IREN Limited Ordinary Shares 2026 Q4 Earnings Call

Good afternoon. Welcome to IREN's FY 2026 results presentation. I am 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 a 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 in that regard. With that, I will turn the call over to Daniel Roberts.

Thanks, Mike, and thanks, everyone, for joining us. Will and I started this business on a pretty simple observation. The digital world scales almost instantly. The physical world does not. Power, land, data centers, these things take years to permit, finance, and build. This was the year that stopped being a thesis and became the defining constraint of the whole industry. If we look at the chart on screen, across eight models tracked by OpenRouter, weekly token usage across large language models increased nearly 17 times in eight months. Every one of those tokens runs on physical infrastructure. It is very difficult to serve demand shaped like that with infrastructure on three-year lead times. This is the start of the cycle, not the end. Cheaper compute doesn't make existing things more efficient. It makes new things economic. Infrastructure enables applications. Applications create demand for more infrastructure.

Every build-out in history has worked this way, and that's the structural disconnect, and it's only getting wider. Let me walk through how we're set up against that backdrop. We operate across three layers from the bottom up. First of all, the data centers, the land, the power, the substations, the cooling. Arguably, the hardest layer to build, and that's where the shortage begins. Then the compute, the GPUs, storage, networking that go inside the data centers, and then finally, software on top. The managed services and enterprise support. That's where Mirantis lives for us. Just today, Mirantis was named an inaugural NVIDIA-certified hypervisor. We've now got NVIDIA validation at the software layer as well as the hardware. Why own all three? Because each layer makes the one underneath it worth more.

A grid connection is worth more with a data center on it, worth more again with GPUs inside, more again with services wrapped around the customer. Most of this market rents at least one of those layers. We own the entire stack. Here's how the year went. Just move on to highlights, please, operator. Thank you. Four things to take away from this update, and I will be quick because there is more detail coming a little bit later. Firstly, customers. New multi-cloud contracts, including Cohere, Prometheus, Perplexity, Figure AI, Fal.ai, Higgsfield AI, and separately, a leading frontier AI lab whose name we are not able to disclose just yet. Revenue. $4 billion of ARR is now contracted for our 2026 capacity, and $1 billion of that is operating today.

This does not include revenue expected to ramp in 2027, such as the $700 million of ARR associated with our NVIDIA cloud contract. Delivering. Horizon 1 was delivered to Microsoft this month, the first of four 50 MW deployments, with Horizon 2 through 4 targeted for the December quarter. Finally, funding. $6.5 billion of GPU financing now in the past three months. With prepayments, that is more than 100% of the associated GPU CapEx funded, and $2.8 billion of it needed no investment-grade off-take and still priced in the single digits. Let me start with customers because everything else follows from them. Our 2026 capacity is largely sold out. The questions we get now are all about 2027 and 2028. We are continuing to contract future capacity deliberately. Every contract opportunity gets weighed on three things. Firstly, who and what does this counterparty add to the platform?

The strategic merit, not just the revenue. Second, what are the economics? Price, prepayment, term, et cetera. Thirdly, what might it open up longer term for managed services and software? We have been saying this for a while now, signing deals is not the bottleneck in this market. Bringing GPUs online is. We also do not need an investment-grade off-take to fund GPUs anymore. We are not chasing headline announcements, we are making long-term decisions about where we want this business to be. When we will sign, we will tell you. We are in late-stage discussions with a range of new customers over a significant portion of 2027 capacity, and 2028 conversations are well underway too, both on customers and financing. Longer term, we want to keep building a deep, diversified base of customers across the AI ecosystem, all who have master service agreements with IREN. Every megawatt we build is scarce.

Before every new cluster switches on, we want it in demand from both existing customers and new ones. Not dependent on any one customer, any one negotiation, or any one point in a pricing cycle. Here is who is on the platform today. As I mentioned earlier, the headline is the new multi-year contract with a leading frontier AI lab. To be clear, this is a new contract. It is separate from Prometheus, who we can now name as the unnamed AI developer from our July announcement. The most sophisticated buyers of AI infrastructure in the world keep choosing us. It also shows the strategy is working. We started concentrated because big customers with immediate demand were the fastest path to scale. As the platform has grown, we have deliberately broadened. Hyperscalers, enterprises, AI developers, now frontier labs, across both training and inference.

But honestly, the part we care about most is the third bullet point. Existing customers keep coming back. Together AI and Fireworks AI have both renewed and expanded. New logos are proving demand, whereas renewals continue to prove ongoing delivery and ongoing execution. Once we have deployed with a customer, we grow with them across sites, GPU generations, and service levels. In terms of who is signing and growing with us, Prometheus and Figure AI are building products for the physical world, robotics, real-world automation, and they are contracting our compute to do it. AI is moving well beyond chat. This is the thing we keep telling people. You cannot size this market off today's usage, and that is exactly why the market keeps getting caught structurally short of compute. Let us move on to pricing. Pricing has moved a lot. Three-year contract pricing is up about 125% since November.

Five-year is up about 70%. Recent three-year contracts are pricing in excess of $20 million per megawatt of IT load, paying back the compute investment in around 2 years. While active discussions are now at around $25 million per megawatt. Recent customer prepayments are funding 45% to 55% of the GPU CapEx. What is behind that? The market has tightened, no question. But it is also who we are signing, how the deals are structured, and what we attach on top in managed services, which Mirantis keeps expanding. Shorter duration and on-demand sit above that as further upside in due course. Revenue per megawatt is shorthand, by the way. We look at the whole return, price, term, prepayment, utilization, customer quality, and expansion opportunity. To be clear about what we are not doing, we are not sitting on capacity to time a spot price.

We allocate capacity to build the customer base and the platform as we see fit. The pricing follows that. From customers to what we are building. The targets, roughly 300 megawatts of IT load delivered in 2026, and another half a gigawatt in 2027. That will take the platform to around 1.2 gigawatts in 2027 of gross capacity, and we are continuing to build across Texas, British Columbia, Oklahoma, South Australia, and Spain. Right now today, there are more than 4,000 people mobilized across our active sites. The best example of that to date is Horizon 1. As we announced earlier, Horizon 1 was delivered to Microsoft. First of four 50-megawatt liquid-cooled deployments at Childress, and it achieved NVIDIA Exemplar Cloud status on GB200 NVL72. Which matters because it proves we can integrate and operate the full platform, hardware, networking, and software, not just build the shell.

The delivery matters, but the template matters more. Every Horizon uses the same design, the same supply chain, the same site team, and each phase carries the lessons of the last. Horizon 2 is in the process of working towards commissioning. Three and four are in late construction. All three are targeting delivery in the December quarter. That approach is running at every site. Quickly on 2026. At Childress, retrofit work and GPU installs are running in parallel on the air-cooled calls. At Mackenzie, GPUs are being racked across the first 2 buildings. At Prince George, the air-cooled fleet is now fully commissioned, and liquid cooled installation is underway. All of that remaining capacity is targeted for the December quarter. Then 2027. Sweetwater 1 is in full swing now. First building is going up. The primary substation is progressing.

At Childress, Horizons 5 and 6 civils are now moving and underway. Another 250 megawatts of air-cooled conversion progressing. It is also worth mentioning in terms of Canal Flats, we have now decided to convert all of that to liquid cooling for GB300s, and that will deliver more value from power in a site we already own. Beyond that, the pipeline steps up again. 2028. Sweetwater 2 and Kiowa in the U.S., Bundey in South Australia, and Badajoz in Spain. Roughly 300 megawatts and the flagship of the Nostrum portfolio of sites we closed during the quarter. All up, more than 5 gigawatts announced with a multi-gigawatt development pipeline behind it, including more Texas power, where a tighter interconnection process should favor real and well-capitalized projects. A quick word on design because it answers a question we are starting to get a lot.

The bear case we hear on this industry is that AI data centers get rebuilt in 10 years. We have spent this year making sure that ours do not. Sweetwater 1 is becoming the reference design for Sweetwater 2, Kiowa, Bundey, and what follows after those. Common layouts, common equipment, more modularization and prefab. Each project is inheriting the last one's lessons instead of starting from zero, and the design is built for successive GPU generations. Evolving cooling, including 800-volt DC, all of which has been developed in collaboration with NVIDIA. We know compute changes faster than buildings, and ours have been designed to adapt to that. We are also getting more out of what we already own. Our existing sites have spare power beyond current deployment plans. That headroom can support more GPUs and more revenue without the need for any new grid capacity.