as well as for Bitcoin production. During Q3, we incurred an all-in cost per Bitcoin of $82,400 from continuing operations. When considering our net gain of $13.3 million from derivatives against our all-in production costs, it would bring the effective all-in cost down to $55,200. Cash G&A for Q3 was $14 million compared to $20 million in Q3 2024. The improvement was largely driven by lower professional services costs. Operating loss from continuing operations was $29 million for the quarter, including impairment charge of $9 million of nonfinancial assets. As a result, net loss from continuing operations for Q3 was $46 million or $0.08 per share. For the third quarter, our adjusted EBITDA from continuing operations was $20 million or 28% of revenue, up from $2 million or 8% of revenue year-over-year in Q3 2024 and up from $9 million or 15% of revenue in Q2 2025. Turning to Slide 17. Before we begin Q&A, I'd like to reiterate our strong financial position and review our expected capital investment plans for the next 12 months. We are extremely well capitalized to fund our HPC/AI growth initiatives. We have a war chest of over $1 billion, comprised of roughly $820 million in cash and Bitcoin and the remaining $200 million available to draw from our Macquarie facility. With these funds, we expect to be able to fully finance the build-out of our Washington site and the initial phases of construction at our Sharon, Sherbrooke and Panther Creek sites. As we advance our development, the actual investment in our projects will be dependent on a number of factors. We are currently focused on executing on the initial phases of our projects, beginning construction and securing long lead time items to ensure our project timelines. We will continuously evaluate a wide range of financing alternatives at both the corporate and project level, maximizing shareholder value with accretive financing will determine our choices as well as the need for a healthy balance sheet. In closing, I'll underscore that Bitfarms is in the strongest financial position in the company's history, and we have a clear vision of how we are going to best utilize this capital to advance our HPC/AI build-outs in North America. The entire Bitfarms team is incredibly enthusiastic and engaged about the opportunities ahead. With that, I'll now turn the call over to the operator for Q&A.

Operator, Operator

Our first question comes from the line of Mike Colonnese of H.C. Wainwright.

Michael Colonnese, Analyst - H.C. Wainwright

Appreciate all the color on the HPC strategy this morning. First for me, Ben, you mentioned that infrastructure for the Vera Rubin GPU should command a premium to the Blackwell infrastructure. Can you share more on how you guys are thinking about economics there and the CapEx differences?

Ben Gagnon, Chief Executive Officer & Director

Thanks, Mike. Yes, happy to speak to that a little bit. There are two driving forces for our expectations on Vera Rubin economics. The first is that as the dynamic continues to play out where infrastructure is an increasingly greater shortage, that will put upward pressure on economics. The second is that the economics around supply and demand imbalance are really specific to GPU models. If you look at H100s, H200s, the GB200s and GB300s and then what's going to be the next series, the Vera Rubin, there's a lot more infrastructure available to support those older GPUs, which have less specific requirements. With the VR series, energy density is increasing from 190 kilowatts per rack with the GB300s to upwards of 370 kilowatts per rack with the VR200s. Much of the infrastructure being built now will not be compatible with the next generation. As companies allocate capital into those Vera Rubin GPUs, they'll be economically incentivized to deploy them. Regarding margins, they are very high on these GPUs, especially when the GPU is the newest, state-of-the-art hardware as the Vera Rubins will be in 2027. The economic incentive to deploy them faster, even at higher infrastructure expense, should drive higher economics. We don't have a firm price point for exact lease rates yet, but we believe the trend is abundantly clear that the economics next year and in 2027 will continue to improve, especially as the shortfall worsens.

Michael Colonnese, Analyst - H.C. Wainwright

Really helpful color there, Ben. And how should we think about the wind down of your mining operations in the coming years, specifically as it relates to the pace and timing of hash rate coming offline as you start to convert and make further progress in converting your data centers over to HPC/AI?

Ben Gagnon, Chief Executive Officer & Director

Yes, happy to speak to that. First, the LATAM exit we've been working on is a key part of this. We shut down our Argentina facility earlier this year. The Paso Pe facility, which is being held for sale, represents a little under 20% of our hash rate and will impact the company's hash rate going forward. When we look at these transactions, similar to the Argentina shutdown, we expect to pull forward a significant amount of expected free cash flow from those operations today so we can reinvest it into North American HPC and AI infrastructure. So while it will affect free cash flow from mining operations, the impact is mitigated because we are bringing forward one to two years' worth of free cash flow and reinvesting it. There's also a derisking factor as we reduce Bitcoin mining exposure. As we move through 2026, the next sites to come offline will be replaced as we develop HPC and AI infrastructure. Washington would likely transition around the middle of the year and that would be about one exahash, and other sites will come off more gradually as we convert facilities. It will be an orderly transformation and we'll continue to update the market as we announce plans.

Operator, Operator

Our next question comes from the line of Brett Knoblauch of Cantor Fitzgerald.

Brett Knoblauch, Analyst - Cantor Fitzgerald

Thanks for a lot of the color on the different sites throughout the call. I guess when it comes to maybe your PA sites and getting additional power, I feel like that's kind of the biggest catalyst maybe over the near term. I believe Stronghold was kind of in queue before you guys went out and acquired it, which was probably, I don't know, over a year ago now. Do you have any idea on an update of when you expect to maybe expand the power capacity at both Panther Creek and Scrubgrass? Is that a couple of months thing? Within six months? How should we think about the timing there?

Ben Gagnon, Chief Executive Officer & Director

Thanks, Brett. It's a pretty exciting development at Panther Creek because just over the last couple of weeks we've received positive indications on the conversion of the ISA to an ESA as well as the expansion with an additional load study. It's a little too early to say exactly when that would come on site. Our plan includes additional Phase 3 and Phase 4 developments, which would likely come after Phase 2. The conversion of the ISA to an ESA could happen quickly because all of the infrastructure is in place; it primarily depends on regulatory approvals and the necessary paperwork. So it could happen within months. Phase 4 is more likely a 2028 initiative.

Brett Knoblauch, Analyst - Cantor Fitzgerald

Awesome. And then on the GPU cloud as a service, the CapEx figure that you've noted on, I guess, maybe converting that Bitcoin mining to host GPUs, that was not including the GPUs, correct?

Ben Gagnon, Chief Executive Officer & Director

Correct. That figure does not include GPUs or some of the construction costs associated with converting the facility. There will be additional expenses at the Washington site. We've had several conversations with leading GPU manufacturers and we think there are very attractive financing options on the GPUs that would keep CapEx requirements down to mainly infrastructure expense. We could potentially fund up to 100% of compute through GPU manufacturers on attractive terms, which would improve returns. We also think it would provide a significantly greater return profile for GPU as a service or cloud.

Brett Knoblauch, Analyst - Cantor Fitzgerald

And from a capital allocation standpoint, what is your preference? Obviously, the PA sites appear to be leaning more towards colocation, Washington site cloud. Do you expect to grow both businesses at the same time? Is there a preference for one to get online sooner than the other?

Ben Gagnon, Chief Executive Officer & Director

The expectation is that the Washington site will be the first site fully online. Sharon will probably be the second site fully online. Panther Creek is split across phases in 2027 with additional Phases 3 and 4 still to be confirmed. Our priority is managing the critical path and project timelines across facilities. When allocating capital, we manage the critical path and ensure cloud opportunities are affordable. One benefit of doing cloud at Washington is its smaller scale makes it cost effective and something we could consider fully funded today. Larger campuses like Pennsylvania are likely better suited to colocation strategies, which are easier to finance at scale.

Operator, Operator

Our next question comes from the line of Stephen Glagola of JonesTrading.

Stephen Glagola, Analyst - JonesTrading

On the $128 million critical IT supply agreement for Washington, can you clarify the counterparty to that agreement? Is that T5? Or is that another firm? And then additionally, just a follow-up to the last one on the GPU cloud model potentially at Washington and Sherbrooke. Can you maybe elaborate on what factors make GPU as a service compelling relative to standard colocation in these markets? And sort of how are you evaluating both potential GPU risk and your return on invested capital IRR hurdle for the cloud opportunity?

Ben Gagnon, Chief Executive Officer & Director

Thanks, Stephen. The supply agreement for the Washington site is not with T5; it is with a large publicly traded U.S. national company that supplies data center equipment and services. The facility is attractive for both colocation and cloud. The benefits of going up the stack include capturing a greater share of free cash flow and demonstrating operational capability across the stack, which can create long-term value and help when negotiating leases on larger campuses. Our modeling and comparable market transactions indicate that this single site could be worth significantly more in cash flow than our entire Bitcoin mining business historically. That would provide a strong cash flow foundation as mining winds down and help us build operational knowledge useful across other sites.

Operator, Operator

Our next question comes from the line of Mike Grondahl of Northland.

Mike Grondahl, Analyst - Northland

Ben, just curious, what would you describe as the two biggest challenges to maybe meeting your timelines for Washington, Sharon and Panther Creek? Like what's going to be the potential bottlenecks and how are you dealing with them?

Ben Gagnon, Chief Executive Officer & Director

Mike, potential bottlenecks in construction are difficult to forecast precisely because construction conditions change daily. The best way to mitigate risk is having strong partners: owner's reps, general contractors, and an internal team of experienced project managers who monitor critical paths and think ahead on potential issues. While it's not possible to identify one single key bottleneck today, we believe our team and strategic partners put us in a strong position to execute.

Mike Grondahl, Analyst - Northland

Great. And then any rough guidelines or framework you can give us for sort of like 2026 CapEx?

Ben Gagnon, Chief Executive Officer & Director

When looking at 2026 CapEx, we have outlined numbers for Washington but are still finalizing guidance as we revise for Vera Rubin. The real challenge is Vera Rubin infrastructure is new and NVIDIA has not yet completed validated reference designs to support that equipment and infrastructure, so CapEx estimates are adjusting in real time. We should expect better clarity in Q1 as NVIDIA finalizes reference designs. NVIDIA is expected to produce the first Vera Rubin GPUs for their own use probably in Q2 next year, and by Q1 we should have clearer indications of CapEx implications for 2026 and 2027.

Operator, Operator

Our next question comes from the line of Nick Giles of B. Riley.

Nick Giles, Analyst - B. Riley

Appreciate all the detail here. Ben, you mentioned the higher rack density of the Vera Rubin gen and that it could make the rack density suited for Blackwells obsolete. It wasn't that long ago that 100 kilowatts per rack was the high end of rack density. So how are you thinking about future proofing as this trend continues? Are there any contract structures that could protect you from the need to upgrade later down the road?

Ben Gagnon, Chief Executive Officer & Director

Thanks, Nick. The evolution of hardware is rapid. GB200s were about 150 kilowatts per rack, GB300s about 190, and Vera Rubins are expected to exceed 370 kilowatts per rack. That increases cooling and electrical distribution challenges significantly. NVIDIA is looking at higher voltages and even DC systems, such as 800-volt DC. That doesn't mean everyone must convert to DC, but as energy density increases, you need to rethink infrastructure. One way to mitigate obsolescence is to build for the technology coming, not only today's technology, and to lock in multiyear agreements which allow you to recover investments over a longer term. Spreading facilities across a multi-year pipeline also helps, so you're not forced to retrofit an entire portfolio at once. Building to future technology where possible and timing agreements to align with deployment minimizes risk of obsolescence.

Nick Giles, Analyst - B. Riley

I really appreciate that perspective. That takes me to my next question. You mentioned the pipeline. Obviously, you have a lot of growth in front of you, but how much time are you spending on M&A opportunities? And where does that ultimately rank in terms of capital allocation?

Ben Gagnon, Chief Executive Officer & Director

Virtually none, Nick. Our focus as a management team is execution. We don't see tremendous shareholder value in pursuing long lead-time M&A opportunities right now compared to executing on our existing portfolio. We continue to get inbound opportunities, but none compare to what we already have in hand. Expanding the pipeline will be an area of focus in the future, but that's likely a year to 1.5 years out.

Nick Giles, Analyst - B. Riley

Got it. That's good to hear. Maybe one more, if I could, just for Jonathan. Sorry if I missed any commentary around this earlier, but how are you ultimately thinking about the Bitcoin treasury? Would you look to liquidate these holdings around the time that mining operations wind down? Or would those be separate timelines?

Jonathan Mir, Chief Financial Officer

Nice to meet you, Nick. We are not operating as a Bitcoin treasury company and do not want to be one. Programs like Bitcoin 2.1 are designed to offset Bitcoin production costs and achieve higher value per Bitcoin sold as a low-risk, low-cost funding mechanism for energy infrastructure investments. The program primarily sells short and long-dated out-of-the-money calls on Bitcoin, for both the treasury and production. Our efforts are focused on maximizing yield and minimizing costs. We expect the Bitcoin treasury to wind down over time as we allocate proceeds to CapEx.

Operator, Operator

Our next question comes from the line of Martin Toner of ATB Capital Markets.

Martin Toner, Analyst - ATB Capital Markets

Congrats on all this progress, guys. My question is around the GPUs. What's your confidence in being able to acquire them on a timely basis? And would you go through a distributor that comes with the financing or who might finance them?

Ben Gagnon, Chief Executive Officer & Director

Thanks, Martin. We've had extensive conversations with leading GPU manufacturers. NVIDIA produces GPUs themselves and also sells to OEM manufacturers. OEMs often have finance programs, which can be attractive if you have the right infrastructure to ensure GPU quality and lifespan. OEM manufacturers can provide turnkey solutions including servers and may offer financing. For the timeline for Washington, we're highly confident in sourcing GPUs and believe there are financing options that could materially improve return profiles.

Martin Toner, Analyst - ATB Capital Markets

Is there a good exahash number to use for Q4?

Ben Gagnon, Chief Executive Officer & Director

Our exahash should stay relatively consistent in Q4 for our continuing operations. It's not possible right now to forecast the impact or timing of the Paso Pe sale. That site continues to run and generate free cash flow but is classified under discontinuing operations per IFRS. If you look at the hash rate associated with the rest of our portfolio, it will remain relatively constant through Q4, and we'll adjust throughout 2026 as we execute on the HPC and AI development.

Martin Toner, Analyst - ATB Capital Markets

Fantastic. Can you give us a sense for initial conversations with customers of the GPU as a service product, reaction and confidence in being able to contract them on a timely basis?

Ben Gagnon, Chief Executive Officer & Director

Conversations on the GPU front are relatively new for us as we've started evaluating cloud in the last month or two. Inbound demand across Washington and Panther Creek has been significant. Different customers want different structures; some want GPUs included, which can command a premium. It's early to provide concrete economics, but from our modeling and recent transactions in the market, we find the economics compelling—particularly for a smaller site like Washington, which could be fully funded today and serve as a proof point.

Operator, Operator

Our next question comes from the line of Brian Dobson of Clear Street.

Brian Dobson, Analyst - Clear Street

I guess more broadly speaking about Bitcoin mining, as more and more miners transition megawatts to HPC, how do you see the global hash rate evolving over the next few years?

Ben Gagnon, Chief Executive Officer & Director

Interesting question, Brian. I think global hash rate will continue to evolve at the historical rate, but if Bitcoin price does not move up meaningfully, that would be a headwind. Likely you'll see mining rotate to lower-cost jurisdictions. Public miners represented almost a third of the network and many are converting megawatts to HPC and AI, which removes a lot of existing infrastructure from Bitcoin mining. That could be a headwind for exahash growth, but mining will continue in markets like the Middle East, Africa, and Russia where costs are lower. The U.S. is the best market to invest in for HPC and AI, whereas Bitcoin mining is more location agnostic, willing to go to cheaper or higher-risk jurisdictions.

Brian Dobson, Analyst - Clear Street

And then just a quick follow-up. So as you're reviewing your portfolio, do you see an opportunity to engage in this type of megawatt redeployment in a broader sense?

Ben Gagnon, Chief Executive Officer & Director

When looking at redeploying our Bitcoin mining assets, opportunities are limited. We believe the best use of management time and resources is to convert existing free cash flow from mining into HPC and AI investments rather than trying to redeploy mining capacity elsewhere.

Operator, Operator

Our next question comes from the line of Michael Donovan of Compass Point.

Michael Donovan, Analyst - Compass Point

Ben, you mentioned dollar per kilowatt trends. Can you quantify a premium on dollar per kilowatt that you're seeing for power secured in Pennsylvania or Washington versus Texas?

Ben Gagnon, Chief Executive Officer & Director

Good question. Several variables drive dollar per kilowatt leases: timeline, location and development risk. It's hard to pinpoint an exact price by location because these factors vary. Based on transactions and what Bitfarms could secure in Pennsylvania before we're even breaking ground at Panther Creek, we could probably lock in $140 to $150 per kilowatt per month. That rate reflects location and the fact that construction hasn't started. If we could shorten the window between signing a lease and generating revenue—reducing the risk and uncertainty priced into a longer timeline—we could potentially achieve $180 per kilowatt per month. Those are internal estimates and there are many factors in play, but the key point is that a higher dollar per kilowatt rate translates into outsized increases in margin and adjusted EBITDA because OpEx is largely fixed.

Michael Donovan, Analyst - Compass Point

That's helpful, Ben. And you talked about connecting data centers to be one campus, and I was hoping you can unpack this a bit more. How can we think about distance between hauls or pods versus theoretical loss and performance for compute?

Ben Gagnon, Chief Executive Officer & Director

There is a strategy called the regional campus strategy pioneered by Amazon. They determined that roughly 300 miles is a cost-effective range to build direct fiber infrastructure, but the critical factor is latency between sites. For high-end GPU performance, latency constraints matter even within racks and across a facility. In Montreal, our sites are much closer than 300 miles—many are within a 90-minute drive and some within 15–20 minutes. With direct fiber, we can reduce latency below 2 milliseconds at those distances. The cost to connect these nearby sites is relatively modest and provides significant scalability benefits given the difficulty of securing new megawatts in the province.

Operator, Operator

I would now like to turn the conference back to Ben Gagnon for closing remarks. Sir?

Ben Gagnon, Chief Executive Officer & Director

Thank you very much. I would like to thank everyone for attending our earnings call this morning. The management team is very excited. Our long-term investment strategy, we believe, is fully aligned with long-term investors. And we are really excited about the future of this company and what we're building at Bitfarms, and we appreciate your continued support. Thank you.

Operator, Operator

This concludes today's conference call. Thank you for participating. You may now disconnect.