Dylan Patel's Podcast and IREN

Investment Analysis Clubs Saul’s Investing Discussions
1

For those who haven’t had a chance to listen to Dylan Patel’s podcast on neoclouds, I highly recommend it — especially if you’re investing in IREN, NBIS, or HIVE.
:headphone: Podcast link

From minute 10 to 18, Dylan discusses the rise of neoclouds — purpose-built cloud platforms optimized for AI training and inference, rather than traditional web workloads.

:magnifying_glass_tilted_left: Dylan’s View: Neoclouds and the Next AI Infrastructure Wave

Dylan sees neoclouds as the next era of cloud infrastructure, driven by the intense demands of AI: compute-heavy training, inference at scale, low-latency networking, and new regulatory requirements. These needs are straining traditional cloud architectures and creating room for new players to emerge.

But the opportunity is not without risk.

:warning: Why Some Small Neoclouds May Go Bankrupt

Dylan points out that many small neocloud providers are taking on significant capital costs to build out capacity — investing in high-end GPUs, networking, power, and cooling infrastructure — often before they have guaranteed demand.

“If these smaller neoclouds don’t have their capacity fully subscribed — and at the right price — the economics just don’t work.”
“You’re making huge capex bets. If you can’t fill it with high-value workloads, or if you underprice to win customers, you burn cash fast.”
“It’s not enough to just build the hardware. You need a full stack that can extract value efficiently — otherwise you risk bankruptcy.”

The takeaway: the business model only works with high utilization and disciplined pricing. Without both, the fixed costs can overwhelm the company — especially for smaller players without the scale or balance sheet of the hyperscalers.

:white_check_mark: Why I’m More Confident in IREN After This

  1. Not all small neoclouds will go bankrupt — IREN is a strong exception.
    Dylan’s warning makes sense, but IREN appears better positioned. It already has a profitable GPU-based business via Bitcoin mining, proving it can operate energy-intensive infrastructure efficiently. That operational track record gives IREN a head start in managing the economics of AI workloads.

  2. OpenAI’s risks don’t mean the whole AI sector is in trouble.
    There are growing concerns about OpenAI’s “circular” investment model — where it engages in complex financial relationships with suppliers and partners like Nvidia, Oracle, and AMD. If revenue doesn’t materialize to support these deals, OpenAI could face financial strain.
    However, even if OpenAI fails, it doesn’t mean other foundation model providers or the broader AI inference market will collapse. IREN, as an inference-focused infrastructure provider, would likely see continued demand — even if the market experiences short-term volatility.

  3. Inference growth creates space for neoclouds like IREN to scale.
    As AI inference workloads explode in size and scale, there’s ample room for new players — especially focused neoclouds like IREN — to grow and compete with the large incumbents. The market opportunity is growing faster than any single provider can serve.

Would love to hear what others think — especially if you’re following IREN or tracking developments in AI infrastructure. Dylan’s podcast gave me a lot more confidence in how this space is unfolding.

37 Likes
2

Isn’t he referring to software advantages here? In that regard, Nebius and Coreweave seem like the full stack players, always talking about software. IREN is “bare metal,” no?

To be fair, I am skeptical of them all…but the fact is, with all the money Microsoft and others have thrown at Coreweave and Nebius, no similar contract seems to have fallen in IREN’s lap yet, right?

Bear

16 Likes
3

Thanks - I closely follow both IREN and Dylan Patel. Just a note that I watched 3 videos this morning, so some of what I’m referencing might not be in this particular video.

But, this sentence got me:
“Many of them [the NeoClouds] can’t even install Slurm [a software package scheduling jobs on GPUs] for you.”

Which brings up my main concern with the so-called Neo-Clouds, that I’ve expressed here before - if they’re not providing any software value add on top of standing up an AI data center, how are they going to survive?

Patel himself points out that the NeoClouds are mostly funded by VCs (Venture Capitalists). That means they need to provide a high rate of possible return. But, he’s now seeing money coming from commercial real estate companies into the NeoClouds. They’re used to low returns (high teens at best) over a long period of time. The difference is VCs are willing to take on more risk than the commerical real estate folks - which is why they demand higher return rates.

As the NeoClouds sign more and more “bare metal providing” deals with MicroSoft, Oracle, Meta, etc. (bare metal means no value-added software on top), they are turning themselves into mere data center providers. Literally. That’s not a high growth business. The economics of these deals is that in 3-5 years when the lease is up, the NeoClouds will own the hardware and can rent it out to anyone at that point. But, who’s going to want to rent 5 year old GPUs? And those that do will want a huge discount. Worse yet, the compute per watt of new Nvidia chips gets better and better, so running older chips may not even be cost effective unless your power is really cheap.

And, as Patel points out, a data center isn’t just a building with power. Power density matters, as does cooling. So, often the new Nvidia servers go into old data centers, and they have to skip racks or even skip rows to both not suck up too much power, or not put too much hot air into the cooling system. And retrofitting for liquid cooling often means structural changes to the concrete building, reinforcing roof capacity, etc. I liked the mention of Meta using tents - they go up more quickly, cost less, and all the other stuff is more easily upgraded later.

Where does this leave IREN? Well, Patel talks about utilization. That the worst thing for these companies is to have GPU capacity that’s not being used and so not earning them anything. I guess that’s why so many of the NeoClouds are doing deals with the larger competitors - it guarantees them utilization and so they won’t lose money. But, with just renting bare metal I can’t believe their profit margins are so great - otherwise the Hyperscalers would just build it capacity themselves.

For IREN, I would think they could take any unused GPU capacity they have and simply mine Bitcoin with it. This also means that IREN can be picky about what deals they do - the income from being a NeoCloud needs to be better than what they can get rom Bitcoin mining on the same equipment. Intersting that Patel points out the CoreWeave already has so much demand that they only want to deal with large companies and/or do long-term deals.

I do wonder if the current supposed advantages of the NeoClouds - AI specific data center designs for power and cooling, plus server set-up in terms of networking, latency, up-time, etc. geared toward AI requirements rather than the Hyperscaler typical CPU tuning, database support, and up-time procedures will last. The Hyperscalers aren’t dumb and I suspect they are already establishing procedures more tuned to AI. Although Patel points out that they’re used to very high margins (especially Amazon with AWS), and they won’t get those margins on AI compute given all the competition.

Finally, Patel talked about AI model changes. This doesn’t affect IREN, but it’s interesting overall in the AI space. We keep hearing about ASICs and even new AI hardware like Cerebus coming to disrupt things. But, Patel says that as AI software is still not just rapidly changing, but changing dramatically, building ASICs is premature - you will end up with data centers that don’t support (or at least optimally support) your soon to be workflows.

He gave an example of memory requirements shooting up. So while companies like AMD put on more memory to gain some benchmark advantage, the benchmarks themselves rapidly became outdated compared to real world workflows, out-stripping the on-board memory and so removing that advantage since it all didn’t fit anymore.

And that this is likely to affect Cerebus as well. Cerebus is literally just a really big chip. Typically, chips are cut out from a wafer. A wafer is like a 12" diameter thin silicon slice. Instead of cutting chips out from a wafer, Cerefus designed a giant chip that takes up the whole wafer, with the idea being that instead of chip to chip networking, you just have core to core communication on the chip, which is faster. But, now you have to have redundant cores on the wafer (since no-one can mass produce a 100% intact wafer), and a bunch of other engineering challenges.

At any rate, once your workflow is bigger than a Cerebus wafer can handle, you’re not gaining much, if any, performance advantage. And, you’ve now, like with ASICs, lost out of being general purpose for development and other purposes.

31 Likes
4

At the risk of redundancy, the contracts Microsoft and others have signed with Coreweave and Nebius (among others) almost certainly do not involve Coreweare or Nebius software. They are bare metal deals. Microsoft wants to run its Azure software stack and sell that service to its customers.

As for IREN, that it’s optimizing for bare metal deals could make it more profitable, and that it could simply mine Bitcoin with any unused capacity seems like good thing. Although I should note that I haven’t heard from IREN that that’s their plan.

21 Likes
5

You’ve made a valid point about the software advantage of companies like Nebius (NBIS) and CoreWeave (CRWV). Especially, he mentioned that NBIS is providing a very good software for the use of HPC, comparing with a lot of new small neoclouds. Like @Smorgasbord1 quoted: “Many of them [the NeoClouds] can’t even install Slurm [a software package scheduling jobs on GPUs] for you.”

However, this observation strengthens my conviction in IREN from a different perspective, in the way that IREN has been in HPC business for some time, quite before this AI boom in the last several months. This practical experience suggests that IREN is less likely to encounter fundamental operational hurdles, such as the “no Slurm” issue, compared to newer entrants with less infrastructure experience. This distinguishes IREN from other companies transitioning from Bitcoin mining to HPC.

21 Likes
6

So, it’s now ok to post about Bitcoin related companies? (Asking for a friend)

3 Likes
7

It has always been OK to post about high growth companies, in fact, that’s esentially what this board is about.

There is an enormous difference between a company that employs crypto as a part of its business strategy and crypto as an investment vehicle.

That really shouldn’t require an explanation.

10 Likes
8

Brittlerock - Last time I asked a similar question, I did not receive as polite of an answer. And my posts were apparently deleted. But I appreciate your current reply.

FWIW, Mike Alfred is a good follow on X related to IREN and CIFR. He is not serious with all of his posts, so just realize that you need to filter what you read from him. But he is on the board of IREN and has been sharing thoughts on it since it was sub $2. He hosts a lot of X Spaces, but does not record them. So if you want to listen you need to be watching for it. He has also been on a few recent podcasts. The main things he personally likes about IREN are the quality of the management and company execution, and the access to power contracts. They also have the lowest cost to mine Bitcoin of the major miners, so that is a big advantage as they transition some capacity from mining to AI. But currently over 95% of their revenue is from Bitcoin - so, if you don’t believe in the long term potential for Bitcoin, then that’s a big risk.

Long IREN

11 Likes
9

I think you may be misunderstanding me.

The Slurm issue is something that a AI cloud service should be able to provide as a basic feature. It’s far below the kind of software that the NeoClouds claim to be providing to their AI developing companies. Not being able to install Slurm is like an old-fashioned cloud provider not being able to install MySQL (a free database program) for you. If they can’t install Slurm, then they are not providing ANY useful software and so they must be just “bare metal” hardware providers.

Now, that’s useful in its own regard: Setting up a data center with power, internet connectivity, having servers installed and running and being maintained/serviced, etc. is all very useful and companies make wish to hire that out to either get up and running faster and/or not have to expend all that Capex up front.

But, it’s not the kind of AI-specific software advantages the Neo-Clouds are insisting they provide. And that the biggest Neo-Clouds (for instance CoreWeave and Nebius) are apparently (and in Nebius’ case predominantly) providing just bare metal to legacy cloud providers (their competition!) is, I think, telling.

As for other Bitcoin miners switching to HPC providing, I haven’t paid much attention. Setting up servers for Bitcoin mining isn’t exactly the same as for general purpose High Performance Computing. Bitcoin mining can be done on custom ASICs, for instance, which are chips designed to optimize Bitcoin mining compute, but are probably lousy at many other kinds of compute. And even if they were using Nvidia GPUs, what’s their setup for networking, data storage, balance between GPU and CPU, etc.? Not to mention energy costs, ability and cost of providing SLAs (Service Level Agreements), customer support, etc.?

That IREN doesn’t pretend to do anything other than what it’s doing is good. Whether what it’ll be doing will be profitable enough remains to be proven, though.

12 Likes
10

Hi All,

Not sure if its been mentioned already but there is a great weekly Spaces podcast on X and its only about IREN.

As mentioned in the podcast, it’s pretty exciting - IREN moved their earnings up so it might be because it gives them a better chance to be included in the QQQ or maybe there is a new deal to announce. :slight_smile: Co-locators don’t have anywhere near the advantages of IREN and also IREN has been doing AI/hpc for a while so I’m very bullish.

I’ve got a 13% position in IREN.

Cheers,

Jeff

18 Likes