IREN (IREN.US) Q2 conference call: The structural gap between supply and demand continues to expand, and vertical integration of full-stack capabilities has become a moat
The Zhitong Finance App learned that data center and AI cloud service provider IREN Limited (IREN.US) released the fourth quarter results report for the 2026 fiscal year after the market on August 27. IREN revealed during the fourth quarter results conference call of FY2026 that the company's AI cloud transformation has entered the harvest period: $4 billion in annual recurring revenue (ARR) has been signed, of which $1 billion has already been put into operation, and the Microsoft Horizon 1 project has been delivered as scheduled. Management emphasized that the structural gap between digital demand and physical supply continues to widen, and IREN has built scarce full-stack capabilities in the industry by vertically integrating electricity, data centers, GPU computing power and software services. GPU financing of 6.5 billion US dollars has been completed in the past 3 months. The customer's advance payment has covered about 50% of the GPU capital expenditure, and the capital flywheel effect is beginning to show. The management made it clear that production capacity was basically sold out in 2026, and they are actively discussing production capacity from 2027 to 2028. Data center financing will be a key breakthrough in the next phase of capital efficiency improvement.
IREN Co-CEO Dan Roberts pointed out at the performance conference that the explosive growth in AI inference demand is being structurally disconnected from the long-term construction of physical infrastructure — the average weekly token usage of the eight major OpenRouter models has increased nearly 17 times in 8 months, while the data center takes more than 3 years from planning to operation. In this context, IREN's vertical integration model of “land - electricity - data center - GPU computing power - software management” has become a scarce ability to meet the immediate needs of customers. Currently, the planned production capacity for 2026 has been basically sold out, and the $4 billion ARR covers diverse customer groups such as Microsoft, Cohere, Perplexity, Figure AI, and an unadvanced AI laboratory. Management specifically emphasized that existing customer renewals and expansions (such as Together AI and Fireworks AI) have fully verified the platform's delivery capabilities, and AI applications are extending from chat to robotics and physical world automation, and this trend will continue to drive demand for computing power.
Chief Commercial Officer Kent Draper revealed that the price of the three-year contract has increased by about 125% since November last year. The recent contract price exceeds $20 million per megawatt IT load, and the price being negotiated is about $25 million. Management attributed this to continued tightening market demand and IREN's added value at the managed services and software layer (Mirantis). In terms of financing, CFO Anthony Lewis said that it has received a total of about US$19 billion in capital over the past 12 months, including US$6.5 billion in GPU financing in the past 3 months — including US$3.6 billion in investment-grade financing (interest rate of about 6%) and US$2.8 billion in non-investment-grade financing (fixed interest rate of 9%). Customer prepayments have reached 45% to 55% of GPU capital expenditure in recent transactions, compounded by a 90% financing leverage ratio, forming a “capital flywheel”. Not only does it fully cover GPU procurement costs, but the excess amount can also be used for data center capital expenses. Dan Roberts compared this process to the process of forming a real estate finance market and emphasized that “demand is real, and financing must follow.”
Looking ahead to fiscal year 2027, the company expects capital expenditure of approximately US$25 billion to US$30 billion, covering Microsoft production capacity delivery, ARR related deployment in 2026, and liquid-cooled data center construction in 2027. Management pointed out that at present, all data center assets (including Horizon 1-4) are not burdened with property rights, reserving huge space for subsequent asset financing. Dan Roberts made it clear that the efficiency of GPU financing has been verified, and that data center financing will be the next stage of “big unlock” — the project is closer to traditional real estate financing logic, the market capital is abundant, and IREN plans to take the opportunity to launch the project after stable operation. He concluded, “We have electricity, land, data centers, computing power and software, and the talent to do it all. For Will and I, this is just the beginning.” Overall, IREN is moving from a production capacity construction period to a large-scale monetization period, and improving capital efficiency will be a key variable in determining whether it can continue to lead the AI infrastructure competition.
The following is a transcript of IREN's FY2026 fourth quarter results conference call:
Executive speeches
Mike Power
Vice President of Investor Relations
Good afternoon everyone, and welcome to IREN's 2026 results presentation. I'm Mike Power, VP of Investor Relations. Joining me on the conference call today are Daniel Roberts, Co-Founder and Co-CEO; Anthony Lewis, Chief Financial Officer; and Kent Draper, Chief Commercial Officer. Before we begin, please note that this session is being webcast and is accompanied by presentation slides. Participants dialing in by phone can ask questions through the moderator after our pre-prepared presentations.
I would like to remind you that some of the statements in this meeting may constitute forward-looking statements. These statements are based on current expectations and assumptions, and are subject to risks and uncertainties that may cause actual results to differ materially from the statements. For more relevant information, please refer to page 2 of the attached presentation and the documents we have submitted to the US Securities and Exchange. With that in mind, I'll leave the meeting to Dan Roberts.
Daniel Roberts
Co-founder, Co-CEO and Executive Director
Thank you Mike, and thank you all for participating. Will and I started this company based on a very simple observation: the digital world expands almost instantaneously, but the physical world doesn't. Electricity, land, data centers — these things take years to license, finance, and build. And this year, this is no longer a theoretical assumption, but a decisive bottleneck for the entire industry.
Take a look at the chart on the screen. According to data from 8 models tracked by OpenRouter, weekly token usage for large language models increased nearly 17 times in 8 months. Every token runs on a physical infrastructure. Faced with this demand curve, it is very difficult to meet it with infrastructure with a 3-year delivery cycle. And it's just the beginning of the cycle, not the end. Cheaper computing power won't make existing things more efficient; it will make new things economically viable. Infrastructure spawns applications, and applications create demand for more infrastructure. This has been the case with every wave of construction in history, and this is a structural disconnect, and this gap continues to widen.
Next, I'll introduce our layout in this context. We operate three levels from the bottom up: the first is data centers, land, electricity, substations, cooling systems — this is arguably the hardest level to build, and this is where the shortage began. Second is computing power, that is, GPUs, storage, and network devices within the data center. Finally, there's the upper layer of software, managed services and enterprise support — that's where Mirantis is in our system. Just today, Mirantis was granted the first batch of certified hypervisor (Certified Hypervisor) qualifications by Nvidia. So far, we have obtained official verification from Nvidia on both the software layer and the hardware layer. Why have all three levels? Because each layer makes the layer below more valuable. Power grid access points are more valuable because they have data centers, and the value is further increased after having GPUs. Coupled with the services surrounding customers, the value is even higher. Most players in this market rent at least one of these tiers, and we own the entire technology stack.
Let's take a look at what we've achieved this year. Please switch to the highlights slide, thank you. There are four main points in this update, and I'll go through them quickly, as more on that later.
First, the customer side. We have signed a new multi-year cloud contract, and our customers include Cohere, Prometheus, Perplexity, Figure AI, Fal AI, Higgsfield AI, and a leading cutting-edge AI lab whose name cannot be disclosed at this time. Second, in terms of revenue. The annual recurring revenue (ARR) corresponding to the 2026 production capacity that has been contracted so far has reached US$4 billion, of which US$1 billion has already been put into operation. That doesn't include revenue that is expected to be gradually realized in 2027, such as the $700 million ARR associated with Nvidia's cloud contract. Third, the delivery aspect. Horizon 1 was delivered to Microsoft this month, the first of four 50-megawatt deployments. Horizon 2 to 4 is targeted for delivery in the December quarter. Fourth, the financing aspect. GPU financing of 6.5 billion US dollars has been completed in the past 3 months, and with customer advance payments, it has covered more than 100% of the corresponding GPU capital expenditure. Of this, $2.8 billion of financing did not require an investment-grade offtake agreement, and interest rates remained in single digits.
I'll start with the customer, because that's where everything else comes from. We have almost sold out of production capacity in 2026, so the questions we are receiving now all revolve around 2027 and 2028. We are strategically continuing to sign contracts for future production capacity. We evaluate every contract opportunity from three aspects: first, who the counterparty is and what it can bring to the platform — that is, strategic value, not just revenue; second, what are the economic terms — price, advance payment, term, etc.; and third, what kind of space this can open up for managed services and software businesses in the long run.
We've mentioned many times before that in this market, signing an order is not a bottleneck; it's just putting the GPU online. And now we no longer need an investment-grade offtake agreement to finance GPUs. So we're not rushing to sign contracts to chase headlines; we're making long-term decisions about where the company will position itself in the future. We will keep everyone informed when we sign the contract.
We are currently in post-negotiations with a number of new customers on a significant portion of our 2027 production capacity, and negotiations for 2028 are already in full swing, including both customer and financing aspects. In the longer term, we want to continue to build a deeply diversified customer base within the AI ecosystem, and all customers have signed master service agreements with IREN. Every megawatt we build is a scarce resource. Before each new set of clusters goes live, we want it to meet the needs of both existing and new customers — so it doesn't depend on any single customer, any single negotiation, or any single price cycle node.
Below is a list of customers currently on the platform. As I mentioned before, the focus is on a new multi-year contract with a leading cutting-edge AI lab. To be clear, this is a new contract and separate from Prometheus — Prometheus is the unnamed AI developer in our July announcement. The world's smartest AI infrastructure buyers continue to choose us, which is proof that our strategy is working. When we started, our customers were concentrated because big customers had immediate needs and were able to scale up as quickly as possible. As the platform grew, we have consciously broadened our customer base to include hyperscale enterprises, enterprise customers, AI developers, and now also cutting-edge labs involving both training and inference. But honestly, what we value most is the third point: existing customers keep coming back. Together AI and Fireworks AI have both been renewed and expanded. New customers have proven that demand exists, and contract renewals continue to prove the reliability of our delivery and execution. Once we've deployed our customers, we'll grow with them to cover different sites, different GPU generations, and different service levels.
About our customers who signed up and grew with us: Prometheus and Figure are building products for the physical world — robots, real-world automation — and they're contracting our computing power to achieve this goal. AI has long gone beyond chat apps. This is exactly what we have repeatedly reminded you — you can't use today's usage to estimate the size of this market, and this is the root cause of the market always structurally lacking in computing power.
Next, let's look at pricing. Pricing has changed a lot. The three-year contract price has increased by about 125% since November last year, and the five-year contract price has increased by about 70%. The recently signed three-year contract is priced at over $20 million per megawatt IT load, and the payback period of computing power investment is about 2 years, while the price currently being negotiated has reached approximately $25 million per megawatt. Recent customer upfront payments have covered 45 to 55 percent of GPU capital expenditure. What's the reason behind it? Market demand is certainly tightening, and it is also related to the types of customers we sign up for, transaction structures, and additional services in managed services — Mirantis is constantly expanding these services. Shorter terms and on-demand pricing will provide room for further upside in the future. Incidentally, revenue per megawatt is just a simplified metric; we're looking at the overall return: price, term, upfront payment, utilization, customer quality, and expansion opportunities. To be clear, we won't be hoarding production capacity to wait for the spot price. We allocate production capacity according to our own judgment to build the customer base and platform, and pricing will follow.
Moving from customers to what we're building. Target: Delivery of approximately 300 megawatts of IT load in 2026, and another 0.5 gigawatts in 2027. This will bring the platform's gross capacity to approximately 1.25 gigawatts (1.2 gigawatts) in 2027. We are continuing construction in Texas, British Columbia, Oklahoma, South Australia, and Spain. Currently, more than 4,000 construction workers are working at our various active sites. Horizon 1 is the best example of this.
As we previously announced, Horizon 1 has been delivered to Microsoft, the first of four 50 MW liquid cooling deployments at Childress, and has received the Nvidia GB300 NVL72 “Exemplar Cloud” certification — which is important because it demonstrates our ability to integrate and operate the entire platform — hardware, network, and software — rather than just building a shell. The delivery itself is important, but the template is even more important. Every Horizon project uses the same design, the same supply chain, and the same field team, and each phase draws on the experience of the previous phase. Horizon 2 is undergoing commissioning, Horizon 3 and 4 are in the late stages of construction, and all three projects are targeted to be delivered within the December quarter. This model is being implemented at every site.
Regarding progress in 2026: At Childress, the air-cooled computer room renovation work and GPU installation are progressing in parallel. At Mackenzie, GPUs from the first two buildings are already on the shelves. At Prince George, the air cooling fleet has been fully commissioned, and the liquid cooling installation is in progress. All remaining production capacity is targeted to be completed within the December quarter. Then 2027: Sweetwater 1 is now in full swing, the first building is under construction, and the main substation is also being promoted. In Childress, civil works for Horizon 5 and 6 have begun, and a 250 MW air-cooled transformation project is also underway. Additionally, it is worth mentioning that with regard to Canal Flats, we have now decided to convert all of them to a GB300 liquid cooling solution, which will release more power value from the sites we already own.
In addition to this, the project pipeline will be further strengthened. 2028: Sweetwater 2 and Kiowa in the US, Bundey in South Australia, and Badajoz in Spain, total around 300 MW. The latter is the flagship project in the Nostrum asset portfolio completed this quarter. Overall, the announced installed capacity exceeds 5 gigawatts, and is accompanied by several gigawatts of development reserve pipelines, including more Texas power projects — projects with real capital strength will have an advantage under more stringent grid-connection processes.
Let's talk a little bit about design because it answers a question we're beginning to get frequently. One pessimistic view of AI data centers in the market is that these facilities need to be rebuilt every 10 years. And this year we've been making sure that's not the case at our facilities. Sweetwater 1 is becoming the reference design for Sweetwater 2, Kiowa, Bundey, and subsequent projects — universal layout, universal equipment, higher degree of modularity and prefabrication. Each project inherits the experience of the previous project rather than starting from scratch. The design is compatible with subsequent GPU generations and evolving cooling technology, including 800 volt DC power distribution, all of which were developed in collaboration with Nvidia. We know computing power is changing faster than buildings, and we're designed to adapt to that change.
At the same time, we are also extracting more value from our existing assets. Existing sites have surplus electricity beyond existing deployment plans. This space can support more GPUs and revenue without additional grid capacity — and additional grid capacity is the scarcest investment in the entire industry. As a result, revenue without additional grid capacity is the highest quality of growth. First, liquid cooling installations were added in Mackenzie, Canal Flats, and Prince George in 2027. In the long run, tools that smooth out GPU power consumption, such as Nvidia MaxLPS, allow us to safely run more computing power at the same power capacity.
But all of this is inseparable from people. Our workforce almost tripled in FY2026, including hundreds of colleagues who joined through Mirantis and Nostrum, and we expect to maintain a similar growth rate in FY2027. The company added five new executives, responsible for development, product, marketing, innovation, and information security, from leading data center operators such as Nvidia, AWS, Oracle, and Google. Building the organizational structure ahead of time is clearly intentional — it requires initial investment before generating revenue, but this is the direction we are moving forward.
There's one last point before we hand it over to Anthony, which is how to finance growth. The first is GPU financing, as the model has now been validated on both ends of the credit spectrum. For the Microsoft contract, we raised $3.6 billion in investment-grade GPU financing at a weighted average interest rate of around 6%. Coupled with the customer's advance payment, it covered approximately 96% of the corresponding GPU capital expenditure. What's really interesting is the non-investment level deployment aspect — that is, the part that targets the broader AI ecosystem. We've just completed $2.8 billion in equipment financing, including $2.4 billion in fixed-interest 9% financing led by Blue OWL and Pimco's funds for the Mackenzie project. The financing included a delayed withdrawal term loan and a preferential guarantee note, which together covered 90% of the GPU's capital expenditure. Coupled with the 45% to 55% upfront payment in recent transactions, the total funding far exceeds the acquisition cost of the underlying GPUs — which is a good thing, as the excess amount is being used to support these sites and data center capital expenses for future deployments.
So far, with regard to data center capital expenditure, we have intentionally kept 100% of our data center assets free of ownership. This is a growing asset pool, and we can finance at the right time. Next, Anthony will explain how this portion of funding supports plans and financial results. Thanks Anthony.
Anthony Lewis
chief financial officer
Thanks Dan and good evening everyone. Over the past 12 months, we have received approximately $19 billion in funding, of which nearly $16 billion comes from customer prepayments, GPU financing, and convertible notes, and about $3 billion from equity. The vast majority of funds are either already in cash or haven't been withdrawn, which provides us with plenty of liquidity. For fiscal year 2027, we expect capital expenditure of approximately $25 billion to $30 billion. The estimate includes contracted Microsoft capacity deliveries, additional deployments required to achieve the 2026 ARR, and GPU and data center capital expenses corresponding to the air-cooled deployment scheduled to be completed in the 2027 calendar year. It also includes investment in the production capacity of Childress and Sweetwater 1's new liquid-cooled data center (targeted for delivery in the second half of 2027), as well as early investments for 2028 and beyond.
Of course, actual capital expenditure will depend on various factors, including final costing, construction schedule, delivery time for GPUs and long-term equipment, overall contract environment, and financing environment. In this capital expenditure estimate, we expect capital expenditure requirements for existing and newly deployed data centers and GPUs to increase by approximately 15% to 20%, and revenue growth is expected to eventually exceed this increase.
Regarding the financing plan, as mentioned before, we had a strong starting position. Current cash and promised GPU financing and prepayments total around $14 billion. This includes $7.6 billion in cash on the balance sheet as of June 30, of which $1.7 billion is restricted capital, and the vast majority of restricted funds have been set aside for Microsoft's GPU capital expenses. We aim to secure an additional $8 billion in GPU financing and prepayments to support GPU capital expenditure requirements — taking into account the healthy level of advance payments in recent contract negotiations and the growing GPU financing market mentioned by Dan. The remaining funding gap is expected to be met through data center financing, operating cash flow, and other funding sources at the company level.
Regarding data center financing, as Dan mentioned, currently all of our data center assets are free of title liability, including Horizon 1 to 4. Thus, it is clear that there is a large and growing pool of assets that can be used for asset financing. Coupled with an increasingly active GPU financing market and attractive customer upfront payments, we believe we are well-positioned to expand the scale of financing to support the rapid growth of the platform.
Now let's take a look at the fourth quarter results and outlook. This quarter's results continue to reflect the transformation to the AI cloud. Revenue for the June quarter was $137.2 million, of which AI cloud revenue was $70.5 million. This figure is down $7.6 million from the previous quarter due to the fact that we retired some of our miners in preparation for GPU installations, an impact partially offset by the growth of AI clouds. Revenue costs also fell by $6.6 million, mainly due to lower electricity consumption due to reduced mining activity. The net loss for the quarter was $684 million, mainly driven by non-cash impairment of $450.4 million (mostly due to decommissioning of mining rigs) and a $92.1 million reduction in fair value of mining rigs for sale, both of which reflect the cost of transforming the existing site into an AI cloud. We expect the mining business to be largely decommissioned by the end of December 2026.
We expect cash sales and management expenses to increase by approximately $40 million to $50 million month-on-month in the first quarter as we continue to invest in sales and marketing, R&D, development, site and cloud operations, and other functional areas to meet significant revenue growth over the next few quarters.
ARR is next. Our ARR at the end of the fourth quarter was around $500 million. After Microsoft accepted Horizon 1, ARR has now reached 1 billion US dollars, and this level will continue throughout the September quarter. We expect ARR to exceed $4 billion by the end of the December quarter, and related contracts have been signed, including the delivery of Horizon 2 to 4. Since most of the production capacity for the December quarter is expected to go online at the end of the quarter, the revenue reflection will mainly be reflected in the March quarter. Now I'm going to give it back to Dan.
Daniel Roberts
Co-founder, Co-CEO and Executive Director
Thanks Anthony. Finally, there's a slide show. Production capacity was almost sold out in 2026, and an ARR of 4 billion US dollars has now been signed. 2027 and beyond is the runway we have deliberately planned. We are continuing to transform this part of our production capacity into a broader customer base and a richer service portfolio. Revenue per megawatt is rising, both from structural demand and from site optimization. And this $4 billion in revenue comes only from less than 10% of our already locked 5 gigawatt or more grid access portfolio.
Eight years ago, we set out to bridge the gap between digital demand and physical supply. Looking at it now, this gap does exist. We have electricity, land, data centers, computing power and software, and the talent to service it all. For Will and I, this is just the beginning. Now ask the operator to start the questioning session.
Q&A session
Michael Ng (Goldman Sachs):
I have two questions. First, could you please talk more about Mirantis and how it can help you commercialize your computing power? Has Mirantis been used to facilitate deals that have been signed or are under negotiation? Do you think Mirantis will eventually be used to sell part of its production capacity to market aggregators like OpenRouter?
Kent Draper:
Ok, happy to answer that question. Regarding Mirantis's service capabilities, I think there are a few aspects that are helpful to our business. First, as you said, it actually opened up a new category of customers for us. As we have mentioned many times before, hyperscale customers or cutting-edge AI laboratories usually prefer to obtain computing power in the form of bare metal, but smaller AI developers, laboratories, and enterprise customers value arrangement layer services provided on top of bare metal computing power. Therefore, the arrangement layer service provided by Mirantis can help us open up this segment of the market. Additionally, it allows us to provide computing power in different ways — not only by providing clusters of reserved managed services, but also by opening up the possibility of providing on-demand computing power. In addition to the arrangement layer itself, Mirantis has long experience in serving enterprise customers in the field of cloud computing, bringing professional capabilities in enterprise support, monitoring, and deployment, all of which help us to monetize the platform faster and continue to provide a high level of customer service. These are just a few of the ways Mirantis adds value to existing platforms.
Daniel Roberts:
Let me add one more point. Thanks to today's announcement, we are now the only “new cloud” service provider to receive Nvidia's certified virtual machine monitoring program. The advantages of this vertical integration and full technology stack are being reflected in real time.
Michael Ng (Goldman Sachs):
I would also like to ask about next year's capital expenditure forecast. Will the 25 billion to 30 billion US dollars all be used to support the 800 megawatt production capacity expected to be signed next year? Or is the longer term part included as well? Also, beyond the $14 billion in cash, GPU prepayments, and other debt financing you mentioned, can you talk about the rest of the financing plan?
Anthony Lewis:
OK, thanks for asking. Regarding the first question. The $25 billion to $30 billion covers the fiscal year ending June 2027, and therefore includes all capital expenses required for 2026 deployment to achieve the $4 billion ARR target, data center and GPU capital expenses for the remaining portion of the air-cooled deployment expected to be completed in the 2027 calendar year, and a significant portion of the capital expenditure required for liquid cooling deployment in the second half of 2027. However, this figure does not include capital expenses for GPU computing power in these new liquid cooling facilities, which will be part of the capital plan for the next fiscal year. Regarding financing, as we mentioned that the $14 billion has already been implemented, I also mentioned the estimated additional $8 billion in GPU-related capital expenses, which will support GPU deployments included between $25 billion and $30 billion. Obviously, we're seeing a very strong financing and prepayment environment, which supports the $8 billion figure. GPU financing market conditions are also strong. Dan mentioned the results of recent transactions, and financing deals for other investment-grade and non-investment-grade clients have also appeared in the market. Furthermore, Nvidia recently announced the establishment of joint ventures and partnerships with the world's six largest capital allocation agencies to further support the financing of AI computing power as an asset class. Additionally, we clearly have Horizon 1 to 4 already basically completed and delivered within the year. None of these are burdened with property rights, and we are also evaluating other asset-backed financing and group-level financing opportunities.
Paul Golding (Macquarie):
Congratulations on making great progress. First, I would like to ask about the decision on liquid cooling transformation in British Columbia. Can you elaborate on the decision-making process? It sounds like there is surplus electricity there, and there are also income opportunities. I'd like to know how you assess relative returns — based on existing demand or a speculative layout? There's one more follow-up question.
Kent Draper:
Mostly driven by the demand we see. We have an efficient liquid cooling design that is ideal for British Columbia conditions. As customers increasingly gravitate towards future workloads, the vast majority of customers are gradually switching to liquid-cooled GPUs, so we're reflecting this trend in our portfolio.
Paul Golding (Macquarie):
OK. Also, Anthony mentioned that the cost of computing power is rising, but the increase in revenue will outweigh the increase in costs. Considering the continued rise in computing power costs and the excess liquidity you get from prepayments and financing, how are you balancing allocating some of the funds to data center construction and longer-cycle computing power purchases? Especially considering the trend in the price of computing power, will speculative purchases be carried out?
Kent Draper:
In short, we're doing both. As Dan mentioned, to sell computing power, you must have data center capacity, and every layer of the entire technology stack is essential. We continue to see extremely strong demand in the industry and a shortage of available capacity, so we will continue to build data center space. We will also continue to procure computing power and continue to superimpose software layers to expand our market reach and obtain better economic returns. As a result, we expect some speculative purchases and some purchases linked to customer contracts. However, as Dan said, contracts are not the driving factor; the key is to bring computing power online, which is what we pay great attention to at the business execution level.
Daniel Roberts:
Paul, with regard to your capital expenditure and revenue issues, let me add that this is also an extension of what Anthony mentioned earlier. Looking back, the contract we are announcing now and the price per megawatt. The corresponding GPU was ordered quite a while ago. As a result, these prices reflect natural price increases in addition to those GPU capital expenditure inflation. GPUs and costs are indeed rising over the next few generations, but the revenue per megawatt we announced today isn't necessarily responsible for those cost increases because they correspond to older generations — you can understand that. Furthermore, our capital expenditure requirements are more nuanced. Yes, we have given total capital expenditure guidance of $25 billion to $30 billion for fiscal year 2027, but we intentionally don't split it up or attribute it to specific megawatts of IT load, because the reality is that data center capital expenditure is often one or two years ahead of delivery — this is determined by the nature of substations and steel procurement. Thus, the composition of capital expenditure in any year actually depends on what is delivered in that year and what is built for the next year. We're looking at how to mismatch data center capital expenditure and how to finance it. As we mentioned in our demo, today 100% of our data center assets are free of title burden, which creates future opportunities. When we receive an advance payment equivalent to 50% of the GPU cost, this is roughly equivalent to 100% of the cost of the data center, because the GPU accounts for about two-thirds of the total cost of the data center plus GPU when fully loaded. If you have already funded the GPU with 90% leverage, then you can see the financial flywheel effect being formed, which can be supported by almost no equity investment in the long run. But we need to add one more factor: our ambitions are growing, the pace of construction is faster than in the past, so the demand for financing continues to increase. However, we are flexible in terms of growth and can be optimized according to the financing terms available at any point in time.
Paul Golding (Macquarie):
Very valuable addition, Dan. If allowed, ask one more brief question. Regarding the ratio of two-thirds computing power to one-third of data center capital expenditure, is this ratio changing as computing power costs rise? Or did the tight labor market and other input costs in data center construction keep that ratio relatively stable?
Daniel Roberts:
No, that ratio has remained fairly stable. We do see that component cost inflation continues to rise, but from a substantive perspective, I think that ratio will remain relatively consistent.
Brett Knoblauch (Cantor Fitzgerald):
Congratulations for this quarter. It's great to see demand trends in 2026 and 2027 and 2028. I'm curious about Horizon 1's construction. Recently we've seen that many deals may have lower PUE (Power Usage Efficiency). Are you adopting design changes to achieve lower PUE in other sites or buildings that will go live in 2027 or 2028? Or is the PUE level of the deal with Microsoft largely unchanged?
Kent Draper:
PUE is never completely static, but it mainly depends on the surrounding conditions of the environment in which the data center is located. As we mentioned before, we use an efficient closed loop liquid cooling system, so we have a very effective PUE compared to the industry as a whole. We do see continued marginal improvements, thanks in part to the Nvidia DSX reference architecture, etc., where they looked at how the chip can operate at higher temperatures without loss of computational performance, thereby allocating more total megawatts to IT load rather than cooling. But these are marginal improvements rather than major drivers over time.
Daniel Roberts:
Let me add that we have always intended to be conservative about publicly advertised PUE. We dedicated a slide in our presentation to discuss this topic—that is, we have surplus electricity in addition to our current deployment plans, which is partly related to optimizing PUE over time. In fact, the average PUE for the whole year is far below the required peak PUE. As a result, a few small adjustments can release quite a bit of power capacity. We also have opportunities with Nvidia DSX MaxLPS, and we know other vendors are using power management tools to oversubscribe megawatts. There's a lot of flexibility within that capacity range. Until now, we've kept it simple, that is, 300 megawatts of gross capacity corresponds to a 200 megawatt IT load, but today we are also making it clear that there is considerable opportunity in the combination to release some of the surplus electricity.
Brett Knoblauch (Cantor Fitzgerald):
Very good. Another issue, about pricing reviews, I think was strong in preparing the statement and in the press release — referring to the $20 million deal, which is now around $25 million. Can you explain how long this is for? Is this an individual case or do you guys think this is average in discussions with customers?
Kent Draper:
This is not an isolated case, and we continue to see this level in our current live conversations with customers. It involves a variety of factors, as Dan said, including terms, advance payments, customer nature, future growth requirements, etc. But the pricing we saw was relatively consistent and continued to show an upward trend. Recently, there has been intense bidding for production capacity, so we are convinced that these numbers represent current market levels, and there is further upward pressure.
Daniel Roberts:
To be clear, these are 3 to 5 year contracts, not 2 years, and they are not spot capacity. If you want to optimize your quarterly income statement, there are plenty of opportunities to sell spot or on-demand computing power, which you might consider in the future as part of a combined strategy. But at this stage, we are very happy to build our customer base, diversify in the AI ecosystem, lock in long-term contracts, and obtain value-added financing through these contracts. In the long run, this will win us permission to intervene in on-demand, short-term contracts and use capabilities such as Mirantis to achieve higher revenue models.
Michael Donovan (Compass Point):
Back to the question of $20 million to $25 million in revenue per megawatt of IT. How do the economic benefits you see in the US compare to initial discussions in Spain and Australia?
Kent Draper:
We think it's very consistent. At the end of the day, it's a global market for computing power. Many customers who get their computing power from North America are overseas themselves, so it has indeed formed a truly global market. There may even be additional scarcity premiums in some other regions due to factors such as sovereign AI. As a result, the price levels we saw in North America were very indicative globally.
Michael Donovan (Compass Point):
One more follow-up question. Regarding the Nostrum acquisition, as far as I understand it, it brought about 490 megawatts. In your preparation remarks, it was mentioned that Badajoz accounts for about 300 megawatts. Are the remaining 190 megawatts distributed across several other sites? How should we understand it?
Kent Draper:
Yes, there are also several development sites in the Spanish portfolio, and we have locked in capacity. But Badajoz is the site we've chosen to focus on right now because it's the biggest one to be built soon.
Nick Giles (B. Riley Securities):
I want to ask a question about Texas. I know this isn't an easy topic, but I'm wondering if these developments have brought commercial advantages considering you already have two large-scale electrification sites there?
Kent Draper:
Are you referring to Governor Abbott's orders? Or some other specific aspect?
Nick Giles (B. Riley Securities):
Yes, Kent, that was exactly that instruction. Given the uncertainty of some early sites, did this push some potential deals in favor of you?
Kent Draper:
Yes. Regarding Governor Abbott's directives, the goals include transparency, grid reliability, water use, and impact on local communities. These are the topics we've focused on since day one and have always focused on. In the case of Childress, we fully funded all necessary grid upgrades for the project. As you know, we are located a few miles from major towns, have no residential neighbors, and our operations have no impact on the community. We use an efficient closed loop water cooling system, and the continuous consumption of water during long-term operation is minimal. Therefore, all of our practices in building and operating sites and combinations are in line with the direction of Governor Abbott's instructions. In this sense, we welcome additional transparency in the market. Also, as you mentioned, we already have two large-scale electrification sites in Texas, which in itself puts us in a very good position in addition to the factors mentioned above.
Nick Giles (B. Riley Securities):
Very helpful. One more follow-up question. We've seen many deals where the initial contracted capacity is followed by expansion options, and possibly with an exclusivity period. As far as I can tell, IREN doesn't seem to include such expansion options. Is this something you're considering, or are you guys deliberately avoiding it?
Kent Draper:
Yes, we've historically been deliberately evasive, but of course that depends on customized discussions with each customer. As you can imagine from Dan's discussion, we're very optimistic about the value of future infrastructure and computing power. Therefore, if we sell our options to someone else, we need to get compensation accordingly. This brings us back to the question of dynamic balance. Overall, we tend to avoid it before because we usually think we see more value than our opponent's bid.
Michael Colonnese (H.C. Wainwright):
Congratulations on the strong momentum of the AI cloud business. I just have one question. Can you provide more information on future uncontracted capacity distribution plans — specifically, which customer segments are you most likely to pursue as you enter 2027? And how to find a balance between pricing and financing costs when evaluating large customer contracts versus small customer contracts? I feel like you're monetizing at a higher level of the technology stack by acquiring Mirantis. I'd like to know your thoughts on evaluating deals.
Kent Draper:
Yes, Dan mentioned many aspects of our considerations in his performance overview earlier. We'll look at customer types, industries, and future growth plans. As you can see from our customer portfolio, we are gradually moving towards end users who directly use computing power. We have now added new customers in the AI native field, including a large leading cutting-edge AI lab and a number of AI developers related to physical infrastructure. As a result, we have seen a number of very attractive customer portraits in this field. For us, as Dan said, our goal is to diversify our customers over time. We already have a range of attractive long-term contracts in the portfolio, which lays the foundation for excellent economic benefits in the future. At the same time, it also provides us with additional flexibility to explore different monetization methods — whether it's improving economic efficiency through short-term reservation contracts, or actually providing on-demand computing power, or managed services rather than bare metal delivery. The position we are in today allows us to be selective in all of these areas, particularly in terms of the development of the non-investment-grade client financing market — something we have always expected to happen, and now we are actually seeing the evidence. This means we are still able to obtain highly competitive financing for different types of clients. Therefore, we are confident about our future prospects, can monetize the platform in different ways, and continue to obtain higher economic benefits.
Benjamin Sommers (BTIG):
I'd like to ask a question about the 2027 and 2028 customer discussions. Looks like we're probably not exploring very short-term contracts right now, but judging from customer feedback on the 3 to 5 year window, which time period do most customers tend to fall into? What do you think would be the ideal length of time?
Daniel Roberts:
Let me answer that, Kent. We're starting to see discussions about a longer period than 3 to 5 years. But that requires balance, right? We essentially operate the business as an investment portfolio—fixed tenants like Microsoft provide large-scale, long-term, and low-cost capital (6% GPU financing), while the broader AI developer market offers slightly higher pricing, good upfront payments, and significant growth potential. So we wanted to be flexible. At the end of the day, the framework for evaluating new customer contracts is what we're talking about today: what counterparties can bring, what are the economic terms, and what space can be opened up in the long term. There is a real trade-off in this. Investment-level fixed clients receive 6% of capital, while non-investment-level customers receive 9%. Looking at this alone, the investment grade seems superior, but if you look at pricing and upfront payments — the down payment covers about 50% of the GPU capital expenditure — those down payments are probably the most exciting part. On top of 90% financing, the customer paid half of the GPU cost in advance, which sends a very strong signal. It's not just about locking up production capacity; they've already started sharing construction funds with us. This is more indicative of the true state of demand than any pricing chart in the presentation.
Benjamin Sommers (BTIG):
Very useful. Just one more question, if you can. Can you provide some information on initial discussions on data center financing? You mentioned that this aspect might be explored in the future, so I'd like to know if there are any initial contacts at this point?
Daniel Roberts:
Yes, there have been lots of initial discussions and we'll keep you posted when we're done. But we've been busy. There is a lot of work on the customer side, and there is also a lot of action in GPU financing—that is a relatively easy fruit to harvest. As we mentioned before, we will consider refinancing the Horizon project after delivery, and it makes sense to wait until the operation is stable before obtaining better financing conditions. We finance the data center after delivery, but we don't rule out starting before delivery. One of the challenges and opportunities we face is vertical integration — data center capital expenditure began two years before production began because steel, transformers, and a large number of long-term equipment needed to be ordered. Despite many payment nodes falling behind, capital expenditure continued to occur before production was put into operation. Find an efficient way to finance that allows us to quickly expand to 5 GW or more — that's the big opportunity we face. Customers do exist, and no one is questioning that now. So for us, the key to everything is the capital flywheel, and how to manage it efficiently, so we can meet market needs, because we are indeed in a very unique position — locked in land and electricity when we started 8 years ago, and we have brought together our team, expertise, and partnership with Nvidia to build and operate these facilities. Capital efficiency is the next major breakthrough. GPU capital efficiency is now realized—the data we published today is objective proof, and it's only getting better. As for data centers, that's the next frontier, and we're actively advancing it, and we're excited about it.
Joseph Vafi (Canaccord Genuity):
Congratulations on all the progress. Dan, you mentioned earlier that electricity is still a major bottleneck. We discussed a lot of financing topics during this conference call. Obviously, you have plenty of resources available, but the entire industry is huge, and many people are also doing all kinds of financing. Could Anthony or Dan talk about the sustainability of the financing environment at a macro level — can the industry continue to fund this large-scale construction at the current rate? Do you guys have any concerns? What advantages can IREN have in different areas of the business? One more quick follow up question.
Anthony Lewis:
Thanks for asking. We just touched on some of the fundamentals. The current evolution of the GPU financing market is like this: initially it was a product dominated by private credit, with a return rate of about ten percent, but now the cost of investment-grade capital has dropped to around 6%. We have seen a significant expansion of the pool of funds involved in GPU financing — private, public, investment-grade, non-investment-level — all of which are positive signs. We also mentioned other factors, such as Nvidia's major announcement with the world's six largest capital allocation agencies, which will provide more support to the market. We also talked about prepayments, which are an important part of the overall financing situation. We're also seeing the data center financing market evolve to support infrastructure construction, covering high-yield and investment-grade sectors. Within the capital community, there is strong confidence in the sustainability and prospects of fundamental changes in technology, the investments required, and the ultimate return. I think it is this confidence that will attract continued capital inflows. Of course, we need to continuously adjust our plans according to market conditions. Market conditions will fluctuate, and we need to be flexible.
Daniel Roberts:
We often hear this question, Joe — can financing keep up with the expansion of asset classes? But look at what happened 12 months ago: GPU financing is almost non-existent. And over the past 3 months, we've raised $6.5 billion on both sides of the credit spectrum. It's not luck, it's a market that's taking shape. Markets are formed the same way every time. Think real estate financing—no one asks if an office with a contracted tenant can get a mortgage loan. Real estate has a complete capital structure, including construction finance, term debt, and institutional funds — due to contractual cash flow and real collateral. That structure took years or even 10 years to set up, and the capital structure of AI infrastructure is being built on a monthly or quarterly basis because the basic elements are the same: hard assets, contract revenue, and institutional counterparties. Blue Owl and PimCo are no small players; they are the world's largest infrastructure lenders, have now underwritten us, and we're maintaining a conversation with their top management. In fact, this analogy is even somewhat underestimated — a building is leased for a few decades, and the yield is only a few percentage points; our contract can recoup the computing power investment in about 2 years, and the customer also prepaid half of the capital expenses. Therefore, lenders can quickly see their funds recovered. Real estate financing has never been this economical. Whenever an emerging market appears, as long as the demand is real, financing will keep up — railways, telecommunications, real estate, electricity, etc. As to whether the demand is real, customers are proving it with action—through prepayments, through their end markets and performance, revenue, and customer appeal. An aspect that hasn't been fully developed is data center financing, but in the traditional sense, a data center is actually easier because it's closer to real estate and physical infrastructure. There's money there, and we know it's waiting for us to launch, and we'll move forward more carefully over the next few months.
Summary statement:
Daniel Roberts: Thank you all for participating. In a nutshell, the 2026 fiscal year is as follows: $4 billion of ARR has been signed, of which $1 billion has been put into operation; Horizon 1 has been delivered to Microsoft; and, as we have just discussed in depth, the financing model works efficiently on both ends of the credit spectrum. All of this was possible on a fraction of the platforms we have today. Finally, I would like to thank the iREN global team and everyone who has recently joined us — Mirantis, Nostrum, and many partners in the market — and keep up the good work. The plan remains the same: we will continue to deliver production capacity, transform it into lasting customer relationships, and finance in a self-regulatory manner. Thank you all, and I'll see you next time with the results.