Dear board members, after listening to the earning call, I have a question that I hope to get insights from the $IREN bulls here.
First of all, congrats on a great quarter! The projected end of 2026 ARR is now officially $3.4B, which is a huge jump up! And the management demonstrated that, the deal with Microsoft, even excluding an internally added colocation cost, can be quite profitable.
However, one thing I was trying to wrap my head around is that, in order to achieve this ARR target, $IREN only needs to deploy 460MW power capacity, while $IREN has 2.9GW contracted power with 1.4GW from the Sweetwater 1 site targeting energization by April 2026. My question is, why wouldn’t $IREN have targeted even higher end of year ARR, with this additional power capacity. Does it take more than half year to start serving GPUs after a data center is energized? Or is $IREN limiting their speed of expansion intentionally so that they have less debts? Or is there any other potential constraint that I’m missing?
Good question there, I do not think you are missing anything there with your analysis. It is just the 2.9 GW is a lot of capacity as it sounds like the 9.4B contract with Microsoft will use 200 MW at the Childress facility. I think the bottlenecks are how fast they can acquire the machines from NVIDIA and also how far they want to reach for additional financing. The 20% upfront payment is a good deal for them.
I was concerned from the other side initially that the company may have over-contracted to get 2.9 GW this soon in capacity, before they had signed any big names hyperscalers or otherwise. This announcement with Microsoft and the details in the conference call have helped alleviate that concern as it does not seem like they should have trouble ramping up and filling out the remaining capacity they have. Additionally, they said on the last call they do not envision stopping building data centers and it sounds like the 2.9 GW will not be an ending point. I expect it may take 3+ years though to fill out all 2.9 GW.
Another point to keep in mind is the Bitcoin mining for with 50 exahash uses 630 MW I believe. I’m guessing these machines may be moved around using the low quality power wattage requirements that are available. It was interesting that Bitcoin was not mentioned a single time in this latest earnings call as it seems to no longer be the focus.
I am pleased with how the transition to HPC has gone so far. There were definitely some real concerns there how they would make this transition, and if their IREN Cloud or other associated services with HPC would work out. A lot of the criticisms I saw regarding the stock were if they could actually sell HPC because they did not have customer service requirements when running Bitcoin capacity. I believe this critique is basically null and void now with signing on Microsoft. I will point out it is pretty curious that Microsoft has signed with Nebius and CoreWeave too, but this leaves me to wonder why it is only Microsoft looking to get in this NeoCloud game early.
Just doing some math on the numbers of about a ~10B contract for 200MW of capacity. This seems to indicate if they fill out at 2.9 GW at those same rates, it would be about 150B worth of similar contracts on those terms. The leadership seems to know what they are doing pretty well with scaling up data centers. By the time they are approaching filling out 2.9 GW a few years from now, I’m assuming they will have some expansions plans underway.
One thing I’ve heard is that the power required to service AI workloads fluctuates. For example, demand for inference is less at night when people are asleep and (hopefully) not posing queries. Bitcoin miner ASICS, which are Tier 1 workloads (don’t require the same level of redundancy as Tier 3) can consume unused power on an intermittent basis as available. Since the availablity of power is one of the main pinch points in HPC, it follows that utilizing ASICS to consume unused power makes a lot of sense. ASICS aren’t the main course anymore, but still have a role to play as side dishes.
I don’t think Bitcoin is going to be as big of a thing as you think it is.
I watched a video of HIVE CEO he said a GPU mining Bitcoin makes 12 cents an hour, where a GPU doing AI makes 1-2 dollars an hour.
So your looking at 5-10% of the AI revenue for Bitcoin. That 12 cents is on a machine designed from the ground up to mine Bitcoin. Your Nvidia AI GPU has lots of parts that are essentially worthless at Bitcoin mining (example memory).
Now the Nvidia GPU is more expensive than the ASCI miners. Your not going to want to waste Nvidia GPUs run hours on Bitcoin. You would rather workload balance at night to extend lifespan of the GPUs.
So now if you want to use excessive power, you can build up different portion of your data centers with Bitcoin mining rigs. But if your goal is AI it seems like you would slow down your AI build out by buying Bitcoin rigs.
Those were my thoughts as well. I was more leaning towards the second reason there - they wanted to control the size of leverage. But, no matter that the reason is, if it indeed takes several years for $IREN to fully monetize the 2.9GW capacity, I would worry that they may lose some of their advantage of having power early. Although many analysts forecast a power shortage until 2029 or 2030, I would not bet on too far away in the future. I believe 2026 will be the most power-constraint year thus is the most important year for $IREN to monetize its power. In 2027 and 2028, many other competitors (e.g. $NBIS, $CLSK, $CIFR, etc) will catch up on power as the projects they established right now will land in that time range. I don’t know what the competitive dynamic will be in that time.
Looking at it from a different angle, when we compared $IREN with $NBIS, we thought the ability to land power early was a huge advantage of $IREN. But, if $IREN ends up 2026 with just 460MW power monetized, then, it might not be actually faster than $NBIS there.
BTW, $IREN did not seem to emphasize in the earning call, but I believe they officially disclosed in their presentation slides that they have multi-GW more power in the pipeline, so at least 2GW more power beyond the 2.9GW contracted power.
The average timeline to build a new power plant is 3 to 6 years. We are not seeing new power plants being built as fast as we are AI data centers for several reasons.
New power plants are not getting commitments yet before construction, so it’s limiting capital investment into it.
Power plants tend to require much longer Permitting processes.
Supply chain limitation for parts like turbines for power plants.
Lots of political resistance for different types of power generation.
So I expect it will be on the longer side of when power stops being the bottleneck.
One other thing to watch here is the rate IREN is able to convert their GPUs from generating BTC revenue to AI revenue. Believe it or not, AI revenue as a percentage of total actually decreased this Q.
That conversion rate will directly correlate to how quickly projected ARR becomes actual revenue. I also couldn’t help but notice the disclaimer used not once but twice in the footnotes concerning ARR:
”It is not fully contracted, there can be no assurance that it will be achieved, and actual revenue may differ materially. Assumes on time delivery and commissioning of GPUs.”
There is no doubting IREN’s opportunity, but there’s also a ton of execution needed to get from here to there. That ratio of BTC to AI dollars should be a useful metric in monitoring the transition as we go.
