Blake Coolidge

Okay, here's my shot at two bold predictions for $IREN this year @mikealfred: 1) they will not sign another hyperscaler or AI lab deal (unless it's a $MSFT expansion of the current deal). 2) their next major surprise will be a government-backed sovereign AI factory, combining NVIDIA architecture, IREN infrastructure, and Mirantis’ secure orchestration layer for public-sector, national lab, defense-adjacent, or sovereign AI workloads. I've tried my best to connect the dots between Jensen's comments in NVIDIA's most recent earnings call, the NVIDIA & IREN partnership, the Mirantis acquisition, the market trends toward sovereign AI & open-source models, and the tremendous community research & support from folks like @PhadsEth, @FransBakker9812, @jiahanjimliu, @bitcoinbutcher1, @ilzmcfly, etc. Over the next 5–10 years, IREN is trying to become the vertically integrated AI factory operator for the part of the market that cannot, or will not, rent its AI future entirely from the hyperscalers. AI is becoming the brain of the enterprise (core workflows, decision systems, customer experiences, research processes, and operating models). And there is an old saying in enterprise IT: "Don’t rent your brain." The first phase of AI infrastructure was about getting as many GPUs online as possible. That phase rewarded whoever could access chips, power, and data center capacity quickly. But the enduring business is about turning GPUs, power, networking, storage, orchestration, software, security, and governance into usable AI factories. Those factories can now be built with three key puzzle pieces: NVIDIA brings the AI factory architecture. It brings the systems, networking, software, and credibility required for enterprises and governments to trust the platform. IREN brings the physical infrastructure: power, land, data centers, GPU deployment, and operating expertise. Mirantis brings the cloud-native software layer: Kubernetes, OpenStack, orchestration, enterprise customers, and open-source infrastructure credibility. Of course, the hyperscalers will remain huge. AWS, Azure, and Google are not going away. But there is a much broader market forming outside of them: enterprises, AI-native companies, governments, regulated industries, industrial companies, research labs, pharma, financial services, healthcare, energy, defense-adjacent customers, and sovereign AI buyers. Those customers have different needs when it comes to AI. Some need dedicated infrastructure. Some need data residency. Some need confidential computing. Some need control over networking, storage, and telemetry. Some need to avoid lock-in. Some do not want the future operating intelligence of their company running entirely inside a shared public cloud controlled by another strategic technology giant. That is where sovereign AI breaks down into a few categories: 1) National sovereignty, where countries want domestic AI infrastructure. 2) Enterprise sovereignty, where Fortune 500 companies want control over their most important AI systems. 3) Industry sovereignty, where banks, pharma companies, healthcare organizations, industrial manufacturers, and energy companies need infrastructure built around regulation, privacy, and mission-critical operations. 4) Model sovereignty, where customers want the flexibility to use open-source models, proprietary models, fine-tuned models, and domain-specific models. 5) Infrastructure sovereignty, where customers want to avoid being trapped inside one provider’s ecosystem forever. To serve these various customer needs, IREN can offer a vertically integrated AI factory platform with NVIDIA-grade infrastructure, industrial-scale power, enterprise cloud orchestration, and open-source flexibility. Inference is bursty. Workloads are diverse. Hardware generations will be mixed. Models will vary. Traditional compute will be pulled into the same environment. Agents will need to call models, databases, APIs, applications, and internal systems. The AI factory becomes a full compute platform built in anticipation of this diversity. And, the open-source angle is critically important. Open source reduces lock-in. That matters because the biggest enterprise and government buyers will not want their AI operating layer controlled by a closed proprietary stack. They want flexibility, standards, and optionality. If you are building tens of billions of dollars of physical infrastructure, you do not want a proprietary software vendor sitting above you with the ability to tax the customer relationship or limit flexibility. Open-source infrastructure gives IREN a way to say: You can run NVIDIA-aligned AI infrastructure at scale, without locking the future of your company into a single closed ecosystem. It also explains why NVIDIA would care. NVIDIA’s challenge is no longer selling GPUs. Every GPU they make will get sold. The bigger challenge is expanding NVIDIA’s AI factory architecture beyond five or six hyperscalers and into the much broader enterprise, sovereign, industrial, and AI-native market. That go-to-market is far more complex. And explains the massive marketing & brand awareness push. IREN is pushing for >90% brand awareness for anyone involved in the AI space. The hyperscalers are easy to identify. The rest of the world is hundreds of thousands of potential customers, each with different industry requirements, compliance needs, software environments, and deployment models. The plan is: 1) Secure massive power. 2) Build AI-optimized data centers. 3) Deploy NVIDIA-aligned AI factory architecture. 4) Layer in Mirantis orchestration. 5) Support enterprise, sovereign, AI-native, and regulated customers. 6) Repeat. To serve thousands of customers in the middle market. If they execute, IREN becomes something very different from what the market originally thought it owned: a vertically integrated AI factory operator for the sovereign, enterprise, and AI-native world. That is the ambition I think they are chasing. And if AI truly becomes the operating brain of every major company and country, infrastructure sovereignty becomes one of the most important themes of the next decade.

Something to consider… what’s the downside of $IREN here? If the AI Cloud bet is a complete failure, and they resorted to signing away their portfolio for colocation deals at the same terms as $CIFR or $WULF, you end up with a $50B - $100B business… I don’t see any of that happening, of course. The IREN team has run the numbers, calculated the risk, and come to the conclusion they can drive many multiples of that. But this would be a horrible time to sell.

@FransBakker9812 @Hak84026661 @bitcoinbutcher1 NVIDIA & IREN will continue to support model-agnostic infrastructure for sophisticated enterprise customers who bring their own model, choose their own model, or run many models. Certainly inclusive of Claude, Opus, etc.

@Hak84026661 @bitcoinbutcher1 I think this is more of a @BlakeACoolidge question

NVIDIA DSX System: IREN's Commercial Potential Under the DSX AI Factory Model (Part 3) Before discussing this topic, I want to first analyze IREN's position within the DSX ecosystem. According to NVIDIA's own description, under the DSX framework, it is collaborating with eight AI infrastructure companies. Besides IREN, the other seven are CoreWeave, Crusoe, Nscale, Lambda, Nebius, Firmus, and Yotta. After reviewing the publicly available information regarding each partnership, the collaboration details can be summarized as follows: CoreWeave — Uses NVIDIA DSX Air to build and test AI Factory digital twins in the cloud, allowing operational simulations before physical equipment is delivered and shortening validation cycles. It is also one of the cloud partners deploying DSX platform components, including DSX Sim, MaxLPS, and DSX OS. Crusoe — Since March 2026, it has expanded its full-stack collaboration with NVIDIA, adopting the Vera Rubin DSX reference design and Omniverse DSX Blueprint to plan and operate gigawatt-scale AI factories. The partnership covers digital twins, AI-driven power and cooling optimization, and mechanical and electrical engineering design. Based on current information, Crusoe is the only company besides IREN that has a dedicated public announcement regarding gigawatt-scale physical engineering collaboration. However, NVIDIA has not designated Crusoe as a "flagship" partner. Nscale — Not only is it one of the co-developers of the Vera Rubin DSX reference design itself, alongside engineering and simulation companies such as Cadence, Eaton, Jacobs, Schneider Electric, Siemens, and Vertiv, but it is also one of the eight cloud partners deploying DSX software platform components. Therefore, Nscale appears to have a certain level of involvement on the engineering design side. Firmus, Lambda, Nebius, and Yotta Data Services — Public information on these four companies largely remains limited to a common press release description: as NVIDIA cloud partners, they deploy the core DSX platform components, including DSX Sim, DSX MaxLPS, and DSX OS, to reduce risk, improve GPU utilization, and accelerate infrastructure deployment. No dedicated announcements have been found indicating deeper physical engineering collaboration. Among these seven companies, only Crusoe appears to have engineering-related cooperation. The rest are primarily involved in specific software-layer integrations, deployments, or operational applications. By comparison, NVIDIA's collaboration with IREN demonstrates several unique characteristics in both depth and breadth. First, logical inference suggests that IREN was likely involved during the initial planning and design stages. The engineering component represents the overwhelming majority of the DSX system. A design framework of this scale must be built around a real, physically deployable site. Without a concrete location, it is impossible to create a truly executable project plan. Engineers must repeatedly validate boundary conditions such as power capacity, land availability, and cooling requirements against an actual site. This is basic common sense in complex systems engineering. In November 2024, equipment procurement had already begun for IREN's 1.4GW Sweetwater campus in Texas. A system design project as large as DSX spans multiple engineering disciplines and requires extensive feasibility verification against real-world infrastructure. Such a reference architecture would likely require at least one to one-and-a-half years of preparation. By the time DSX was officially announced in March 2026, the overall framework was already largely complete, and the project had reached the stage where a flagship AI factory was ready for deployment. Working backward from that date, the timeline aligns closely. As the only known gigawatt-scale site in the United States over the past two years with a confirmed power energization schedule, Sweetwater has never publicly announced a specific monetization plan. To this day, no customer has been formally attached to the site. The only disclosed information is that it will serve as the flagship deployment site for the DSX AI Factory system. The other company involved in DSX engineering collaboration, Crusoe, appears to have a significantly different relationship with NVIDIA. While Crusoe also controls gigawatt-scale infrastructure, its sites already have clearly defined end users. The 1.2GW Abilene campus is leased to Oracle and OpenAI. Crusoe's business model with these customers is itself an innovative undertaking. Because these facilities already have committed users and contractual obligations, it seems unlikely that they would have been available from the outset as dedicated engineering testbeds for NVIDIA's DSX development efforts. As a result, NVIDIA's engineering collaboration with Crusoe is more likely focused on specific project phases or selected engineering disciplines. Another important distinction is that Crusoe is a leading representative of the Behind-the-Meter (BTM) power generation model. Different power architectures create different engineering requirements. Therefore, the NVIDIA-Crusoe collaboration may be more focused on BTM-specific technologies. BTM data center infrastructure remains a relatively new industry segment. Significant challenges still exist regarding reliability and regulatory compliance. Crusoe's history of project delays and cancellations reflects some of these realities. Second, IREN's $3.4 billion services agreement with NVIDIA is unique. Based on currently available information, no other DSX cloud partner has been disclosed as having a similar direct paid customer relationship with NVIDIA. NVIDIA is effectively outsourcing part of its internal AI research workloads to an external operator. This arrangement appears to be unique within the DSX partner ecosystem. This also indirectly demonstrates both the depth and duration of the relationship. The deployment location for this agreement is IREN's Childress campus. Childress combines high-capacity fiber connectivity, liquid cooling infrastructure, high-density rack deployments, and both air-cooled and liquid-cooled environments. Across virtually every infrastructure category, it represents a best-in-class configuration. It is also arguably the most complete single-site environment for designing and validating the supporting components required by the DSX AI Factory architecture. Again, this appears to be unique to IREN. Putting all these factors together, Sweetwater's long-standing strategic ambiguity, combined with the recent announcement identifying it as the flagship DSX AI Factory site, suggests that cooperation between NVIDIA and IREN likely began at the earliest design stages. NVIDIA's direct paid customer relationship with IREN is unique among all eight partners. The flagship AI Factory under the DSX framework is being co-developed with IREN. A flagship designation fundamentally implies uniqueness, exclusivity, and top priority. Its deeper significance is that IREN gains earlier access to NVIDIA's core DSX architectural concepts and technical standards. This has major implications for planning the future development of IREN's 5.8GW pipeline of secured power capacity. It also reinforces an important lesson in this industry: Being first is not what matters most. Getting it right is. Patience can create enormous value. So what ultimately defines success for the DSX system? In my view, it can be summarized in one sentence: Lowest cost, highest output. Jensen Huang has repeatedly expressed a similar objective: "Maximize token performance per megawatt at the lowest token cost." Anyone familiar with my long-term analysis of IREN will immediately recognize a striking alignment. $IREN's operational philosophy has always been: Lowest cost, highest output. For NVIDIA, $IREN may represent the ideal partner for achieving this objective. This helps explain why IREN alone has received flagship partner status. From a broader commercial perspective, NVIDIA will certainly not limit itself to working exclusively with $IREN. The DSX system will likely expand commercially across all seven remaining partners. However, the amount of value ultimately created by each partner will depend on its own business model and execution capabilities. NVIDIA undoubtedly hopes all partners succeed. Yet there remains only one ultimate metric: Lowest cost, highest output. Whether it is called an AI Factory or a Token Factory is largely irrelevant. If you achieve the lowest cost and highest output, you become the industry's dominant leader. Naturally, NVIDIA's DSX resources will gravitate toward the highest-performing operators. This is inevitable because it reinforces NVIDIA's leadership position across the broader AI ecosystem. Once the lowest-cost, highest-output model is achieved, IREN's commercial opportunity could change dramatically. Today, much of the market still evaluates AI infrastructure through an internet-era lens, assuming that software stacks are the primary value anchor of AI cloud businesses. In reality, software moats are gradually weakening. The future benchmark for AI cloud infrastructure may ultimately become: The lowest token cost per watt and the highest output per watt. To approach that standard, operators may adopt DSX. To reach the highest levels of performance, IREN could become the reference model. Achieving this, however, depends on an extremely difficult-to-replicate integrated system. Within that system, the decisive factor is not software alone. It is not power alone. It is not HBM memory. It is not optical networking. It is not any individual bottleneck currently receiving market attention. Rather, it is the combination of all these capabilities. The bottlenecks people focus on today are simply challenges that must be solved while building a much larger integrated capability. Many IREN investors still believe that the company's competitive advantage comes primarily from controlling scarce resources such as power and land. Some worry that if power shortages disappear after 2030, IREN's moat will disappear as well. In my view, that interpretation significantly underestimates the company. That is not what is happening. To achieve the goal of the lowest token cost per watt and the highest output per watt, IREN has already built a three-layer system consisting of: The energy layer The chip layer The infrastructure and software management layer Its deep collaboration with NVIDIA may eventually allow participation across the broader AI stack. Over time, IREN could gradually extend its influence toward the model layer and application layer. Direct participation appears less likely in the near term. However, strategic partnerships could provide indirect exposure and additional leverage opportunities. As a result, value creation could become much larger than currently appreciated. The market may eventually stop viewing IREN as simply an AI cloud company. Instead, it could be redefined as: The most important infrastructure platform company of the NVIDIA AI Factory era. If that happens, its commercial opportunity set could expand dramatically. First, IREN could provide DSX-standard technology licensing, facility engineering consulting, and managed operations services, generating high-margin fees and recurring operating profits. Rather than selling compute alone, it would sell complete AI Factory templates, standards, and operating frameworks. This could position IREN as a key partner in global Sovereign AI initiatives and place it at the center of the emerging AI infrastructure supply chain. Second, following IREN's acquisition of Mirantis, and with the support of DSX Flex, the company could connect directly to NVIDIA's DSX Exchange communications backbone. This would create seamless integration between the physical and software worlds. From substation switching and microgrid dispatching at the lowest infrastructure layer, all the way up to enterprise Kubernetes deployments at the application layer, a unified control chain could emerge spanning power systems, networking, GPUs, containers, and applications. Through extreme optimization, every watt of electricity could be converted into the maximum possible density of AI tokens. This is infrastructure-level software optimization rather than traditional application software. It represents a unique advantage for IREN. Under the same electricity cost structure, IREN could produce more AI tokens than competitors. This, in turn, could enable a unique toll-road-style economic model with substantially higher returns than peers. Third, flexible power arbitrage. This is already an area where IREN has extensive experience. Once DSX Flex becomes operational, IREN's advantages as a large-scale owner and controller of power infrastructure could be maximized. Fourth, becoming a key enabler of compute financialization and an important anchor asset within that ecosystem. This topic deserves a separate dedicated analysis in the future.

@franklee6924T Love this. Seems to be the key. Keep it coming

@FransBakker9812 @_Sgr_A_Star Do we know why / what happened?

@_Sgr_A_Star @BlakeACoolidge And the DC is immaterial. It was also downscaled to 1,000 GPUs ;)

Excited for the discussion. $IREN is already leading the Sovereign AI market: “IREN’s Prince George data center will be Canada’s first cluster with Dell PowerEdge servers with NVIDIA GB300 NVL72 AI infrastructure and NVIDIA Mission Control software. These advanced facilities, powered by the Dell AI Factory with NVIDIA, will allow users to train powerful AI models and run inference at scale while offering superior performance with built-in security. This collaboration helps to advance Canada’s Sovereign AI Compute Strategy, which aims to strengthen national AI capabilities by expanding secure, domestic compute capacity and ensuring that data and innovation remain under Canadian control.” Let’s continue across the globe! 🇺🇸🇪🇸🇦🇺 dell.com/en-us/dt/corpo…

@jiahanjimliu @FransBakker9812 @Agrippa_Inv Most likely. If not now, in the future It’s been very clear that Dan & team are hitching their wagons to Dell & NVIDIA as part of the AI Factory offering

Question for @FransBakker9812, @Agrippa_Inv - is $IREN using Dell Software for some customers? Ben is very credible and although he has $IREN rated lower than $CRWV and $NBIS by a lot, Dell is rated higher than all 3. Software fees for Dell is most likely much cheaper than Colocation fees, and favorable if software is superior. Overall, I think $IREN’s bread and butter customer should be FireworksAI and Anthropic that bring their own software stack

Helpful scorecard when evaluating the standalone offerings. $IREN, however, is not trying to solely position themselves to prospective customers. A more accurate comparison would be to look the NVIDIA AI factory offering, which is ripe with software capabilities provided by NVIDIA, DELL, and other partners.

$NBIS $CRWV $IREN Very interesting scorecard I will for now not comment that $NBIS x $CRWV are almost identical. I disagree there heavily, but the main takeaway is what I think everyone feels right now $IREN has the best setup in terms of data center “road map” but their technical capabilities are by far the worst That’s basically the whole thesis about $IREN. They are behind but they can catch up or at least close the gap in the following months/years. If they succeed, the stock is extremely undervalued here I will make a longer post about this scorecard because it is very detailed and I have a lot to say about it, but this is the first thing I wanted to adress when I saw it

@jukan05 Chip demand far outweighs supply. He’s not “threatening”. NVIDIA would and should naturally favor the companies leaning into them most

"Some neo-clouds worry that they can’t stray from buying Nvidia’s full stack of hardware for fear of being put in “Jensen jail,” meaning they might lose their allocations of Nvidia chips, said Adam Fisher, a partner at Bessemer Venture Partners." It seems Jensen Huang is effectively threatening neo-clouds implying that Nvidia could cut their chip allocations if they don’t buy Nvidia’s full stack.

Done writing novels for now, until there are more meaningful updates. For now, I'm holding both with patience, and would love to hear community feedback.

Why both can be major winners The ACIE opportunity is too large and too varied for one model to dominate. Sovereign AI clouds need local infrastructure. Enterprises need secure and controllable AI environments. Industrial companies need AI factories connected to real-world operations. AI-native startups need high-performance training and inference. Hyperscalers need partners that can bring new capacity online. NVIDIA needs an ecosystem of companies that can absorb GPUs, deploy DSX-aligned systems, and turn hardware into functioning AI factories. That market will not be served only by one kind of company. IREN can win by becoming one of the world’s best power-to-compute conversion platforms. If it executes, every incremental megawatt becomes a monetization option: AI cloud, sovereign AI, enterprise AI, managed inference, or future AI factory services. Its power portfolio becomes a strategic asset base that compounds as AI demand grows. Nebius can win by becoming the AI-native cloud platform for builders, enterprises, and regional markets that need more than raw capacity. If it executes, it becomes a new kind of AI hyperscaler: not necessarily bigger than AWS or Azure in general-purpose cloud, but purpose-built for the workloads that define the AI era. The bull case is not that one beats the other. The bull case is that NVIDIA has defined a new industrial category, demand is already overwhelming supply, and IREN and Nebius are two of the few public companies with credible paths to become foundational platforms in that category. IREN is building from the grid up. Nebius is building from the cloud down. IREN’s advantage is physical scarcity. Nebius’s advantage is AI-native usability. IREN is closest to the power bottleneck. Nebius is closest to the customer workflow. Both understand the endgame. Both are moving toward vertical integration. Both are aligned with NVIDIA’s need to scale the AI factory ecosystem beyond traditional hyperscale. And both are still early enough that the market may be underestimating how large the opportunity becomes if ACIE is not a niche, but the next major compute platform. If the next decade is about manufacturing intelligence, then the winners will be the companies that can industrialize the full system. IREN and Nebius are two different paths into that future. Both matter. Both will compound. Both are much earlier in the journey than the market understands. And, if they execute, both will win. A lot. I’ll end with another recent quote from Arkady: "When we talk about competition, we don't see a red ocean. We see a new market being created. The hyperscalers will remain dominant and continue to grow, but there is an opening. These new workloads, new users, and new requirements don't belong to anyone yet. We're not trying to take business away from the hyperscalers. We're building new business alongside them. This isn't a consumer market where one company wins everything. The opportunity is large enough for multiple winners."

IREN and Nebius 🧵 I know, I know. And?! Doesn’t it have to be OR?! You would think so, given all of the fighting back & forth on this platform. Personally, I’m long both. And plan to be for a while. I started my IREN position at $16.97 and my NBIS position at $53.87. And I believe both stocks are still Strong Buys, if you’re holding for 3+ years and believe what’s laid out below. I believe IREN, in particular, is a screaming buy right now because let’s start with a few quotes: “You need to have a pipeline of construction and know that the next gigawatts are coming.” “To build what we want to build, you have to be extremely efficient, and that means starting from scratch. You need to understand everything from land acquisition and power permits to data center construction, rack design, and the software that runs on top of it. If you outsource those layers, you give away both margin and control. We build across the entire stack. That gives us economic efficiency, but more importantly, engineering efficiency." “When we first started, we took whatever capacity we could get, including smaller colocation sites. They were more expensive, but speed mattered. We needed to get into the market quickly. As we've scaled, we've built more of our own capacity. Those aren't small deployments. They're large campuses measured in hundreds of megawatts, and increasingly in gigawatts." “We're building bare metal capacity for Microsoft AI lab. But it allows us to go to the banks and use this contract to finance construction of our own cloud cloud. So, we use these big bare metal contracts not as our main line of business, but as one of the ways to finance building our own cloud for the rest of the market.” These are all direct quotes from Arkady Volozh, CEO of Nebius, during his most recent interview with Accel. Why are these quotes so bullish for IREN too? Nebius is racing to build physical scale underneath an already sophisticated cloud platform, and their stock has been rewarded handsomely for that engineering prowess & early GPU access. IREN spent years assembling the hardest pieces of the AI factory stack before the market understood their value: power, land, grid interconnections, construction expertise, and operating infrastructure. And that bet will pay off.