I read somewhere that if we were to service all of the consumer inference demand and exclude energy required for AI-training, we're still at an enormous gigawatt shortfall which implies this supply / demand crunch persists longer than everyone thinks. This is a pretty interesting bottom up analysis: (drive.google.com/file/d/1n8W…). With stuff like this I'm beginning to view $IREN now not as a multiyear hold but a 5-10 year old. I think it's becoming increasingly dubious that enormous gigawatt shortfall we have to service AI demand will be able to serviced. $ORCL / $CRWV, for instance are probably going to lose a lot of customers in their backlog. If everyone including your mother (Hut8 + Bit Digital / White Fibre) can obtain colocation deals for their gigawatts, it means: - hyperscalers are so desperate for power that they would sign with companies with absolutely no track record of operational excellence - market is not discerning at all between good operators and bad operators (just like it didn't do so for bitcoin mining) What this means over time is that: - the bad operators are going to become very apparent over time - backlog from these operators will be back onto the market and will find their way to the best operators ($IREN, $CIFR, etc.) - market will panic in the short term such that $IREN's stock price will crater - market will understand that $IREN is amongst the best operators that has scale - $IREN achieves scale such that it can afford to finance with little debt and no dilution Capex heavy projects related to its IaaS platform while incumbents with no scale will struggle to compete due to massive, upfront CapEx - market understands that $IREN has probably one of the best site procurement teams (extremely underrated fact) in the world such that they are perceived not as an outsider that will jack up electricity prices, but a friendly counterparty that is beneficial to the community - intense re-rate of $IREN stock price bringing it up to a multi-hundred billion dollar valuation There's a world out there where this crunch extends beyond what everyone, even within this group believes which means: - US is unable to solve the gigawatt shortfall with much disdain from $NBIS bulls who assign too high of a probability that their software will outcompete tech giants like Databricks, AWS, GCP, Azure overnight and who massively underrate the difficulty of developing infrastructure, securing power + sites, building next generation AI Factories that have never been done before in the history of mankind such that the only ways these datacenters will work is that if you have liquid cooling to service rack densities at 80 kw+, 130 kw+ , 200 kw+, 300 kw+ over time. I'm pretty sure very few companies, if any, in the world have a datacenter that will be able to service rack densities of 200+ kw by end of Q1 2026, $IREN will have that. - energized power at multi-hundred megawatt scales is valuable such that you can continue to build greenfield datacenters without tearing down old ones which will turn out to be economically viable for inference - because of $IREN's vertical integration and its superiority at fostering relationships with communities, securing power, securing land, having energized power and land, developing infrastructure with impressive speed / efficiency (including securing long-lead items), developing AI Factories, it will outrun both incumbents + existing hyperscalers over a long-time horizon. There is a chance 2026 / 2027 $IREN obtains escape velocity such that it's status as a datacenter company of the AI Era will be unrivaled. This is obviously making a lot of assumptions and ignores things like investor willingness to fund AI companies with no clear path to profitability + the overall demand side picture continue to play out + many other assumptions, but what if this does play out? What is the likelihood of this happenning? No one can be 100% sure, but my view is that the probability of this awesome but ridiculous scenario playing out over time has increased.

I just want everyone to know that my company has contracted 45 GW+ of grid-connected capacity coming online. That is, after checking every bureaucratic box, sitting through the interconnection queue, working with the utility to upgrade high-voltage transmission lines, having the utility build a new substation, and building my own bulk substation, I'll have this power ready for you in 2045. So I'll just pay a gazillion dollars for Bloom Energy cells and air turbines. Wait, nevermind, lookings like I don't need those BE cells anymore. I won't be needed in Wyoming. But it's okay, my private market valuation is still 40 billion. Thank you for listening to my Ted Talk.

This whole $IREN vs $NBIS argument is actually pretty simple. If you literally believe $NBIS will become the next AWS which means: - it becomes a vertically integrated hyperscaler that not only provides next-gen compute GPU clusters but also services the long-tail of customers who just want to plug into an API for open source models and build applications on top of its platform Then yes, you should invest in Nebius, even at this price. I don't believe it has enough differentiation from AWS, GCP, Azure. And I fail to see why customers already locked into those contracts will want to take the humongous effort of migrating their entire tech stack to another vendor when open source models also exist on AWS, GCP, Azure. I also am doubtful a ton of value when migrate to watered-down open source when it's pretty clear all of the value today is largely captured by Claude and then GPT. Keep in mind that if Anthropic / Open AI wanted to distill their models for more efficiency, they could just ... do that. Also just look around in enterprise: - nobody who codes use open-source -> it's all Claude or GPT (Meta / Google all use Claude instead of their internal models) - in enterprise workflows it's really just claude. everywhere. - the solo-pretreneur might use open source but open source does dominate enterprise. Claude /GPT (mostly Claude) does. Why do you think Open AI / Anthropic are starting to hire all of those FDE's? Levels of customers: - layer 1: Hyperscalers - these are the googles, Microsofts, Amazons of the world who just want bare-metal - layer 2: AI Labs like anthropic + open AI, also just mostly want bare metal - layer 3: Jane Street, FireworksAI, TogetherAI, Tik Tok, Palantir -> Some want bare-metal-like performance with a managed Kubernetes/Slurm layer; some want full managed AI cloud primitives; some might just want pure bare-metal. (Here $CRWV > $NBIS > $IREN) - layer 4: larger enterprises from finance / tech companies / real estate like JP Morgan to Salesforce to DR Horton. Salesforce uses all of AWS, GCP, Azure. Specifically, it builds a number of applications on top of AWS Sagemaker. My view is that most of the value will come from layer 1 to layer 3 because the vast majority of the value from layer 4 will be captured by AWS, GCP, Azure. And since this is my view, $NBIS can be reduced to an infrastructure play. Seems reductive, but I largely think this will be true over the next few years. And if you view Neoclouds largely as an infrastructure play, then you're better off investing in $IREN. And because that is my view, no amount of genius mathematician programmer CEO, or mid-tier software engineers will really make a difference will lead to higher-quality infrastructure + faster data center development timelines. I'll also mention that the smartest engineers in the world don't go and say "I want to work for Nebius". They work at these places, in higher to lower talent density (all high): - anthropic -> today, everyone wants to work here - open AI - cursor - databricks - fireworks AI - google - meta - stripe - ramp - decagon - etc. It's actually pretty simple. You can say your a CEO is a genius at programming, you can say you have smart engineers, you can see Yandex is awesome, you can say you have cool subsidiaries, but it doesn't matter when you're engineers really aren't world-class and when that has nothing to do with your core-business: building real-world infrastructure to power AI workloads. What has happened so far is that Nebius has done a good job of plugging in GPU's for Microsoft, even if they haven't come from Vineland. I applaud them for that. At the same time I think it's sus this compute is not coming from Vineland. But that's the easy part of their portfolio. Someone tell me how they will deliver on their backlog when a site like Missouri doesn't even have an interconnect agreement there. Pretty sure this site will have to use BTM to bridge the gap as well. Someone tell me whether paying Data One for colocation and then paying Bloom Energy on top of Data One is a good or bad look. I'll also be candid here about $IREN: - they have had some growing pains - horizon 1 is by far the hardest part of their portfolio because they're improvising as they go - getting GPU's has been a bitch - they still have to prove they can scale AI revenue from a meager 33 million last quarter to billions over the course of the next few quarters - they still have to prove they can deliver and operate liquid cooled datacenters at scale It's not like $IREN is perfect, but because I have an edge in knowing how the Horizon development is going + company roots + ability to source long-lead items + the genuine differentiation it has with its huge portfolio, I see it as more de-risked from an infrastructure standpoint than Nebius. For all of these companies - Nebius, Coreweave, Iren, my view is that they should not be evaluated as software companies, but as infrastructure first companies. It follows that if you don't have an infrastructure first background, then you might win the early game, but you won't win the late game. Good luck.

Not that you need advice from me @mikealfred but if someone starts going on about an eminent bubble in AI and come to the conclusion efficiency gains leads to less compute overall, I would just use this example: - anthropic came out with Opus 4.5 / 4.6 in the early half of this year - this was followed by GPT 5.4 and 5.5 - during this time frame, we saw Anthropic have an incredible rise in ARR at a rate we've never seen before in history - both the chips used to serve these models are significantly more efficient and powerful than they were one year ago - this has been followed by token consumption exploding across the board for various enterprises, particularly enterprises involved knowledge work - Anthropic has had to throttle some of their enterprise plans because of their inability to serve demand which has lead to them losing some customers to Open AI - coincidentally, GPU prices have gone vertical during the same time period - hyperscalers and neoclouds across the board have insane RPO's and capex commitments year into the future Coincidence? I think not. If someone were to read this and somehow still draw the conclusion that model efficiency gains has lead to less overall consumption of compute, they can't be helped. This is jevon's paradox happening in real time.

No doubt having an orchestration layer ("software") is valuable. But try acquiring 500 MW + of power and see how much that'll cost you. How much did Coreweave try to acquire Core Scientific for? And these were just bitcoin mining sites. What does this say about what is more scarce and valuable? Why does it matter if you have all the "software" in the world when you have nowhere to plug in your GPU's? Why sell most of your services as bare-metal (i.e. software-less) if the professed software in the first place is so valuable? What cannot be produced in the real world, irrespective of how smart your management allegedly is?

I looked at the price action today. Since $NBIS is up 12%, and $IREN is only up 1.5%, I determined that $NBIS is the better company. Because of this relative out-performance, I also determined: - time to compute no longer matters - software will solve the power bottleneck - Dan Roberts is a bad CEO - Iren needs to sign deals now or the sky will fall - the company I invest in needs to protect the value of my call options Price action never lies. It's Game over.

How to identify a true market genius: - only buys calls - says “wen deal” daily - thinks price action is analysis - permanently upset at “manipulation” - communicates exclusively in “ASAP” - high signal if you invert it - says $NBIS > $IREN and calls it “obvious” - reads tea leaves (calls them charts) - says “long-term conviction” but checks price every 5 minutes

I've done my research on $IREN, and here's what you need to know: AI Compute demand has fully gone vertical. Anthropic doesn't have enough compute. H100, B200, A100, H200 spot prices are the highest they have ever been. Scaling laws have been upheld. Opus 4.6 + claude code does my job for me. AI is being embedded into every possible enterprise workflow. GB300 models are coming. Sweetwater I is completely de-risked with long-lead items on site. Execution Horizon 1-4 is progression is going beautifully. Canada sites can service inference in British Columbia and Seattle. But the shooting star reverse pikachu triangle tells me that $IREN is using all of its 6 billion dollar ATM at this moment to destroy shareholder value. So I'm bearish, and you should be too.

This is the $IREN thesis playing out in real-time. As models become both more capable and cheaper, demand will grow exponentially while real-world supply can barely scale linearly. What matters from here on out is who can actually execute in a time frame that matters, and who can actually capture the economics on the forward curve as the demand-supply constraint gets worse. Eventually, the market is going to figure out who's real and who's not.

Some important things to understand if you don't want to get shaken out while holding $IREN: 1. Jevon's Paradox: cheaper cost per token leads to astronomically more demand 2. Model Improvement - each step change in model improvement leads to solving the long-tail of edge cases. even if this is only a 1% improvement, going from 98% efficacy to 99.95% efficacy is arguably bigger than going from 0% efficacy to 75%. 3. Cloud Economics: Cloud economics are vastly superior to edge inference. Firstly because GPU's in a cloud will have superior utilization rates and have superior networking that allow for much better throughput. Imagine running Claude Opus 4.6 at 80 tokens per second on a rack of interconnected GB300's vs Claude Opus 4.6 on a mac studio (oh wait, you can't). 4. Elasticity: The relationship between cheaper cost and better models and usage is NOT linear. It is highly elastic. A 2x improvement in token cost or model does not lead to a 2x increase in usage. It leads to a 4-5-6-7-8-10-20-30x increase. 5. Ignore Deepseek: Deepseek, or any chinese model is generally noise from an investment standpoint. They can continuously distill weights from US models but if Chinese labs were in the lead, do you really think their models would still be open-source. Ask yourself: If you had a decisive lead in developing a military weapon of mass destruction, would you open source this know-how to adverse nations? If you think Anthropic, Open AI, and Google should open-source their weights, then maybe you should join the CCP or marry Eileen Gu. 6. Agents are Coming: outside of coding, agentic workflows have a massive TAM and is largely unsaturated. There is so much left to do in finance, PR, administration, science, math etc. 7. Time to Power: At some point, every investing goon that is chasing memory, copper, chips, etc. will realize that time-to-power is an insurmountable bottleneck. There are a ton of GB300's are still just sitting there that can't be plugged in. And people realize what this implies. Right? Right?? 8. K-shaped Economy: Average normie's quality of life may get worse and AI-exposed part of the stock market can still go boom. Welcome to the k-shaped economy and the era of AI. 9. Prisoner's Dilemma: Not only are American labs and major tech companies in an existential loop of -> if they don't invest then they lose relevance so they must invest to win. There are winner takes all dynamics here since any company that can develop a decisive model advantage can dramatically accelerate their lead. 10. National Security: If you believe in the national security if the US, then these companies have to be at the forefront of the AI revolution. It follows that if these companies are critical to national security, then in a way, they have to be propped up by the US government (they're too big to fail). 12. TAAM for Iaas >> Colocation: If you're an IaaS provider, and assuming you can prop up kubernetes + some observability environment, then your total pool of customers expands from being primarily Google, Microsoft, or Meta to everyone on the long-tail of successful tech companies. This can be Tik Tok, Palantir, Uber, databricks, X. They all need GPU's, and most of these companies don't want to purchase and run their own GPU's. This gives you more pricing power. 13. Exponentials: People really don't understand exponentials, and Michael Burry doesn't know what he's talking about, particularly when it comes to understanding throughput and performance tradeoffs when running Chinese models locally. Point is, we are only seeing the tip of the spear. We should really be applauding @danroberts0101 as we are now in a window where $IREN can monetize its megawatts at a higher price point. Can you say this if you're $CIFR or $WULF? With most of their megawatts already under long-term contract, they have limited near-term exposure to rising market rates and therefore little opportunity to reprice capacity at higher points along the forward curve. The success of a company like $IREN is really just downstream from demand, and I'm pretty sure everyone is underestimating demand. This whole thing is way bigger than any of us can imagine. Zoom out and ignore the noise.

The funny thing here is that if no company is willing to invest massively in research (which encompasses massive pre-train compute + RLHF + SFTL) there there are no SOTA models to begin with. Paradoxically, if there are no state of the art models to begin with, then Chinese models will have no models to distill off of and consequently, there models won't be as good as they are today. This whole situation is analogous to compounding pharmacies making GLP-1's. If there is no patent protecting pharmaceutical companies from investing heavily to develop revolutionary drugs like GLP-1's, then where are the incentives to develop technologies like this at all? Am I missing something here?