We're looking at maybe the largest physical infrastructure build-out in generations, and I don't think anyone outside the industry has internalised the scale. What's the single most underappreciated thing about this data center cycle that the market still hasn't priced in?
A couple of things. Everybody appreciates that demand currently exceeds supply, but I don't think people quite grasp how quickly the demand side is advancing. In the digital world there are very few impediments to growth, so you just see this explosion of demand. Every time you get an improvement in a model, or compute becomes cheaper and you can open it up to more use cases, it results in more demand, and that keeps building on itself.
You're getting a genuine exponential curve, and humans struggle to comprehend exponential growth because almost everything we see in our lives is linear. When you're confronted with an exponential demand profile, it's very hard to get your head around it.
On the supply side, it's getting harder and harder to deliver this infrastructure at scale. When the industry was 5 gigawatts, going from 5 to 10 was very doable, but doubling the data center industry from where it is now is extremely difficult. Whether it's grid connection processes, finding additional generation, finding land, or permitting, you're getting more and more of a wedge between the demand profile and the supply that can be brought online. That was really the entire reason for being for IREN. The co-founders had the view that everything was going to digitize and you would need real-world infrastructure to service that, but the real world runs on multi-year development cycles versus the instantaneous exponential demand curves you see in the digital world.
Where would you identify the main bottleneck? We've had very different answers. Applied Digital pointed to five or six-year gas turbine lead times, while TeraWulf said that wasn't the case at all because of their brownfield work.
There really are bottlenecks throughout. Power absolutely is one. If you want to start from scratch today with no land and build a grid-connected data center, you're looking at a minimum of five to seven years before you can bring anything online. Behind-the-meter generation like gas turbines has very long lead times today. People are starting to use fuel cells, smaller turbines, or even diesel generators at scale, but none of those are efficient, several haven't been proven out at large scale, and many of them pollute. So access to power is a challenge, but you also need the data center equipment to utilise that power, and there are supply chain issues and labour shortages. Specific items like high-voltage circuit breakers, high-voltage transformers, and chillers for liquid cooling all carry multi-year lead times.
Once you have the equipment you have to build the centres, and in certain markets labour is now hard to find. One of our advantages is that we've been building consistently for six years, so we have the relationships to get incremental labour over time. But if you're starting today with no presence in a market and you need to find 4,000 skilled electricians and mechanical labourers, that is an absolute challenge. On the compute side you're also seeing shortages in networking equipment and, as is now well understood, on the storage side with DIMMs and SSDs. So it's not one bottleneck, there are a number, and you have to attack all of them to deliver projects.
You've secured roughly five gigawatts of grid-connected capacity. What was your source thesis, and where does the marginal megawatt come from now that every hyperscaler is hunting the same interconnects?
The original thesis was this dislocation between the real world and growth in the digital world. Going right back to the founding of the business, we'd identified a number of use cases: machine learning and AI, rendering for videos and movies, oil and gas reservoir modelling, and cryptocurrency mining. We expected the world to continue to digitize, and off the back of that thesis we got started very early on development. A number of the projects in our five-gigawatt portfolio we began working on in 2020 and 2021, pre-ChatGPT, when there wasn't as much demand, and that's why we've been able to establish such a large position.
If you're starting from scratch today it's very, very difficult to do anything on a decent timeline, so that's one of our genuine competitive advantages. There is a time-locked window now to establish yourself as a major player where new entrants simply can't come in because of those bottlenecks. The five gigawatts is just the capacity where we've secured binding, executed grid connections, but we have a very large pipeline of development projects at various stages sitting behind that as well.
There's been a lot of discussion about circularity with NVIDIA backstopping demand for its own GPUs. How do you push back on the view that this overstates true end-customer demand?
NVIDIA has a view of the entire ecosystem. Much of the software stack has been built around products designed to run NVIDIA GPUs, so they see the whole demand picture, all the new AI startups and enterprises. The backstop arrangements are often criticised, but they're designed to bridge a very specific window where you have new AI labs or enterprises that are relatively small and not themselves investment grade, yet have large demands for compute. If they're your customer as a neocloud, it's harder to finance buying compute for them, so NVIDIA has stepped in to provide a backstop that is unlikely to ever get hit, because the real end-user demand will be there. That allows people to finance bringing compute online in the meantime.
So a lot of the perceived circularity I don't think will ever actually happen. What NVIDIA is doing is establishing the ecosystem and providing a financing mechanism for natural levels of demand. Some of it is also misunderstood. We recently entered into a three-and-a-half billion dollar contract with them for cloud services for their own internal R&D use cases. That's real compute they need and are procuring from us, so it isn't circular in any meaningful sense.
Microsoft and NVIDIA make up most of your contracted ARR. What does the third leg look like in 2027, and where does the Mirantis acquisition fit?
Microsoft and NVIDIA are our two largest single customers, but our mix is already close to 75 to 25, with the remaining quarter being enterprises, AI labs, and aggregators selling to enterprises. So we already service a meaningful share of non-investment-grade customers, and we're continuing to see further demand there. Most enterprises are adopting AI in some way, whether internal use cases like optimising manufacturing or drug R&D, or external customer-facing products, and those real-world use cases are driving more and more demand for inference. The hyperscalers are effectively playing a middleman role, on-selling much of the compute they procure to end users, and we naturally want those end-user relationships to be direct, so we expect that 25 percent proportion to increase over time.
The Mirantis acquisition is absolutely part of that. They've been operating CPU-based private clouds for decades across a customer base of more than a thousand enterprises, so they bring a wealth of knowledge in software development, monitoring, faster system startup, and enterprise and customer support. That brings immediate value, but on top of it they're developing their own software stack that makes it easier to directly target enterprises over time, which was certainly part of the thinking behind the deal.
What's the binding constraint to keep on upscaling capacity?
Our biggest constraint as a business is how quickly we can build the data center capacity. There is no shortage of demand. We're constrained by the amount of labour we can get out to sites and the timing of equipment. What we're spending a lot of time on internally is how to improve those bottlenecks, particularly labour at the site level. We're doing a lot of work on modularisation and pre-assembly offsite, doing as much as possible in controlled manufacturing facilities rather than at the site. That is our number one constraint, we cannot build capacity fast enough to meet demand today.
The bubble question is a popular preoccupation. Do you see any scenario where you overbuild and the demand isn't there?
It's very hard to see in the near to medium term how that could happen. As I said, there's this widening wedge between demand and supply, where it's getting incrementally harder to bring on supply while demand keeps going up every single day, so it's hard to see what changes. The real-world use cases coming through give particular support to that position. If you rewind even 18 to 24 months, the vast majority of demand came from a few very large AI labs training big models, but today a lot of the demand is real-world business applications. For the short to medium term, we think that demand-supply imbalance is only going to exacerbate.
In the opposite case, where supply can't keep up with demand, it will be a constraint. People are creative, so over time you'd expect different solutions like space-based data centres, which I think are particularly challenging, but it's an example. People will find creative ways to bring supply online, but in the interim there could be a massive dislocation, and that, coming back to the founding thesis, is exactly what we were set up to address.
You run US and Canadian infrastructure from Australia. What's the operating model that makes that work?
We've always had operations in multiple places. The company was founded in Australia but all of our operations are in North America, so we were well set up to accommodate this because we've done it from day one. With our recent acquisition of Nostrum, a data center development platform in Spain, we were responding to a distinct set of customer requests and growing demand within Europe, whether sovereign AI or the AI labs coming out of London and Paris, which are both major hubs today.
The opportunity was clearly there, and the way we targeted it was buying a development platform with a 60-person team already on the ground. We take that local team and overlay our existing risk management and funding capability at a global level, which lets us proceed very quickly. That's one of the advantages of having always been an international company. When we do expand, it's probably a little easier than it is for some others.
