Frans Bakker@FransBakker9812
“The most important is the narrative violation on time to compute.”
This is your choice to make the most important, the sentence in itself is predetermined, not objective on both counts.
“It turns out that if you don’t have a lot of experience running a large scale cloud you run into delays. IREN is amazing in constructing the powered shell, but clearly has more difficulty in bringing GPU’s online. Hence the delays. Also because they rely on third party server installs from Dell & co.”
Running a large scale cloud is happening after the GPUs are installed and commissioned, a “large scale cloud” implies coherent working GPUs connected by infiniband or similar, delays don’t happen ahead of running a large scale cloud while commissioning GPUs, but during the process of receiving them from an OEM that is chip constraint. You conclude that experience in running was because of the difficulty of bringing GPUs online “hence the delays”. This completely acknowledges, and literally omits the delivery delays from Dell. But you do want to blame Dell to emphasize “server installs”, I wonder why the servers are being included here, server installs are the absolute easiest part of the job. But I know the answer, and you bring it forward soon enough (hint: Nebius).
“Next the power narrative requires a reset with moving SW completely to 2027. Smart move to get VR’s, but a violation of the power now narrative. For me its quite clear that they cannot bring the capacity online fast enough to let this be a heavy advantage versus other neoclouds.”
You front load your take with a conclusion that you do not deserve to draw. First of all, the “time to compute” narrative does not mean the fastest time to install any chip. It is the fastest time to the highest possible token/MW within a certain window of time that matches customer demand within that window.
Let’s turn this take around, if $IREN could have stood capacity up sooner, and would have data centers ready now in Sweetwater, this would have been a 2025 deal with 2025 economics, and for GB300s. As you know, the 2025 economics weren’t great (well you pretend not to know because you believe Nebius got a great deal without knowing the unit economics), and the GB300 isn’t a very efficient product on a token/MW basis.
What IREN instead did, was build out the high voltage step-down infrastructure to the tune of 750MW worth of primary substations and a bulk substation all tier III redundant, and now have a blank canvas to sell 2027 VR200 capacity with guaranteed chip delivery from Nividia.
Would IREN have opted for GB300 and a 2025 “backlog” for 2026, we would all be crying now because the deal economics would probably not be much better than the $MSFT deal. So instead of being an early mover that goes deep to bring current generation compute online, they chose to go wide instead, and can bring multiples more of next generation compute online a year later.
I think you have to realize that this was not a lack of ”ability bringing capacity online fast enough”, but a strategic decision to maximize the token/MW out of their flagship site, by conceding a slower ramp up in initial compute, but catching up by a “go wide” strategy with VR200s into 2027.
You can call it what you want, but it certainly is not a shortfall in the "ability to bring capacity online". If you want to quantify “speed of bringing capacity online” you cannot dismiss a strategy that was opted to promote the tokens/MW and the $/MW output, albeit later.
Time to compute is not a static slogan that you can apply to fit your narrative. Time to compute is a flexible concept that does not mean a race to indiscriminately bring as much single chips online from now until the end of the site’s capacity. Therefore, there is no “violation” of “power now”. Because that was never a thing in the first place.
“Finally it is also clear that even if you design your datacenters for HPC, the organization was not. The Mirantis deal was really neccesary to bring in mature cloud experience and enterprise sales support”
“The organization was not ready for HPC”
Here you go again steamrolling over the fact that there is a wide variety in what you can label as HPC.
IREN has been running their AI cloud for almost 2 years, be it in a very small GPU quantity. The organization has been slowly adapting to this over the course of the last 2 years, and prior to Mirantis, IREN was able to contract Microsoft, contract $500m of AI ARR, and grow their AI revenue from <$1M per Q to $33.6m per Q.
IREN has always been ready for AI cloud, which is a form of HPC, you need to call the activity what it is, and using a general term like HPC is purposely misleading because it gives the idea that there is only Bitcoin mining, and HPC.
Mature cloud experience, Enterprise sales support, are all words that fit Nebius very well, and yes Mirantis will certainly bring that experience to IREN as well. But nothing from the earnings call screams that IREN needed Mirantis to run AI cloud, let alone HPC. IREN has acquired Mirantis to complement their AI cloud offerings, and connect to a growing TAM of enterprise customers and tap into the sovereign market.
Just because IREN decided to do this now, doesn’t mean that everything they did prior to the acquisition should be dismissed as “not ready for HPC”. That’s just disingenuous from you.
Finally you say this:
“but the execution premium in the valuation versus Nebius for me makes more sense now”
This implies that IREN has no execution premium? What exactly has Nebius actually delivered?
I agree with you that the valuation of Nebius is rich, and IREN is certainly being discounted. But the bulk of the valuation disparity is because Nebius has agreed to sign more deals for 2027 and beyond. Where IREN has not signed anything for 2027, outside of a 3.4B deal with Nvidia for early 2027.
So is Nebius getting an execution premium, or are they just being rewarded for a higher RPO and backlog?
You have applied a lot of bias in your take. I have taken the time to give you a rebuttal that carries my personal view on IREN. I really don’t understand how you could ever be in my subscriber group, and still have such bad takes man.
It’s a waste of time to try and bring this type of material out there, because you honestly have no idea what you’re talking about.
I have talked to IR at IREN and the delays were for the most part due to the GPUs being delayed. ‼️‼️
Why don’t you compliment IREN on their partnership with Nvidia where they get on par with Nebius in GPU priority? This literally solves the issue that you have wrongfully described in your post.
If I can give you some advice, try to be a bit more objective next time, and consider that there are many reasons why revenue can be delayed, and by the same token, why deals can be signed later.
