Earnings Call Transcript - IREN Q4 2025

Operator, Operator

Good day, and thank you for standing by. Welcome to the IREN FY 2025 Results Conference Call. Please be advised that today's conference is being recorded. I would now like to hand the conference over to your speaker today, Mike Power, VP, Investor Relations.

Mike Power, VP, Investor Relations

Thank you, operator. Good afternoon, and welcome to IREN's FY 2025 Results Presentation. My name is Mike Power, VP of Investor Relations. And with me on the call today are Daniel Roberts, Co-Founder and Co-CEO; Belinda Nucifora, CFO; Anthony Lewis, Chief Capital Officer; and Kent Draper, Chief Commercial Officer. Before we begin, please note that this call is being webcast live with a presentation. For those that have dialed in via phone, you can elect to ask a question via the moderator after our presentation. Before we begin, I would like to remind you that certain statements that we make during the conference call may constitute forward-looking statements, and IREN cautions listeners that forward-looking information and statements are based on certain assumptions and risk factors that could cause actual results to differ materially from the expectations of the company. Listeners should not place undue reliance on forward-looking information or statements. Please refer to the disclaimer on Slide 2 of the accompanying presentation for more information. Thank you, and I will now turn the call over to Dan Roberts.

Daniel Roberts, Co-Founder and Co-CEO

Thanks, Mike. Good afternoon, everyone, and thank you for joining our FY 2025 earnings call. So today, we will provide an update on our financial results for the fiscal year ended June 30, along with some operational highlights and strategic updates across our business verticals. We'll then end the call with Q&A. So FY '25 was a breakout year for us, both operationally and financially. We delivered record results across the board, including 10x EBITDA growth year-on-year and strong net income, which Belinda will discuss shortly. Operationally, we scaled at an unprecedented pace. We increased our contracted grid-connected power by over 1/3 to nearly 3 gigawatts and more than tripled our operating data center capacity to 810 megawatts, all at a time when power, land, and data center shortages continue to persist across the industry. We expanded our Bitcoin mining capacity 400% to 50 exahash and in the process, cemented our position as the most profitable large-scale public Bitcoin miner. At the same time, we made huge strides in AI, scaling GPU deployments to support a growing roster of customers across both training and inference workloads. We also commenced construction of Horizon 1, our first direct-to-chip liquid cooling AI data center, and Sweetwater, our 2-gigawatt data center hub in West Texas, one of the largest data center developments in the world and a cornerstone of our future growth plans. These achievements underscore the strength of our execution and the earnings potential of our expanding data center and compute platform. We expect this momentum to carry into FY '26 and beyond as we realize the revenue potential of our 50 exahash platform and advance our core AI growth initiatives. So reflecting on current operations. Our AI cloud business is scaling rapidly with more than 10,000 GPUs online or being commissioned in the coming months, backed by multiple tranches of non-dilutive single-digit GPU financing, this rollout will feature next-generation liquid-cooled NVIDIA GB300 NVL72 systems at our Prince George campus. This strengthens our position as a leading AI cloud provider and a newly designated NVIDIA preferred partner. In parallel, while we have paused meaningful mining expansion, our 50 exahash platform continues to generate meaningful cash flow, over $1 billion a year in annualized revenue at the current economics, supporting our continued growth in AI. Together, these operations are approaching annualized revenue of $1.25 billion. That's the scale we're delivering today. However, the clear visibility to continued growth ahead is something we're quite excited about on that note. Our strategy is focused on scaling across the full AI infrastructure stack from grid connected transmission line all the way down to the digital world compute. With a strong track record building power dense data centers and operating GPU workloads, we are uniquely positioned to serve AI customers end-to-end from cloud services to turnkey location, capturing a broad and growing addressable market. Beyond the current expansion to 10,000 GPUs, our 160 megawatts of operating capacity in BC provides a path to deploy more than 60,000 NVIDIA GB300s. With Horizon 1 then offering the potential to scale that further to nearly 80,000. As we continue to assess market demand, this gives line of sight to billions in annualized revenue from our AI cloud business alone. In terms of AI data centers, we're progressing 3 major data center projects to drive this revenue growth as well as provide scope for future expansion. At Prince George in BC, we're continuing to transition existing capacity from Bitcoin to AI workloads with retrofits for air-cooled GPUs and the construction of a newly announced liquid-cooled data center underway to support our GB300 deployment. At Childress, Horizon 1 continues to remain on schedule for Q4 2025. Given strong demand signals, we have also begun site works and long lead procurement for Horizon 2, a second liquid-cooled facility. Together, these projects can support over 38,000 NVIDIA GB300s. At Sweetwater, our flagship 2-gigawatt data center hub in West Texas, Sweetwater 1 remains on track for energization in April 2026. Construction is progressing well with key long-lead equipment, either on-site already or on order. Upgrades to the utility substation have now commenced. So in summary, we've delivered record performance this year. We've got a clear AI growth path with near-term milestones. And most excitingly, we continue to position our platform ahead of the curve to monetize substantial opportunities in the AI infrastructure and compute markets. I'll now hand over to Belinda, who will walk through the FY '25 results in more detail.

Belinda Nucifora, CFO

Thank you, Dan. Good morning to those in Sydney, and good afternoon to those in North America. As noted in our recent disclosures, we've completed our transition to a U.S. domestic issuer status from the first of July of this year. And as such, we've reported our full year results for the period ended 30 June 2025 under U.S. GAAP and the required SEC regulations. For the fourth quarter of FY '25, we delivered record revenue of $187 million being an increase of $42 million from the previous quarter, primarily due to the record mining Bitcoin of $180 million as we operate at 50 exahash. During the quarter, we also delivered AI cloud revenue of $7 million. Our Bitcoin mining business continues to perform strongly, supported by best-in-class fleet efficiency at 15 joules per terahash and low net power costs being $0.035 per kilowatt hour in Q4. Whilst our operating expenses increased to $114 million, primarily due to overheads and depreciation costs associated with our expanded data center platform and increased Bitcoin mining and GPU hardware, we've delivered a strong bottom line of $177 million. High-margin revenues from our Bitcoin mining operations were a key driver of this profitability with an all-in cash cost of $36,000 per Bitcoin mined versus an average realized price of $99,000. Noting that these all-in costs incorporate expenses across our entire business, including the AI verticals, underscoring the strength of our platform. We closed the financial year with approximately $565 million of cash and $2.9 billion in total assets, giving us a strong balance sheet to support the next stage of growth. I'll now hand back to Dan to discuss the exciting growth opportunities that continue for IREN.

Daniel Roberts, Co-Founder and Co-CEO

Thanks, Belinda. So I think it's fair to say that the market backdrop for our AI cloud business is pretty compelling. Industry reports demonstrate accelerating enterprise adoption of AI solutions and services with the percentage of organizations leveraging AI in more than 1 business function growing from 55% to 78% in the last 12 months alone. As almost all of us would know, demand is accelerating faster than supply. New model development, sovereign AI programs and enterprise adoption are driving a step-up in GPU needs and the constraint is infrastructure and compute, not customer interest. Power availability, GPU ready high-density data center capacity remains scarce with customers prioritizing speed to deploy and the ability to scale. IREN is uniquely positioned to meet this demand. Our vertical integration gives us control over the key bottlenecks, significant near-term grid-connected power with data centers engineered for next-generation power dense compute. This enables accelerated delivery timelines and rapid low-risk scaling. Because we own and operate the full end-to-end stack, we are able to deliver superior customer service, tighter control over efficiency, uptime, and service quality translating directly into a better customer experience for our customers. We are leading our service with a bare metal service because it gives sophisticated developers, cloud providers, and hyperscalers what they want most, direct access to compute and the flexibility to bring their own orchestration. As and when customer needs evolve, we have the flexibility to layer in software solutions to provide additional options to the customer. Our new status as an NVIDIA preferred partner is helpful in that regard. It enhances supply access and helps broaden our customer pipeline, supporting expansion across both existing relationships and new end users, platforms, and demand partners. So the market is large, it's accelerating. Supply is constrained, and we have the platform to meet market demand for AI cloud and meet that reasonably quickly. That is why we're immediately scaling to more than 10,000 GPUs, but also now importantly, focusing on what comes next. So 10,000 GPU expansions underway. With it, we will be positioned at the front of the Blackwell demand curve, delivering first-to-market benefits. We saw this with our initial B200 deployment several weeks ago. Upon commissioning, it was immediately contracted on a multiyear basis. Importantly, we are funding growth in a CapEx-efficient way. In the past week alone, we have secured 2 new tranches of financing which have funded 100% of the purchase price of new GPUs at single-digit rates. Anthony will touch on this shortly as well as what's next. In terms of revenue, these GPUs will be delivered and progressively commissioned over the coming months, targeting $200 million to $250 million of annualized revenue by December this year. Approximately 1 exahash of ASICs will be displaced as a result, which we plan to reallocate to sites with available capacity, minimizing any impact to the overall 50 exahash installed hash rate. Finally, we also expect the strong margin profile of our AI cloud business to continue, underpinned by low power costs, but importantly, full ownership of our AI data centers, eliminating any third-party colocation fees from our cost base. Our Prince George Campus will anchor this next phase of our AI cloud growth. So as I alluded to earlier, we are pleased to announce today that construction is well underway on a new 10-megawatt liquid-cooled data center at Prince George, designed to support more than 4,500 Blackwell GB300 GPUs. Following this buildout, half of Prince George's capacity will now be dedicated to AI cloud services. There has been clear runway to double capacity to more than 20,000 GPUs at this site alone. Procurement is also in progress to equip every GPU deployment at Prince George with backup generators and UPS systems. Beyond Prince George, Mackenzie and Canal Flats. Our data center campuses in each of these locations create an even larger opportunity with powered shells, existing and designed to the same architecture as Prince George, these sites offer a straightforward and replicable pathway to more than 60,000 GB300s. Horizon 1 and our broader portfolio of data center sites in Texas opens up a further path to continued AI cloud growth. It's fair to say we're incredibly excited by the AI cloud opportunity. It's a business line that many are simply unable to pursue due to the significant technical expertise and requirements involved with 2- to 3-year payback periods and the low-cost GPU financing structures we are securing, we see this as a highly attractive pathway to continue compounding shareholder value. Our ability to build and operate world-class AI services, all the way down from the transmission line to the compute layer uniquely positions IREN at the forefront of this digital AI transformation. Now on to the major projects driving our AI expansion. Childress continues to show strong on-the-ground momentum with Horizon 1 construction progressing according to schedule and remaining on track for this year. As you can see in the progress photos, the data center buildings are nearing completion and the installation of the liquid cooling plant on the south side of the halls is underway. Based on customer feedback, we have also upgraded certain specifications, including introducing full Tier 3 equivalent redundancy across all critical power and cooling systems. Due to the expected timing gap before NVIDIA's Rubin GPUs are available, we have also reconfigured the design to be able to accommodate a wider range of rack densities while preserving the flexibility to accommodate next-generation systems when they are available. Even with these adjustments, we expect to remain a very competitive build cost target, reflecting the efficiencies of our in-house design, procurement, and construction model. Finally, we are also moving ahead with certain tenant-scope work to derisk delivery timelines and provide additional flexibility, including the potential to monetize the capacity via our own AI cloud service. In that regard, engagement remains active with both hyperscaler and non-hyperscaler customers across both cloud and colocation opportunities. Site visits, diligence, commercial discussions, and documentation ongoing. Building on this strong customer traction at Horizon 1 and general overall market momentum, we are pleased to announce that we've commenced early works and long lead procurement for Horizon 2, a potential second 50-megawatt IT load liquid-cooled facility at Childress. Together, Horizons 1 and 2 will have capacity to support over 38,000 liquid-cooled GB300s creating one of the largest clusters in the U.S. market. In saying that, it's still modest compared to the capacity of our Sweetwater hub, which could support over 600,000 GB300s, which is a good segue, Sweetwater. So both construction and commercial momentum continue to build at the 1.4 gigawatt Sweetwater 1 site, still scheduled for energization in April 2026. As you can see in the progress photo, construction of the high-voltage bulk substation is underway and key long-lead equipment continues to arrive at the site. On the commercial front, we're advancing discussions with prospective customers for different structures. The campus is inherently flexible by design so we can meet demand across the entire AI infrastructure stack, powered shells with partners who want to self-operate, turnkey colocation for customers seeking speed, and cloud services for those who would like us to run it end to end. While we have a multitude of other exciting growth opportunities preceding this, Sweetwater's combination of scale, certainty, and flexibility positions it as yet another growth engine for IREN in the accelerating wave of AI compute. So where do we sit today? Industry estimates call for more than 125 gigawatts of new AI data center capacity over the next 5 years with hyperscale CapEx forecast supporting the credibility of that trajectory. Yet, as most of us know, existing green capacity is well documented and has been far from sufficient to meet this demand. Against that backdrop, we have expanded our secured power capacity more than 100x since IPO. We've built over 810 megawatts of operational next-generation data centers. In the process, demonstrating our ability to not only secure our valuable powered land but also deliver next-generation data centers and compute at scale in some of the most demanding markets. It's a really exciting time for the industry, and it's a really exciting time for us. With that, hopefully providing a reasonably comprehensive overview of the opportunity in front of us, I'll hand over to our newly appointed Chief Capital Officer, Anthony Lewis, to discuss financing.

Anthony Lewis, Chief Capital Officer

Thanks, Dan, and good morning or good evening, everyone. This slide shows how we are funding growth in our AI divisions through strategic financing and strong cash flows from our existing operations. The table to the right, which many of you will recognize, illustrates the cash flows from our current Bitcoin mining activities. With the current network hash rate and a Bitcoin price of $115,000, we project over $1 billion in mining revenue. After deducting all costs and overhead from our business, we estimate close to $650 million of adjusted EBITDA. Additionally, we expect an annual revenue increase of $200 million to $250 million from the expansion of our AI cloud business, which will contribute more over time. While there is some sensitivity to the assumptions, the main takeaway is our expectation of substantial operating cash flow to support our growth initiatives under a variety of operating conditions, supported by low-cost power and top-tier hardware. These cash flows, along with current cash reserves and recent financing efforts, completely cover our near-term capital expenditures, including the cloud expansion initiatives like liquid cooling and power redundancy at Prince George, which will boost our GPUs to 10,900, completing Horizon 1 and energizing the Sweetwater 1 substations. Now, let’s discuss our overall funding strategy. As a rapidly growing capital-intensive business, we are focused on diversifying our capital sources to maintain a strong and efficient balance sheet. The $200 million GPU financings we announced this week exemplify this approach. These transactions financed 100% of the upfront GPU capital expenses, allowing us to accelerate the growth of our AI cloud business at a favorable cost of capital, with payback periods of 2 to 3 years that align well with the quick returns on the hardware. Options like these also provide us with added operational flexibility. We are witnessing strong institutional interest in asset-backed and infrastructure lending within the AI sector, and given our asset portfolio and growth prospects, we believe IREN is positioned to tap into that capital. We are currently exploring various financing opportunities to support further growth, which may include additional asset-backed financing, as well as project-level and corporate-level debt. We have also successfully accessed the convertible bond market with two well-supported transactions this financial year, which presents another funding avenue for us. Of course, we will remain focused on maintaining a prudent equity capital level as we continue to scale, ensuring our balance sheet remains resilient. In summary, with a solid foundation from strong operating cash flows and a diverse array of capital sources at our disposal, we are well-equipped to fund the next stage of our growth. Now, we will open the floor to questions.

Operator, Operator

Our first question comes from Paul Golding with Macquarie.

Paul Golding, Analyst

I wanted to ask on efficiency at these sites. I noticed that PUE in the British Columbia sites is down at 1.1, which is a very impressive efficiency ratio versus Sweetwater being about 1.4. Those may be peak numbers as opposed to average. But I was wondering if you could give some color around how that might influence the thought process around rollout or concentration of sites receiving GPUs initially versus others as you think about the efficiency? And then also just along the lines of this infrastructure being developed with PUE that low being cited, how are you thinking about the backup generation for the existing pods that you have outstanding? I only ask that question in relation to the on-demand versus contracted customer dynamic and how you're seeing that evolve?

Kent Draper, Chief Commercial Officer

Paul, happy to jump in and take that one. So as you mentioned, across the BC sites, we're operating at a PUE of 1.1. That's on an air-cooled basis. Once we install the liquid-cooled facilities there, we expect that to be operating on an average slightly higher than that, but still well under 1.2 PUE across the year. At Childress, the Horizon 1 liquid-cooled installation, the number that you mentioned is much closer to a peak PUE number, although we actually expect it to be less than 1.4 and the average PUE over the year to be around 1.2. In all cases, I think those are extremely competitive numbers across the industry. We are more led in terms of our deployments across the different sites by what our customers are ultimately demanding. Within British Columbia, the ability to scale extremely quickly on an air-cooled basis has been a significant driver of demand for us. And again, that PUE level is extremely competitive regardless. And so that is where we are seeing some of the primary interest from our customer base. At Horizon 1, that liquid-cooled capacity, in particular, is extremely scarce in the industry at the moment and the ability to locate a single cluster of up to 19,000 or just over 19,000 GB300s is significantly attractive and driving high levels of customer interest. So I think less driven by PUE overall in terms of deployments and more driven by the customer side of the equation. To your question on redundancy, as Dan mentioned in his remarks, we're introducing redundancy across the entire fleet of GPUs that we have in our existing operating business as well as for the new GPUs that we purchased. While we believe that for many of the applications that these clusters are used for, it's not necessarily the case that it's required to have redundancy. We have seen some of our customers wanting that redundancy. And for us, we ultimately want to be driven by providing the best customer service, and that's really what's driving us to install that redundancy across the fleet.

Paul Golding, Analyst

And if I could ask one quick follow-up on the GB300 NVL72 capability that has been incorporated or retrofitted into the original plan for Horizon 1, I believe. If you could just give us any incremental color around what that may have entailed and any impact that may have had on how financing availability or future financing plans may be impacted as you think about incremental cost for that density and in particular, maybe as you plan for Rubin given this preferred partner status now?

Kent Draper, Chief Commercial Officer

Yes. So I think what you're referring to with Dan's comments around introducing flexibility for a wider range of densities. And for us, that actually comes more towards lower densities, so being able to operate at densities that are under what the Vera, Rubins would require. So the base design, as we had it, could handle up to 200 kilowatts of rack easily able to accommodate the next iteration of GPUs. But what we're seeing in the market today is that many customers actually want flexibility to be able to operate not only at the rack densities for GB300s, which are around 135 kilowatts a rack, but actually at even lower densities to accommodate additional types of compute within the data center infrastructure. And so what we've done is gone back and reworked some of the electrical and mechanical equipment to be able to actually accommodate lower rack densities. So as it relates to accommodating Rubins in the future, no change from our perspective.

Operator, Operator

Our next question comes from John Todaro with Needham.

John Todaro, Analyst

Congrats on a very strong quarter. First question on the cloud business. And apologies if I missed this, but just the average duration of the contract, kind of trying to determine given the 3-year payback with the GPUs plus infrastructure, the overlap there with the customer contract division? And then I also have a follow-up on the HPC side of things.

Kent Draper, Chief Commercial Officer

Yes. We've got a range of contract lengths across our existing asset base today, all the way from 1-month rolling contracts out to 3-year contracts. For the newer gen equipment, including the Blackwell purchases that we've made, we've typically seen demand in slightly longer contract lengths whilst those Blackwells are new equipment on the market, and so a good indication of that is the initial portion of our B200s that, as Dan mentioned, as soon as they were installed, we were able to contract them on a multiyear basis. So we do have contracts across the spectrum, but we are seeing fewer gen equipment often longer-term contracts being available.

John Todaro, Analyst

Got it. That's great. And then just with the success you're having so far in the cloud business, you could take a step back and think, do we need to sign HPC colo capacity? Would you be more comfortable kind of continuing at this at even a bigger scale? And then as it relates to just kind of thoughts on the CapEx to get you there? Any targeted leverage ratio or a threshold on debt, too.

Kent Draper, Chief Commercial Officer

Yes. We're constantly evaluating the opportunities as it relates to both colocation and cloud. I think we're uniquely positioned in the sense that we are able to take advantage of both opportunities which we think is quite differentiated to a number of others in the industry. They obviously have very different profiles in terms of the risk-adjusted returns. So colocation, longer-dated contracts, typically in the range of 5 to 20 years, but lower payback periods often higher than 7 years before you can get your capital back. And in many cases, because of the nature of the debt financing associated with those, there's very little actual cash flow coming out of the business during that finance period whereas cloud shorter-dated contracts, but much stronger margins and shorter overall payback period. So we typically see around 2-year payback periods on the GPUs alone and 3 to 4 years on the GPUs plus data center infrastructure. So it is something that we're constantly evaluating. And overall, we're looking to maximize risk-adjusted returns across both models. I think you can tell from the comments today, as it stands, we do find the cloud opportunity extremely compelling. Anthony, did you want to touch on the comments around financing?

Anthony Lewis, Chief Capital Officer

Thanks, Kent. Yes, obviously, we have very modest debt servicing requirements today. And I guess as we scale the business, obviously, where those opportunities have developed and the nature of the cash flows and the security of those cash flows will ultimately drive what an appropriate level of leverage is for the business. So the capital structure will continue to evolve as we continue to grow, but we'll obviously be focused on maintaining a strong and resilient balance sheet as well as an efficient cost of capital.

Operator, Operator

Our next question comes from Darren Aftahi with ROTH.

Darren Aftahi, Analyst

Congrats on all the progress. A couple, if I may. So on Horizon 1 and 2. I guess there's commentary in the press release about what theoretically Horizon 1 could support in terms of GPUs, but you kind of left the door open that there may be other uses. So I'm kind of curious on strategic thinking there. And then on Horizon 2, I think, if my math is right, you guys only have 25 megawatts left at Childress, and you're talking about, I guess, 50 megawatts of critical load. Will you be borrowing from your Bitcoin business to kind of get there? And are there expansion opportunities beyond that? Second question, I guess, on Slide 9, you have one of your demand partners is Fluidstack. I'm more curious on the neo cloud side and maybe that entity in particular, given one of your peers signed a deal with them and another partner there, just kind of what the demand drivers are with Fluidstack in particular?

Daniel Roberts, Co-Founder and Co-CEO

Thanks, Darren. Appreciate that. So 3 questions I hear in there. Horizon, we mentioned 19,000. It's just tick over based on the NVL72 configuration, GB300s. The project has been engineered specifically for liquid-cooled GPU. So there is no other use case as an end market other than that. In saying that there's a couple of different ways we might monetize that capacity. A is through different types of GPUs. So as we mentioned during the presentation and Kent reiterated, we've now introduced the flexibility to accommodate a wider range of rack densities. We actually discovered during this construction that the issue is we were building rack densities that are too dense for where the industry is today. So we've had to dial it back a little bit. So accommodating lower rack density gives us the ability to accommodate a wider range of different GPUs whilst preserving the ability to service the Vera Rubins as and when they're released and potentially beyond that. So that's exciting. In terms of monetizing the capacity, there's been colocation versus cloud. So we may buy, own, operate the 19,000 GPUs, and we're having conversations with a variety of potential partners for that, including hyperscale customers. We're progressing financing work streams in parallel. That's a real option. If the risk-return balance is right, as Kent mentioned, then absolutely, we're in a unique position where not many people can build, own, and operate a cloud service. So we're pursuing that, and we're excited about that. But equally, we're seeing a lot of demand for colocation and that would deliver more of an infrastructure return on capital, and we'll remain open to that structure, but we want to see a risk-return framework that is compelling. And to date, I guess we haven't yet seen that. In terms of potentially displacing additional mining capacity. You referenced 25 megawatts potentially being displaced from Childress. Look, that's a cost of business. As we said 7 years ago when we started this business, Bitcoin mining will help bootstrap the business and help us build out the data center capacity. As and when higher, better value use cases come along, then we have the ability and the flexibility to swap those in and monetize our data center capacity differently. In saying that we don't envisage stopping building new data centers. We've got 2 gigawatts at Sweetwater. So it may simply be the case where we just reallocate capacity across different sites. And perhaps there's some relocated action of mining capacity to Sweetwater at some point, but that's something we're working through. Finally, Fluidstack. Yes, we've known the Fluidstack guys for quite some time. We've got a good relationship there. We speak to them. We speak to Google. We know what deals are being done. We look at the deals. As of today, we find a 3-year payback on data center and GPU infrastructure pretty compelling, particularly when Anthony is lining up 100% GPU financing at single-digit interest rates. So look, we'll remain open to colocation opportunities, but the devil is in the detail. And the high level is not always what you end up carrying over a longer period of time, so I might leave that there.

Operator, Operator

Our next question comes from Joseph Vafi with Canaccord Genuity.

Joseph Vafi, Analyst

Congratulations on the impressive progress in fiscal Q4 and so far this quarter. I have one question, which is somewhat two-part. I would like to delve deeper into the financing related to the Blackwells. You mentioned there are some options at the end of the lease financing period, and I was hoping we could explore your thoughts on that. At the end of this lease financing period, what factors might influence your decision on how to proceed? Additionally, it appears that initially, creating your own clusters with this financing seems attractive in terms of payback and time value of money. How much financing do you believe is available in the current market compared to the project financing that you and others have mentioned for more extensive colocation projects?

Anthony Lewis, Chief Capital Officer

Thank you for the question. There are various leasing structures in the market. Some are more traditional full payout finance leases, while others resemble a technology rotation model with fixed lease payments and a fair market value option to acquire the equipment at the end, typically capped at a percentage of the initial price. This structure provides flexibility to return the equipment if we wish to invest in the next generation of GPUs or to continue using the equipment based on the conditions at that time. Could you please remind me of the second part of your question?

Joseph Vafi, Analyst

Just the amount of financing capacity you see out there on the GPU side versus colocation.

Anthony Lewis, Chief Capital Officer

Yes, I think they are obviously quite different asset profiles and the amount of obviously leverage and the cost of that leverage depends greatly on the specific situations. On the cloud side, it's obviously focused on the underlying portfolio of customers, the diversity in the customer mix, the credit quality, the duration of the contracts that will all drive both the sort of pricing and leverage that you can secure. And I guess, similarly on the colocation side, obviously, you can obtain very attractive cost of funds and very meaningful leverage against high-quality offtake such as high-scale offtakes. And as you come down the credit spectrum or the duration of the contract, that will obviously flow through into the cost of the finance and the leverage that you can obtain.

Daniel Roberts, Co-Founder and Co-CEO

And maybe just to add to that, Anthony. Joe, the two are not mutually exclusive, cloud and colocation, in the sense that we are arranging these 100% financing lease structures, as Anthony mentioned, over the GPUs, but that doesn't preclude us then financing the asset base and the infrastructure base at a data center level similar to how you would finance the colocation. It just happens to be the case that the colocation partner is an internalized IREN entity. So that market is open. We're talking to a number of a vast number of potential providers of capital for that. But as Anthony has mentioned, we're looking up and down the entire capital stack to optimize cost of capital at a group level. So you've got these asset-level options, but then you've got corporate options as well. We mentioned the buoyant convertible note market that continues to look quite prospective. We've been prosecuting bond-type structures at a corporate level as well. So there's a whole different array and every week, depending on the level of demand, our revenue profile, how we're building out different elements of the business, the jigsaw puzzle from a financing perspective falls into place and help support that. So it's that reflexive wheel of sources and uses of capital, and that's the benefit of now having Anthony on and dedicated full time to optimizing the cost of capital while Kent runs around North America looking to deploy it.

Operator, Operator

Our next question comes from Reggie Smith with JPMorgan.

Charles Pearce, Analyst

This is Charlie on for Reggie. Can you talk a bit more about some of the key hires you've made in building out the cloud and colocation businesses and where, if anywhere, there is still some room to go? And then as a follow-up, digging in a bit more on the sales side, can you provide a bit more on how you're getting in front of and winning some of the AI clients that you called out in the slides?

Kent Draper, Chief Commercial Officer

Yes. Happy to jump in there on the resourcing question. So we've been hiring across the stack, as Dan made clear, at a level of vertical integration that we have. We continue to need resources across all areas, including data center operations, networking, InfiniBand experts, developers on the software side. We also continue to build out our go-to-market function. So that consists of hiring additional sales executives as well as solutions architects, and we're also expanding the marketing team in parallel with that. So there is an ongoing level of hiring across the business to support the additional customer-facing work that we're doing. And sorry, there was a last part to your question that I missed, it was breaking up a little.

Charles Pearce, Analyst

Yes. Just more on the sales side, like how you're getting in front of the clients, what are you competing on? Why are they choosing IREN, things like that?

Kent Draper, Chief Commercial Officer

Yes. So we get a mix of inbound and outbound customer demand drivers. We have been active recently in the conference space. So we have been getting out telling our story, showing why we are differentiated. As I mentioned, we've been expanding the marketing team and our efforts there to help drive inbound, particularly our activities across all social platforms have been ramping over the past 12 months, in particular. And we're seeing a high degree of interest there. And as that gets out into the public sphere as well as our ongoing provision of cloud services and customer word of mouth, we are starting to see more inbound inquiries as well around both our cloud services platform and the potential colocation platform. So it is a bit of a mix there in terms of what we're seeing.

Daniel Roberts, Co-Founder and Co-CEO

I want to emphasize that the demand-supply situation in this industry is unbalanced, with limited supply. When people need something that is scarce, they tend to seek it out. Word-of-mouth from demand brokers, conferences, and existing customers helps spread the word. We have three distinct competitive advantages compared to others in the neo cloud space like AI scale. Firstly, we manage our infrastructure comprehensively, allowing us to adjust capacity across our existing data centers as well as new ones we are developing. Secondly, vertical integration is crucial, giving us direct control over every layer of our operations. This leads to tighter management of performance, reliability, and service, resulting in increased uptime since we are not reliant on colocation partners or constrained by third-party service level agreements. Finally, from a cost standpoint, we benefit from no colocation fees and enhanced operational efficiency. This positioning also benefits our sales and marketing efforts because we are actively engaged in the industry, have available capacity, and are experiencing significant interest in joining IREN, unlike other providers who lack capacity and have sales representatives with limited opportunities.

Operator, Operator

Our next question comes from Brett Knoblauch with Cantor Fitzgerald.

Brett Knoblauch, Analyst

Maybe on the cloud services front. Is the strategy to go out and order or purchase GPUs with a customer already in mind? Or are you buying those GPUs and then trying to find a customer? And then could you maybe just elaborate on the power dynamics per GPU? I think the 19,000 GB300s for Horizon 1 implies that you can be 380 of them per megawatts of critical IT load. So you have like maybe a similar metric or GB300s or GB200s, if you can provide any color there, that would be helpful as well.

Daniel Roberts, Co-Founder and Co-CEO

So I might take the first half. Kent, if you can do the second half. The prospect of ordering GPUs before or after a contract, this is the nature of the industry where companies want compute, they want it now, but they don't want to wait 2 to 3 months. You think about an enterprise that's made the decision. You think about an AI scale-up or start-up that's raised a bunch of capital. Very few companies are in a position where they can plan out and map out a 2- to 3-year timeline of GPU needs. Often it's we need GPUs, we need them for a project, we need them for today. So the world wants on-demand compute and we almost use this as a universal motherhood statement to guide what we do. The world doesn't really want data center infrastructure. The world at its core wants compute and it wants it now and when it needs it. That's the first element. The second element is I feel like it's groundhog day. We're back in this world, and it takes me back to Bitcoin mining where every man and their dog promises certain amounts of capacity online by a certain date, and no one does it. No one hits the schedules; everyone revises them downwards, stretches them out, and experiences cost blowouts, et cetera because the real world is hard. Dealing with large-scale infrastructure projects, large-scale workforces, complex project delivery, safety. It takes a lot of work and systems and structures to deliver that. This is why we're in such a good position. We never missed a milestone on Bitcoin mining. We are the most profitable if only profitable Bitcoin miner because we did things properly from the start. And we're now sitting here. And as I said, it's groundhog day with the cloud business, where again, all these companies, neo clouds, and otherwise promise capacity online by a certain date, and they rarely hit it. And as a result, customers get a bit gun shy. So the best thing you can do is to continue ordering the hardware. If it snapped up and bit me as soon as it's commissioned, that's a pretty good sign that you're doing the right thing. And as and when we install hardware and the sales cycle starts slowing down, then you know, okay, well, maybe we've just got to slow down on the orders. But each incremental order from here is a relatively small portion of our overall risk so we can afford to take it.

Kent Draper, Chief Commercial Officer

And with respect to the power question, yes, we do continue to see the overall power usage per GPU ticking up with each incremental release from NVIDIA and the other manufacturers. I think using some of the examples of the numbers that were presented earlier in the presentation on an air-cooled basis for B200s, we can fit over 20,000 GPUs into the Prince George site, which is 50 megawatts. At Horizon 1, 50 megawatts of IT load, you're looking at around 19,000 GB300s. So yes, it's not exact math there, but it does give you an idea of what we're seeing in terms of the amount of power per GPU going up over time.

Operator, Operator

Our next question comes from Nick Giles with B. Riley.

Nick Giles, Analyst

I wanted to revisit how the capacity of Horizon 1 will be used as you approach its completion in the fourth quarter. At what point will you decide to fill Horizon 1 with your own GPUs instead of pursuing a colocation deal? Alternatively, if you choose to use your GPUs, should we assume that will apply to the entire capacity, or is it possible that you could mix your GPUs with those from a third party?

Kent Draper, Chief Commercial Officer

Yes. I think that's one of the advantages of where we're at is they're not mutually exclusive options for us. So as we mentioned earlier, we are in a unique position that we can monetize that data center capacity in a number of ways, and it doesn't have to be 1s or 0s. We don't need to do all of it as cloud or all of it as colocation. It could be a combination within Horizon 1. As Dan mentioned, we've started building out Horizon 2. Again, that gives us significant optionality where we could potentially do Horizon 1 under 1 methodology and 1 type of monetization, Horizon 2 under another. But what we will continue to do over time is try and maximize the risk-adjusted returns for how we monetize the assets. And that may fluctuate over time. We're in an obviously incredibly dynamic industry here. And at different points in time, we may see very different risk/reward propositions in colocation versus cloud, but we do have significant flexibility as to how we utilize the capacity.

Nick Giles, Analyst

Thanks for that, Kent. Just on the cloud services, you're focused on bare metal today, but I think you did make some comments that you could expand your software offerings or integrate, if needed. What should we be looking for there? Or what would the incremental revenue opportunities be if you were to integrate?

Kent Draper, Chief Commercial Officer

Yes. Today, as Dan mentioned, the vast majority of the customers that we are dealing with, which make up the majority of the compute market. These are highly experienced AI players, hyperscalers, and developers. They are, for the most part, demanding bare metal because it actually suits them better to be able to bring their own orchestration layer where we see benefits over time from adding incrementally to the software layer is being able to serve a slightly different customer class, which might be smaller AI start-ups or enterprise customers who are looking for a simpler single-click spin up, spin down type service. But today, where we see the demand-supply imbalance, that bare metal offering that we have has a significant level of demand for it. And so we feel like we're well-positioned where we're at today.

Daniel Roberts, Co-Founder and Co-CEO

I think, again, look, just to reiterate, this notion that software is required and these large sophisticated end users of GPUs want a third-party provider to staple their own software and make them use it like these guys are sophisticated. They just want compute. They want to run their own stuff. And at the end of the day, software is eating the world. We know that. Software is not difficult to overlay. The large customers don't want your software. They want their own software. And we are hearing it also firsthand from executives and employees at some of these companies that offer their own software, that it's a nightmare because every time the GPUs change, they need to update the software and rewrite it. And it's this constant evolution of code, bugs, rewriting, updating, et cetera, all for an area of the market that, yes, it might seem good as a narrative, but fundamentally and substantially in terms of revenue opportunity is quite small today.

Operator, Operator

Our next question comes from Stephen Glagola with JonesTrading.

Stephen Glagola, Analyst

As IREN is now recognized as the preferred cloud partner on NVIDIA's website. I was hoping, Dan and Kent, maybe you could provide more detail on your participation in the DGX Cloud Lepton marketplace. And specifically, how do the economics of working through the Lepton marketplace compare to maybe operating your own independent cloud offering? What advantages does IREN get from being on that platform? And any insights into sort of NVIDIA's fee structure or take rate for participants there.

Kent Draper, Chief Commercial Officer

Yes. Happy to give some more color there. So we're not currently participating in the Lepton marketplace. But as an NVIDIA preferred partner, we continue to evaluate platforms like that could expand how we're able to get customers access to our infrastructure. So it may offer us broader reach into developer communities, simple onboarding. So again, to come back to the previous comments that I made on software, it may open up some of the smaller areas of the market, with smaller AI start-ups and the enterprise customers who are looking for a simpler solution. So we continue to monitor this. We are seeing an increasing number of these types of offerings coming to market. And for us, we think it will be an additional demand driver for the underlying compute layer that we are providing.

Stephen Glagola, Analyst

And if I could just ask one more about Horizon 1. Is the growth of your cloud services business influencing which partners you're willing to consider for colocation at Horizon 1, even if they may be competitors?

Kent Draper, Chief Commercial Officer

Yes, we are constantly assessing the mix. I believe you are likely referring to neo cloud customers in relation to colocation. The majority of neo clouds exhibit a different profile compared to hyperscalers in terms of colocation. Within the broader colocation market, there is considerable differentiation. Hyperscalers generally seek longer-term contracts, often ranging from 10 to 20 years, and are highly creditworthy but negotiate hard on financial terms and economic returns. In contrast, neo clouds usually have shorter-term needs, typically between 5 to 15 years, and are less creditworthy than hyperscalers. This factor is something we consider in the risk-reward analysis we spoke about earlier. Regarding the concern that our cloud service impacts our colocation capabilities, I would argue the opposite. Many colocation customers significantly appreciate our expertise in operating these clusters at scale, along with our data center knowledge. We have demonstrated our ability to design data centers for effective cluster operation, and our own cloud service has shown that we can manage them efficiently. Therefore, I don't see any conflict, and it has not posed a significant issue for us historically.

Daniel Roberts, Co-Founder and Co-CEO

Sorry, just to jump in on the Lepton cloud as well. It hasn't really been live functionally. So NVIDIA has been working through a number of items in relation to making that available. I think some of it is now live early access, and we're in direct conversation with them about integration at the moment. So it is a demand partner that we can absolutely envisage using.

Operator, Operator

Our next question comes from Ben Sommers with BTIG.

Ben Sommers, Analyst

So kind of more on the colocation side, just curious what went into the decision to start developing Horizon 2. And if that was a lot of potential customers were thinking about potentially scaling beyond the initial 50 megawatts of Horizon 1. And then I think kind of more broader picture, as we progress towards getting Sweetwater online. What's the different customer profile, if any, for more larger scale sites versus potentially just wanting 50 megawatts or 100 megawatts and just kind of any color on the counterparties that you're having conversations with.

Daniel Roberts, Co-Founder and Co-CEO

So we haven't committed the full CapEx to building out Horizon 2. So importantly, over the last 7 years, our whole business model has been around cheap optionality. And sitting here right now today looking at the bigger picture, and I can drill into that, it just makes sense to order long lead items and start moving the ball ahead on a potential commissioning of the Horizon 2 facility. So a lot of the way the S curve works for CapEx in respect to these facilities is you've got long time and small cash outlays that build up over time before the larger CapEx commitments come in. So it makes sense to put down deposits on long lead items, get the ball rolling, so that we can maintain a really competitive fast time to power for Horizon 2. Now sitting here today relative to 3, 6 months ago, we're seeing further validation of a decision to spec a relatively small amount of capital. We are seeing demand take up for AI cloud. We're seeing the number of inbounds for colocation. We're seeing better visibility on the overall demand supply imbalance for liquid-cooled chips. So it's a bit of a no-brainer to be honest. And in terms of committing full CapEx to that, we've got time, and we'll just continue to monitor the market live because things are changing week-to-week in this industry. And that flexibility, having a governance structure that is founder-led, the ability to make quick decisions, work with the Board and adapt to where the market is going is really important because it is super dynamic.

Operator, Operator

Thank you. I would now like to turn the call back over to Daniel Roberts for any closing remarks.

Daniel Roberts, Co-Founder and Co-CEO

Thank you very much. Thanks, everyone, for dialing in. It's obviously been an exciting quarter and exciting year. We're thrilled about expanding to 10,900 GPUs in the coming months and really putting our AI cloud service further on the map. But for us, most of our time is now focused on what lies beyond that. So expanding our 3 gigawatt power portfolio, we're working hard on. That's exciting. That's many years away, but it was many years away the 3 gigawatts where we started 7 years ago. So continuing to position ourselves ahead of the curve in every respect is just critical. And it's really important when you're fighting this real-world digital world imbalance where digital demand increases overnight, it goes exponential. Your ability to service that demand with real-world infrastructure and compute works in a linear fashion, it's harder, it takes longer. So the ability to preempt those digital demands and build for tomorrow, position for tomorrow rather than where we are today is a key competitive advantage and something we will maintain. And it manifests itself in us building 200-kilowatt racks, but the industry can't support 200-kilowatt racks. So Horizon 1, we're having to reconfigure to make it smaller. So we'll continue to keep that in mind. We're excited about the future. We appreciate all of your support and can't wait for the next quarterly earnings. Thanks, everyone.

Operator, Operator

Thank you. This concludes the conference. Thank you for your participation. You may now disconnect.