Q3 2026 Earnings Call Transcript

Good day, and thank you for standing by. Welcome to IREN Q3 FY '26 Results. [Operator Instructions] Please be advised that today's conference is being recorded. I would now like to hand the conference over to your first speaker today, Mike Power, Vice President of Investor Relations. Please go ahead.

Thank you, operator. Good afternoon, and welcome to IREN's Q3 FY 2026 Results Presentation, and thank you for your patience as we get assembled. I'm Mike Power, VP of Investor Relations. And with me on the call today are Daniel Roberts, Co-Founder and Co-CEO; Anthony Lewis, CFO; and Kent Draper, Chief Commercial Officer.

Before we begin, please note that this call is being webcast live with an accompanying presentation. For those dialed in by phone, you can elect to ask a question through the moderator after our prepared remarks. I would like to remind everyone that certain statements made during this call may constitute forward-looking statements. Those statements are based on current expectations and assumptions, and are subject to risks and uncertainties that could cause actual results to differ materially.

Please refer to Slide 2 of the accompanying presentation and our SEC filings for more information. During today's call, we will also refer to certain non-GAAP financial measures. As a reminder, a reconciliation to the most directly comparable GAAP measures is included at the end of the presentation. So with that, I will turn the call over to Dan Roberts.

Thanks, Mike, and thank you, everyone, for joining us today. 8 years ago, when Will and I founded this business, we spent a lot of time thinking about what the digital future actually meant for the physical world. We talked about films like the Matrix and Ready Player One, not as science fiction, but as a signal. Worlds where digital adoption was total, instantaneous and infinite. The insight we kept coming back to was this.

Digital adoption curves can go from 0 to 1 overnight, but the real world doesn't scale that way. Power, infrastructure, land, data centers, these take years to permit, finance and build. The bigger the demand, the harder delivery becomes. That gap between exponential digital growth and the physical world's ability to service it, that structural disconnect is exactly what we set out to solve.

That scarcity is now defining where AI infrastructure gets built and who can build it. 8 years later, that thesis is playing out exactly. And this quarter, we demonstrated what disciplined execution against it looks like at a global scale. In AI infrastructure, secured power is only valuable if it can be converted into customer-ready compute. That conversion is hard.

It requires site control, grid connection work, permitting, design, procurement, construction, GPU installation, networking, commissioning, financing and customer delivery, all coming together on tight time lines. IREN's strength is bringing those pieces together. We have experienced site teams, standardized designs and repeatable construction processes that allow us to build across multiple sites in parallel. As we scale, each phase builds on the prior phase.

The template becomes more repeatable, the procurement and construction process becomes more efficient, and the site teams carry that experience forward. That is where IREN has built its moat, and why real assets and real capabilities are harder to replicate than they might appear. That execution capability is showing up in the numbers, more capacity, more revenue, stronger funding certainty, and in the partnerships we are announcing today. This was a significant quarter and a significant week.

Let me run through the highlights. On capacity, we increased secured power to 5 gigawatts, added new sites in Europe and APAC, energized Sweetwater 1 on schedule, and have Horizon 1 GPU commissioning now underway for Microsoft. On customers, all of our operational capacity is fully contracted. We are not chasing demand.

We are racing to build supply fast enough to meet it. In this market, the moment compute comes online, it goes to work. That is the nature of the structural imbalance between AI infrastructure supply and demand, and it is why time to compute is the most important metric we track. We increased ARR under contract to $3.1 billion, remain on track to hit $3.7 billion exiting calendar 2026. And this week, signed a $3.4 billion 5-year AI cloud contract with NVIDIA.

The first step in a broader strategic partnership I will come to in a moment. On capital, we had $2.6 billion of cash at April 30, and we continue to progress GPU, data center and corporate level financing initiatives to support the next phase of build-out. But the headline today is the NVIDIA partnership, and it deserves a little more than a bullet point. Let me explain what this partnership actually means.

We are working with NVIDIA to support deployment of up to 5 gigawatts of NVIDIA DGX-aligned AI infrastructure across our global data center platform, alongside DGX environments and the DSX AI Factory reference architecture. The $2.1 billion NVIDIA investment is structured to reflect that. Their rights to invest only vest as NVIDIA GPU infrastructure is deployed across IREN campuses, and only fully vest upon deployment of 600,000 GPUs. NVIDIA's capital is directly tied to execution.

That's not a passive financial investment. NVIDIA is a partner who wins as we deliver. The $3.4 billion AI cloud contract announced today, supporting NVIDIA's own internal workloads is the first step in that partnership. 8 years ago, Will and I set out to build the infrastructure the digital world would need. Today, the world's leading AI infrastructure company has chosen IREN as the partner to help build it.

This next slide shows exactly how we build against this. So here is our plan. In 2026, we are targeting 480 megawatts of AI cloud capacity, 150,000 GPUs and $3.7 billion of ARR by year-end. That is the near-term plan and the clearest bridge from capacity to revenue.

In 2027, we are scaling to 1,210 megawatts, with an additional 730 megawatts currently under construction across British Columbia and Texas, including Childress and the initial phase of Sweetwater 1. The construction flywheel we are running in 2026 carries directly into this next phase. Beyond 2027, we are building against a 5 gigawatt global power portfolio. North America, our new European platform in Spain and an APAC pipeline anchored by large-scale Australian opportunities. The sequence of delivery matters because it dictates time to compute, and time to compute is what drives revenue.

Each phase supports the next, that's how the platform compounds. One more thing before we move on. This week, we welcome Mirantis into the IREN family. 650 engineers, operators and customer support professionals who have spent more than a decade running cloud infrastructure for over 1,500 enterprise customers globally. To Alex and the whole Mirantis team, welcome.

I'll come back to what this means for our delivery capability later on. But let me start with 2026, where construction and customer demand are coming together most visibly. The 2026 expansion is focused on delivering 480 megawatts of AI cloud capacity across Childress, Prince George and Mackenzie. This is where the road map translates into near-term deployments, customer handoffs and ARR conversion.

We'll start with the largest and most complex 2026 work stream, the 300-megawatt Horizon 1 to 4 liquid cooled deployment at Childress, where NVIDIA GB300 NVL72 installations are now underway. Horizon 1 is scheduled for Microsoft handoff in Q3, and Horizons 2 to 4 remain on track for delivery by the end of this year. This is a major execution milestone. It demonstrates our ability to design, build, fit out and commission large-scale next-generation liquid cooled infrastructure for a hyperscale customer on an accelerated schedule.

We have around 3,000 workers on site right now. That level of activity reflects both the urgency of AI infrastructure demand, and also the depth of our execution capability on the ground. Importantly, the model is repeatable. Horizon 1 establishes the build template.

Each subsequent phase benefits from the same design, supply chain, construction sequencing and site team. That is how we drive faster deployment phase every time. Alongside the liquid cooled build, we are also converting existing air cooled capacity into AI cloud deployments across British Columbia and Childress. In British Columbia and Childress, we are progressing 180 megawatts of air cooled AI cloud capacity by leveraging existing infrastructure.

At Prince George, all air-cooled GPUs have now been delivered and are either operating, or undergoing commissioning across the 50-megawatt site. At Mackenzie, 80 megawatts of data center capacity has been prepared for GPU installations commencing in the second half of 2026. And finally, at Childress, data center retrofits are underway across an initial 50 megawatts ahead of GPU deliveries in the second half of this year. This is a capital-efficient part of the road map. We are taking existing sites and converting them toward higher-value AI cloud workloads.

And it works because we already have the operational teams, infrastructure and site control in place. Air cooled capacity can come online faster than liquid cool. In a market where time to compute is everything, that speed is a commercial advantage, and we are using it. We're already seeing this dynamic play out commercially with capacity continuing to be contracted ahead of commissioning as customers prioritize speed to market.

We now have $3.1 billion of ARR under contract, including approximately $700 million of ARR associated with a $3.4 billion 5-year contract for Blackwell GPUs to be deployed across 60 megawatts of air cooled capacity at Childress for NVIDIA. Against the full 2026 expansion, we are targeting $3.7 billion of ARR by year-end across 150,000 GPUs. The remaining uncontracted capacity represents approximately 50,000 air cooled GPUs scheduled for delivery in phases through the second half of this year. Demand for that capacity is robust.

Our focus is on using our time to compute advantage to secure the right customer mix. With the 2026 plan on track, let me turn to what comes next. The 2027 expansion where the pipeline -- the platform rather scales to 1,210 megawatts. So the 2027 plan is about demonstrating that we are -- that what we are building in 2026 is not a one-off.

It is a repeatable, scalable model that should accelerate over time. Here's what that looks like in practice. In British Columbia, Canal Flats is another example of converting existing infrastructure into AI cloud capacity. We plan to retrofit all 30 megawatts of existing air cooled capacity to support AI workloads.

Capital efficient, fast to execute and consistent with the same model we are running at Prince George and Mackenzie. In parallel, Childress continues to be the largest single contributor to the 2027 setup, with both new liquids cooled capacity and additional air cooled retrofits adding a total of 400 megawatts of gross capacity. At Childress, the 2027 plan includes 100 megawatts of additional liquid cooled IT load for Horizons 5 and 6, as well as retrofitting an additional 250 megawatts of existing air cooled capacity. Of that 250 megawatts, approximately 60 megawatts will be deployed to support the NVIDIA AI cloud contract.

The combination of new liquid cooled data centers and air cooled retrofits gives us real flexibility. We can support next-generation high-density deployments while continuing to use existing infrastructure where it is the right technical and economic fit. That flexibility is part of what makes Childress such a productive campus. In parallel, Sweetwater becomes the next major Texas campus in the 2027 plan.

At Sweetwater 1, the high-voltage substation has been energized on schedule, and construction is now underway for the initial 200-megawatt IT load phase of liquid cooled data centers. Energizing the substation is an important milestone. It moves Sweetwater from development into execution, and establishes the electrical foundation for the broader site build-out. Sweetwater 1 is being designed for next-generation chip architectures, including the NVIDIA Vera Rubin.

Like Childress, we are deliberately sequencing the build so that the first phase creates the backbone for faster subsequent phases. The first 200 megawatts is not just the first 200 megawatts. It is the foundation for a much larger site. The commercial pipeline for our 2027 capacity is anchored on the same principle that is driving everything we are building.

Our vertical integration is a genuine advantage for customers because we control more of the critical path than anyone else in this market. Power, land, data center construction, the pieces that cause delays for others are the pieces we own and control. Customers want certainty that capacity will be available when promised. The phase 2027 build-out plan gives us a concrete basis for those conversations, and we are having them.

We are in the process of negotiating large-scale AI cloud deployments across our 2027 capacity today. Demand is not the constraint, however. It is highly unlikely to be the constraint. The priority is delivering capacity on schedule and converting our time to compute advantage into durable long-term customer relationships.

We do expect the customer mix to evolve over time. hyperscalers, AI natives, enterprises and on-demand use cases, but we do not need to force that outcome. The platform will attract the right customers as it continues to scale. Beyond 2027, the same execution model extends into a much larger 5-gigawatt global platform. We now have 5 gigawatts of secured power.

To put that in context, that is not a pipeline number or an aspiration. That is secured power, and it represents one of the largest portfolios assembled for AI infrastructure anywhere in the world. The question now is how we build against it. The answer is a phased global platform across North America, Europe and APAC with additional development opportunities beyond that.

Let me walk you through each region. We'll start with North America, which remains the largest component of the long-term platform. In North America, the next major phase is driven by Sweetwater and Kiowa, our flagship gigawatt scale campuses in Texas and Oklahoma, where data center capacity is expected to commence ramping across 2027 and 2028. We also have multiple development projects advancing through the connection processes, including batch zero candidates in Texas, which represents some of the most strategic valuable interconnection opportunities in the country. The North American pipeline has a natural progression of scale.

Childress demonstrates the operating model today, Sweetwater expands it across an even larger campus and Kiowa provides the path to another hyperscale tier opportunity as power ramps from 2028. Every campus builds on the last. That's the compounded effect of having secured the right land and power positions early. At the same time, we're expanding the platform into Europe through Spain. Today, we announced the acquisition of Nostrum Group and with it our entry into Europe.

The transaction adds 490 megawatts of secured power in Spain, a gigawatt scale development pipeline and a team of more than 50 people across development, engineering, construction and operations. But what it really adds is a platform and the right people to build it. I want to acknowledge Gabriel Nebreda and the Nostrum team. Gabriel spent nearly 2 decades in European energy at EDP Renewables managing gigawatts of operating assets across multiple European markets and most recently as CEO of EDP Solar.

He understands European power infrastructure as well as anyone, and we are excited to have him leading IREN's European platform. Spain is the right place to start. Supportive AI policy, abundant renewables, lower build costs and strong connectivity into broader European demand. Europe is a market where power availability and grid time lines are increasingly shaping where customers can actually deploy.

And Spain gives us a credible scalable answer to that question. This is not just a power acquisition. It's the establishment of IREN's European platform. From Europe, we move to the other side of the world and an opportunity that matches the scale of everything we've just described.

Australia is obviously not a new idea for us. We have been progressing large-scale Australian projects towards secured grid access for some time. And we think the opportunity here is as significant as anywhere in our portfolio. And this is why.

Asia Pacific is home to roughly 4.8 billion people, around 60% of the world's population. That includes some of the fastest-growing AI demand markets on earth. Indonesia, Singapore, Japan, Korea, the infrastructure requirement to service that demand is enormous, and it is largely unmet. Australia is uniquely positioned to serve it, abundant renewables, a trusted jurisdiction, strong rule of law and as the submarine connectivity map shows, direct fiber links into major demand centers across the region.

It is the natural anchor point for AI infrastructure service in APAC. We are already seeing hyperscalers and frontier labs make significant commitments to Australian operations, and we intend to be a major part of that story. Beyond Australia, we continue to progress global development opportunities that extend IREN's runway further still. The platform we are building is designed to create scale into demand wherever it develops, and the pipeline gives us the flexibility to do exactly that.

That is the global platform, secured power across North America, Europe and the development pipeline extending into APAC and beyond. But securing power and building data centers is only part of the equation. The other part is what happens when the compute goes live. How it is deployed, managed and supported for customers at scale?

That is where I'd like to spend a moment on Mirantis. This week, we welcomed Mirantis into the IREN family, and I want to take a moment to acknowledge that. 650 people joined IREN this week. Engineers, operators, customer support professionals, a team that has spent more than a decade building and running cloud infrastructure for over 1,500 enterprise customers globally. That track record speaks for itself.

And what they bring is specific. Their k0rdent AI platform manages AI infrastructure across bare metal, virtual machines and Kubernetes environments, exactly the complexity our customers are dealing with as deployments scale. They are also a founding ISV partner of the NVIDIA AI Cloud-ready initiative, which means they are already deeply embedded in the same ecosystem we are building into. As we scale, delivery is not just about bringing GPUs online.

It is about what happens after, provisioning, monitoring, supporting customers through increasingly complex environments. Mirantis strengthens all of that. We are already seeing and they will play a central role in supporting our NVIDIA AI cloud contract. To Alex and the whole Mirantis team, a big welcome.

We are super excited to have you. So what you have heard today is a company that has secured power at scale, is contracting revenue at scale and is now building delivery capability at global scale. Anthony will now walk you through how we are funding it.

Thanks, Dan. The capital strategy is designed to support the phased build-out of capacity Dan discussed, while maintaining flexibility and capital discipline. As of April 30, we had $2.6 billion in cash and cash equivalents. We expect this, together with operating cash flows, GPU financing and additional financing initiatives to support our near-term CapEx program, which includes delivery of the Microsoft contract and deployment of aircool capacity across Mackenzie and Childress.

For GPU CapEx, we are leveraging secured debt and customer prepayments. As we have noted previously, approximately 95% of Microsoft GPU-related CapEx is expected to be funded through prepayments and GPU financing. And we have work streams underway for additional GPU financing to support upcoming deployments. On the data center side, we expect our financing approach to evolve as projects move from development to construction and contracting, and ultimately to stabilized operations.

Early-stage development can be supported by balance sheet capacity and corporate level sources. As projects reach construction and customer contracting milestones, asset and project level financing can be introduced. And as assets are stabilized, refinancing and capital recycling can help support future builds. And as I note we will continue to maintain a disciplined balance of debt and equity as the platform continues to scale.

I will now turn to the financial results, which continue to reflect the transition underway from Bitcoin mining to AI cloud. Revenue was $144.8 million for the March quarter, compared to $184.7 million in the prior quarter. Within that, Bitcoin mining revenue was $111.2 million, down from $167.4 million, driven by a lower average Bitcoin price and the ongoing decommissioning of mining hardware ahead of GPU installations. This was partially offset by continued growth in AI cloud services revenue, which increased to $33.6 million, compared to $17.3 million in the prior quarter.

Cost of revenues decreased by $25.9 million, primarily due to electricity costs from reduced Bitcoin mining capacity. Net loss for the quarter was $247.8 million, impacted by noncash impairments of $140.4 million, primarily related to the decommissioning of mining hardware, as well as $23.7 million of unrealized losses related to cap calls associated with our convertible notes. As we continue to transition our remaining Bitcoin mining operations towards AI cloud, we expect to incur additional noncash impairments associated with decommissioning mining hardware. These outcomes reflect the strategic reallocation of infrastructure toward AI cloud growth, which we believe is the higher-value long-term opportunity.

Adjusted EBITDA was $59.5 million, compared to $75.3 million in the prior quarter, primarily on account of the revenue and cost of revenue items noted above. So as noted, the quarter reflects the ongoing transition from Bitcoin mining to growing AI cloud. As Dan noted earlier, we continue to target $3.7 billion in ARR by the end of calendar 2026. We expect that ramp to be back-end weighted with Microsoft revenue, and revenue from the additional 50,000 GPUs procured during the quarter expected to begin ramping in Q3 2026. I will now turn back to Dan for closing remarks.

Thanks, Anthony. So 8 years ago, Will and I asked a simple question. What does the world need to build the right digital future? The answer was power, land, data centers and compute and the ability to bring them all together at scale faster than anyone else. Today, that thesis is playing out, and we are just getting started. With that, we will open the call for Q&A.

[Operator Instructions] First, we have Mike Ng from Goldman Sachs.

Congratulations on all the progress. I just had two questions, if I could. First, on the 5-year NVIDIA AI cloud contract, I was just wondering if you could talk a little bit about how many GPUs are being supported by the 60 megawatts and the cost per GPU? And then second, for Sweetwater and Oklahoma, I think you mentioned the data center capacity is coming in, in '27 and '28. I was just wondering if you could talk a little bit about like at what point do those sites become marketable, or maybe they already are?

And what milestones do you typically need to hit to increase the likelihood of a tenant being willing to take that out?

Let me take that one, Dan. So with respect to your first question, we haven't disclosed the specific amount of GPUs. But as we mentioned on the call, approximately 60 megawatts of air cooled Blackwells. And we think that the contract value that we're getting and obviously, the relationship that we continue to build with NVIDIA is very beneficial coming out of that contract.

Importantly, this is a managed services deployment. And so it shows our ability to be able to service different segments of the market as we move forward. With respect to your second question, as Dan mentioned earlier, we are still seeing extremely strong levels of demand within the industry, certainly outstripping supply. And what we continue to see as we move forward is that capacity becomes increasingly scarce further out than people were expecting.

So if we rewind even a number of months ago, '27, people were thinking that there was relatively decent amount of capacity available. We're already seeing that capacity available in '27 is extremely scarce, and that is continuing to push into '28 now as well. So for us, there is certainly the ability to market those sites for '27 and '28 online dates. As Dan mentioned earlier, we're working through the type of customers that we bring into the mix and making sure that we are structuring the contracts in the right way to enable a flywheel at our end.

But certainly, the demand signals are very strong.

And maybe just to add to that quickly, Kent. I think to directly answer the question, there's nothing stopping us contracting that capacity today. It just gets easier, the closer you get. So the focus is on time to compute.

The demand we know is there, and all it does is make the conversations and the negotiations that we are having live time for a lot of that capacity much easier when you've got a defined construction and delivery plan, rather than trying to make things up on the fly in parallel with a full form agreement.

Next, we have Paul Golding from Macquarie.

Congrats on all the progress and the new relationships coming in-house. I wanted to ask about air cooled GPUs in general. So it sounds like with the 60-megawatt deployment at Childress for NVIDIA that will be an air cooled deployment along with the rest of the uncontracted capacity that you're deploying across British Columbia and Texas. Air cooled is going to represent a meaningful part of the strategy.

I just wanted to ask how you see efficiencies, as well as hardware performance looking so far based on the deployments that you've planned for? And how we can look at that from a financial perspective as well as we think about the model and the air cooled opportunity?

So in terms of efficiency and performance, I mean, what we're deploying across the air cooled portfolio is the latest generation of NVIDIA air cooled GPUs being Blackwells. So they perform extremely well. There is very high demand for those across all Blackwell GPU types and certainly continue to see customers finding a very good degree of performance versus cost efficiency from those units over time. And sorry, Paul, I didn't quite understand the second part of your question in relation to how that converts into revenue over time.

That's right, Kent. So just wondering with, sort of, retrofitting and repurposing of Bitcoin mining infrastructure for these air cooled deployments, how that seems to be working out maybe from a margin perspective relative to some of the liquid-cooled deployments that you're doing around the Horizon projects, just given the simpler cooling opportunity there?

Yes. From an operational margin perspective, it is slightly more efficient than the liquid cooled deployments. But where we get the real benefit is, as Dan mentioned earlier, it's very capital efficient because we're taking existing air cooled data centers that require relatively little CapEx to retrofit them compared to brand-new build liquid cooled facilities. So that is the major difference in terms of the two.

At an operating margin level, yes, air cooled is probably slightly higher but immaterial.

And if I could just sneak one more in around Europe and the Nostrum acquisition. As we think about the road map there, are you looking to use a similar form factor to what you've used either at Horizon or with air cooled facilities? Or is there a bespoke form factor you plan to leverage from that platform as you do the European rollout?

Yes. So one of the things that attracted us to the Nostrum opportunity, and we've been looking at Europe for a while is that they did have significant land holdings that came as part of that and access to a large amount of secured power. So that gives us quite a large degree of flexibility as we build out that platform over time as to the form factor that we use. Typically, in Europe, you do tend to see slightly more condensed build-outs, but we do have the ability there to utilize our typical modular design that we use across North America, which obviously may well bring construction advantages with it.

So that was one of the key elements that we saw in terms of the platform that they have and the projects they've developed.

Next, we have Brett Knoblauch from Cantor Fitzgerald.

Congrats on the, I guess, multiple acquisitions over the last week and the NVIDIA partnership and deal. I wanted to touch on Mirantis a bit because I thought that was important to the long-term story. Could you maybe just elaborate how that fits into your go-to-market motion, how it might accelerate your go-to-market motion when it comes to landing these enterprise deals, which is also what it seems like the NVIDIA partnership wants you to do as well?

Yes. Happy to take that one initially, Dan, and then you can add. So it brings with it a number of elements that we think are significantly attractive to our business, the ability to deploy quickly, the ability to service enterprise customers that may require a high level of software over and above bare metal. They also, as a large company that has very big internal engineering resources, bring very good capability on the software development side, and that can flow through to the business, not only in terms of the software stack, but also the operations of these large clusters more generally.

And further to that, again, having serviced customers for decades, they have an extremely well-built out customer support function internally. So all of those elements are things that attracted us to the Mirantis team and -- are able to add to the existing skill set and customer service support that we've already built up internally.

Awesome. And then if I could maybe just do a follow-up. Just double-clicking on the capacity ramp for '27. Am I right in thinking that of the 730 megawatts, 450 will come from the remaining Childress capacity and, I guess, the 280 would be coming from Sweetwater?

That's correct.

Next, we have Nick Giles from B. Riley Securities.

Congrats on all the developments here. I know the IREN team has a lot of experience in developing infrastructure in Australia, but maybe less under the IREN platform. So I was curious if you could walk us through some of the key differences, specifically in power procurement, maybe commercial strategy, so on and so forth?

Sure. Look, in some ways, Australia is very similar to other markets. And the operation of the electricity market in Australia managed by AEMO is very similar to what we see in Texas as ERCOT. There are markets in Australia, which resemble Texas in other ways, lots of land, good transmission line capacity, good fiber connectivity and abundant renewables, which isn't located close to other demand centers similar to what we see in West Texas.

So there are a lot of parallels. The reality is Texas is just an easier place to do business, and we've been able to accelerate faster there. But it hasn't stopped us continuing to incubate projects down in Australia, and we're getting far closer to those projects becoming a bit more of a reality. And I think the demand environment and the ability to service APAC, and the demand constraints that we're seeing and hearing in our conversations with hyperscalers, means that Australia looks like a fantastic frontier for us, and we'll look to accelerate that in parallel with North America and Europe.

Next, we have Michael Donovan from Compass Point.

Congrats on the progress. How should we think about regional customer mix as the platform expands? Are certain markets globally better suited for enterprise and sovereign AI customers versus hyperscalers? And does that change the expected contract structure or margin profile?

Look, it's going to evolve, and there's a lot of unknowns around this. But if you break it down, hyperscale contracts can mean two things. They can mean hyperscalers using capacity for their own purposes in terms of training and servicing workloads such as their own AI models. Or it can mean they're just acting as intermediaries to aggregate capacity for end customers that we're talking to directly.

So obviously, in the case of the latter, whether you're dealing with a hyperscaler or going directly to the end customer, the end demand is the same. Then you've got different types of workloads, so inference and training. Inference is a little more latency sensitive. Training, you can probably afford a bit more latency.

And indicatively, we've had conversations around training models in Australia. Yes, the U.S.A. to Australia is a long geographic distance, but it's actually not that far over fiber, particularly where you're talking about training models. And given where inference sits today as well, we're all using ChatGPT or Claude, the response times are still, I guess, adjusting to the level of demand and the supply to service it. So look, it will evolve over time, and our objective is to build out an expansive ecosystem of end customers.

The partnership with NVIDIA is designed around that. The Mirantis and integration into our business is designed to help facilitate that over time in addition to all the near-term operational capabilities that it brings out. So the goal is very much to build out that diversified customer base over time across all of those markets.

Appreciate that. And a follow-up, if I may. Can you help bridge the 490 megawatts in Spain from secured power to time to first token? And what has to happen before construction begins?

So that is secured power and the sites across the portfolio there are secured as well. So from here, it's a matter of working through final design permitting, which is already well advanced at a number of those sites and then ultimately, construction of those facilities. But one of the elements that we found very attractive was the near-term security of power. So that is power that is available on a time line that we think is going to tie in very well to general European demand, and we are already seeing a number of direct requests from existing and new customers for European capacity.

Next, we have John Todaro from Needham & Company.

This is Austin on the line for John Todaro. Maybe just a quick question on how do you intend to finance the build-out for the recently announced NVIDIA deal? It seems to be around 5 gigawatts. So just any color on that would be helpful?

I can take that. Yes. So the CapEx involved for the retrofitting of the air cooled data centers in Childress is pretty modest in the scheme of things. In terms of the CapEx for GPU, obviously, we've got a range of financing sources available to us.

That obviously includes initiatives at the corporate level, but we can also look to finance GPU acquisitions in various ways in the debt capital markets through debt capital as well. So we'll be looking at all those initiatives.

In terms of the 5 gigawatts more broadly, maybe just to address that in the plan. So that's obviously a lot of capital today, but the reality is you don't need all that capital day 1. There's an S-curve of construction that takes time. It takes years to deliver this. This is the whole point around time to compute.

It's not just a case of getting power and land. It's assembling multi-thousand construction teams and actually delivering it. And the funding for that just is progressive over time. So as we've seen, as we continue to deliver, we continue to drive revenue, we can reinvest that revenue in CapEx, and it continues to unlock more and more financing sources over time.

And part of the partnership with NVIDIA, we've announced they've got the ability to invest in IREN and as we commission GPUs. But equally, there's other support mechanisms being discussed to the extent that we need them. But the reality is capital markets are open. They've been very supportive of our plan, and we anticipate that continuing at the moment that, that changes.

There's a whole world of capital out there in terms of other options, whether you're creative around private markets or otherwise. And when you look at the GPU financing, which is the lion's share of that CapEx, the Microsoft contract is a great template. We financed 95% of that CapEx at an average interest rate of about 3% through prepayments and GPU financing. So the capital is out there as long as you sign good contracts and you show that you can execute and operate this capacity.

Next, we have Joseph Vafi from Canaccord Genuity.

My congratulations here as well on the great progress. Just a couple of thoughts, or some of your thoughts here just to gauge demand out there. I know you threw out a $3.1 billion contracted going to $3.7 billion contracted in ARR here exiting the year. Your confidence in that uncontracted capacity and signing contracts, how is the demand out there for that, say, extra $0.5 billion of ARR?

And what kind of clients you may be looking to bring on board there? And then I'll have a quick follow-up.

I think -- sorry, go ahead, Dan.

Sorry, Kent. Look, again, we're trying to reiterate this as much as we can, and I'm very happy for someone to point it out, but there are no idle GPUs. And the prospect of there being GPUs sitting there unused, given how structurally constrained this market is a little over the near term, but in the medium term, it's not the focus. We are having a lot of customer conversations, but all of our operational capacity is fully contracted.

We're contracting substantial portions of capacity before it even arrives. And we're in discussions with a variety of customers all the way from hyperscale clients down to AI native labs all of that 2026 and 2027 capacity. So when a signature is put on a paper, it just flows naturally. Our conviction is around the demand supply, and you cannot tap into that unless you bring the capacity online.

And this is a customer contract doesn't deliver revenue. Having compute online delivers revenue, and that has been the focus.

And then I was going to say many of the same things. So the one addition I would add is particularly for our air cooled capacity where we are adding substantial amounts across second half of '26 and into the early part of '27, there is very significant demand on those time lines. That is the most constrained portion of the market, and that is directly what is leading into the dynamic that Dan discussed where they just are not idle GPUs that are not being used in this market. Everything on shorter-term time lines is extremely attractive to counterparties.

Got it. That's great color, Dan. Just then on your strategy and philosophy around customers and diversification there, if you are in the catbird seat here relative to fulfilling demand from multiple parties. How are you looking at your broadening, deepening, diversifying that customer mix over time?

It's something that we're looking closely at, Joe. Like there is no set formula as to the proportional splits between different types of customers. There are benefits in having hyperscale clients in terms of financeability, contractual certainty, but there are also consequences in terms of price because you're not servicing the end customer in many of those instances. But the ability to service the end customer has been something we've focused on since day 1. All of our early deployments have been very focused on non-hyperscale customers and getting as close to AI natives and enterprise as we can.

So the Mirantis acquisition certainly helps that. I'm not going to sit here and say we're going 100% hyperscale. We're going 100% AI native end market. The reality is that blend will just emerge organically over time.

And this, again, is part of the close working relationship we've got with NVIDIA. We've spent a lot of the last fortnight in their San Jose office working through how we service all types of customers, all the way from the $1 trillion hyperscalers through to the emerging AI scale-ups where a lot of this innovation and development is taking place. And it's funny speaking to someone the other day, you don't need a sales team in this market, particularly when you've got NVIDIA, they see the whole ecosystem. The introductions, the referrals, putting us in touch with anyone that needs capacity.

It's just happening so organically, so quickly live time that it will just play out a good way. But I think a combination of hyperscale, a hot combination of other is absolutely the goal.

Our last question comes from Ben Sommers from BTIG.

So I was curious a little bit on older generation GPUs. I know you've talked in the past as you've seen the useful life of older generations for like H100s extend out further than maybe people had originally thought. So kind of curious what you're seeing on the demand profile there, just what potentially type of workloads are going on for those older generation GPUs?

Yes. The comments that we made about no idle GPUs that applies to all GPUs, not just latest generations. So older generations, A100s, H100s, H200s, all fully -- effectively fully utilized across the industry. So the demand picture continues to be strong.

In some instances, you're actually seeing pricing for older generation units climbing significantly, and there's a number of observable pricing points out there in the market where you can see that happening. Yes, the type of demand may shift over time. You may have older generations being used more for inference, but also those older generations are equally suitable for certain types of training. So we just see strong demand across the board, both on the inference and the training side, and that continues to drive demand and [ elongated ] life cycles for those older generations of equipment.

Great. And then just on potential future conversations that you're having for potential contracts down the line, is there any talk of prepayment structures similar to that of Microsoft? Or just kind of curious what you're hearing in the market on that end?

Yes. It certainly plays a role in a number of those conversations, and we are still seeing prepayments being on the table in a large number of instances. Now it obviously factors in as part of the overall equation. So it's not the single factor that you're looking at.

Everything has to go together with a combination of term length, prepayment, creditworthiness, price, but prepayments are certainly very much on the table in the current environment.

I see no further questions at this time. I'll now pass to Dan for closing remarks.

Thanks, operator. Thanks, everyone, for joining us today. We remain focused on execution, delivering the 2026 plan, advancing the 2027 build-out and positioning our now global platform for the opportunity beyond that. And we look forward to updating you as we deliver. Thanks, everyone.