GroGuy @GroGuy2

@TheBigDegen Spain although I would prefer Belgium.

@Bella_91m7u 15

@Stella_5253 1.8

@W3_Brain 21

@Simi_2210_ 18

@tugce190333 26

@Daphne_lkm 1

@medoliyy 40

@MustafaCharts 7

@accounting_ds Trossard

@Kaizen_Investor Congrats. Hospital in Belgium I assume.

@MMatters22596 Bloom Energy

Who Is Building High-End AI Smart Factories?(I) 2025 is the first true year of AI development—fast and intense, yet deeply chaotic. Never before has demand reached such unimaginable levels of commercial scale, and never before have the physical world and the virtual digital world been so disconnected and mismatched. This contradiction has no immediate solution. It resembles a raging forest fire met with rescue water that is utterly ineffective. As 2025 draws to a close, I would summarize this year’s AI industry with a single word: chaos. The major contradictions currently facing the AI industry are: 1.Unlike the CPU era, the exponential growth in GPU compute density does not match the practical, physical growth constraints of data centers. 2.The exponential growth in compute demand is mismatched with the configuration and performance improvements of physical infrastructure; even exponential increases in GPU compute density cannot satisfy the growth in compute demand. 3.The “energy bottomless pit” of compute power is mismatched with—and in some cases directly conflicts with—global goals of “carbon neutrality / grid capacity.” 4.The “infinite marginal value” of AI applications conflicts with the “finite ROI (return on investment)” of the real world. Further elaboration of these four core contradictions reveals that, between the digital virtual world and the physical real world; between power systems designed for civilian use and their excessive occupation by high-performance computing; and between early movers who believe AI is omnipotent and therefore worth aggressive, cost-insensitive investment, and professional investment institutions with rigid requirements for capital returns—there exist three major dimensions of conflict that are difficult to reconcile in the medium to long term. The sharp decline in AI-related stocks over the past month has been primarily triggered by the third contradiction. No one is questioning the long-term prospects of AI. The real issue is how to do it in a way the physical world can accept. Once investment exceeds the tolerance of capital, severe volatility inevitably follows. Among the contradictions outlined above, data centers are the industry with the highest degree of linkage. They are not merely infrastructure. Across the early, intermediate, and advanced stages of AI development, data centers will consistently play a critical role. This is because computational energy efficiency is not determined by GPUs alone. Data centers with different performance characteristics and different physical infrastructure configurations can exhibit enormous differences in realized compute efficiency. Especially at the advanced stage of GPU cluster efficiency, the decisive factor is the technical capability of the data center that supports cluster operations. It can be said that, within the AI industry, data centers and traditional data centers are fundamentally different things—and data centers that serve large-scale and ultra–large-scale clusters are yet another category altogether. We may call them smart factories. The exponential growth in GPU unit compute density has caused the developmental ambitions of the virtual digital world to collide head-on with the physical constraints of the real world. Under the Hopper architecture, traditional data centers could still function. Under the Blackwell architecture, however, the vast majority of GPU models require entirely new infrastructure standards, effectively excluding almost all existing traditional data centers worldwide. Yes—starting with Blackwell, what is required is a new type of data center designed to match the development of the AI industry. Its entry-level requirements include: 1.120 kW high-density racks 2.Liquid-cooling thermal systems 3.High-load data center floors rated at ≥1.3 tons per rack 4.Support for NVLink / NVSwitch / PCIe 6.0 high-speed interconnects 5.High-density fiber and short-reach high-speed cabling capability 6.Overall infrastructure scalability and upgradability As GPU performance continues to advance—for example, after the emergence of the Rubin architecture—the requirements for data centers will rise to an even higher level and should include: 1.200–300 kW per-rack ultra–high-density racks 2.Liquid cooling systems (DLC / immersion cooling) 3.Ultra–high-load floors rated at ≥1.8–2 tons per rack 4.Suppo rt for NVLink 6 / NVSwitch 6 / PCIe 7.0–class high-speed interconnects 5.CPO (Co-Packaged Optics) and silicon photonics interconnect infrastructure 6.Rack-level AI Fabric support (low latency, ultra–high bandwidth) 7.High-density fiber trays and hybrid electro-optical cabling capability 8.Power and cooling redundancy re-architecture at the full–data-center level 9.Topology support for large-scale AI training clusters (ten-thousand-GPU class) 10. End-to-end energy efficiency optimization capability for 1–2 MW per row. I cannot verify the precise accuracy of each of these parameters, but they clearly illustrate that the requirements for the overall supporting infrastructure are continuously rising. In theory, such escalation could continue indefinitely. In reality, however, overall social resources and capital return requirements will not support it. Improvements in GPU efficiency will ultimately converge on an optimal curve balancing efficiency, cost, and performance. AI development, in turn, will seek its direction under the combined constraints of resources, energy consumption, and reasonable input–output ratios. At present, the path most likely to prevail on a comprehensive basis is GPU cluster scaling, which is also a necessary route toward AGI. Whether the AI industry will fully move toward AGI, or remain on a trajectory of specialized intelligence combined with human–machine collaboration, remains uncertain. Many social and even ethical issues are still under debate. However, from a profit-maximization, commercially driven perspective, the probability is high that development will move toward AGI. I personally define AGI as an inorganic humanoid brain—a silicon-based, inorganic humanoid brain created by organic brains that evolved over millions of years in carbon-based life. Returning to GPU clusters: their significance is not a simple arithmetic sum of GPU counts. Based purely on GPU physical performance, a clustered compute system can, in some cases, produce less output than the same number of GPUs operating independently and then aggregated. The true power of a cluster comes from socialized cooperation. All organisms capable of social cooperation are extraordinarily powerful—ants are a classic example. GPUs are no different. Once intelligent units form clusters that cooperate and communicate, they can unleash enormous power. This also vividly demonstrates the uniqueness of intelligence itself: a cluster composed of inorganic matter can exhibit social attributes. The larger the cluster, the stronger this capability becomes. However, the larger the cluster, the greater the difficulty of managing it. Large and ultra–large clusters impose extremely high requirements on safety and stability. These cannot be achieved by software alone. The demands on physical hardware infrastructure become even higher. The logic is the same as with high-end computers and smartphones: many applications simply cannot run on low-spec devices. From this perspective, debating whether a company’s development path is IaaS or PaaS at this stage is of limited significance. Before ultra–large-scale, AI campus–level infrastructure—power, land, fiber, cooling, scheduling, and reliability—is built, the PaaS layer remains too low-level and is, in essence, merely IaaS wrapped in a software layer. Now we can turn to AI smart factories. I asked four different large language models to provide definitions and extracted the most consistent elements. The result can be summarized as follows: An AI smart factory is a new generation of data center centered on ultra–large-scale accelerated computing. Through a complete software and systems stack, it converts power and data into intelligent output, undertakes both training and inference, and forms large-scale clusters via high-speed networks to enable the industrialized production of models, tokens, and intelligent services. It operates continuously on a 7×24×365 basis, transforming physical resources into digital intelligence (tokens) and becoming the core production facility of future enterprises. Scaleis its most fundamental characteristic. AI smart factories also exist at foundational, intermediate, and advanced levels—they form a capability curve. The stronger the capabilities, the more complete the closed loop, the higher the intelligent output, and the greater the economic value. Compared with data centers that merely adapt to the exponential growth of GPU energy efficiency, AI smart factories represent an upgrade in both hardware infrastructure and overall operating systems. They are required to possess: 1.Training capability (Training) 2.Inference capability (Serving) 3.Data feedback capability (Feedback) 4.Model iteration capability (Continuous Learning) 5.Unified orchestration capability (Fabric-aware Scheduling) 6.Scalability (Scalability) 7.Reliability, security, and isolation (Enterprise-grade) Explaining each of these technical terms in detail would be an enormous task. It is sufficient to understand the level of technological sophistication and requirements they represent. In simplified terms, the future AI data center landscape divides into two major tiers: those that provide compute power, and those that provide intelligence. Let me repeat the key question: Are you providing compute power, or are you providing intelligence? Or are you somewhere in between? Future data centers adapted to the AI industry can be broadly classified into several categories. The largest group consists primarily of facilities that provide high-density power and cooling infrastructure; their products are physical space and energy leasing. The second category builds on AI-compliant data centers by deploying bare-metal GPU clusters and directly offering raw compute leasing services to the market; their product is compute. The third category builds further on the second, focusing on intelligence processing and model management. These operators do not necessarily pursue extreme training capability, but they can provide enterprises with APIs, model fine-tuning, and full-scale inference services. They possess data feedback loop capability, rely on upstream models as foundational layers, and then productize and service these capabilities to deliver turnkey intelligent solutions downstream. The fourth category is the AI smart factory, the highest form. It achieves an industrialized closed loop of “energy input, intelligence (tokens) output,” integrating everything from ten-thousand-GPU–class cluster training and full lifecycle management to sovereign-grade security isolation. The above organization of the industry’s development trajectory is the result of analyzing and interpreting massive volumes of information and has already been significantly distilled. Understanding this development path is critically important. Without a macro-level perspective, it is difficult to evaluate investments in related companies, identify where true advantages lie, or focus discussions on what actually matters. It should be noted that these are conceptual classifications that approximate the future. The ultimate outcome will still depend on real-world commercial execution, but the overall direction is unlikely to be wrong. Based on the above, several points about the development of AI infrastructure and the AI industry can be stated with confidence: · GPU and GPU cluster compute efficiency grows exponentially. · Improvements in GPU efficiency will converge on an optimal curve balancing efficiency, cost, and performance. · AI development will find its direction through the combined constraints of resources, energy consumption, and reasonable input–output ratios. · AI smart factories represent the high-end form of the AI industry: they have high technical barriers, require massive investment, possess strong closed-loop capabilities, can industrialize the production of intelligence, apply across broad scenarios, and exhibit high value density. They are the highest-margin segment of the value chain. · At every stage of development, data centers are indispensable—and the more advanced the stage, the more critical they become, with correspondingly higher barriers to entry. Now I can classify the related stocks and examine who has the potential to evolve toward the highest form of the AI smart factory, and why. First, we exclude companies that do not touch GPU computing at all, or have only just begun to do so, and that primarily focus on leasing or hosting their own infrastructure. These are mainly Bitcoin miners, with representative names including CIFR, WULF, HUT, APLD, GLXY, CORZ, and BITF. The second group consists of what are commonly referred to as neo-cloud companies: CoreWeave, Nebius, IREN, as well as Crusoe and Lambda. With the exception of Nebius and Lambda, all of the companies listed above share a common trait: they are, or once were, cryptocurrency miners. Moreover, with the exception of @IREN, they were largely unsuccessful—or not particularly successful—as crypto miners. In the early years of the crypto mining boom, dozens of mining companies emerged. Some went bankrupt, such as CORZ; some transitioned very early, such as CoreWeave and Crusoe; the remainder stayed in Bitcoin mining and failed to achieve operating profitability. It can be said that more than 95% failed in their crypto businesses. This judgment of failure is not based on crypto prices not having risen “high enough,” but on operating fundamentals: their treasury and holding strategies violated basic principles of corporate operations, placing hope in massive uncertainty while neglecting operating cash flow and surviving on financing. This unhealthy approach was carried forward by CoreWeave after its transition to AI. Relying on strong demand and supplier dominance, it continued to downplay financial risk and has now arrived at a situation of sharply rising financing costs and potential default pressure. Crusoe, which followed a similar path, is facing comparable pressures. By contrast, Nebius and Lambda have much stronger foundations—at least in terms of corporate operating mindset—and do not exhibit major issues on that front. However, all of the above companies, including even the three hyperscale cloud providers, share a common problem: in the capital-intensive business of data centers, none of them possesses a vertically integrated business model. From the data center layer onward to the end of the value chain, varying degrees of partnership and outsourcing are required. Even for the three major cloud providers, pre-AI data centers primarily followed a hybrid model dominated by leasing, with ownership as a supplement. After GPUs entered the Blackwell era, the data centers required by the industry effectively need to be rebuilt from the ground up. Although the three major cloud providers have substantial financial resources, they are still constrained by multiple real-world factors. Combined with large-enterprise inertia, aversion to heavy assets, and a long-standing preference for stability, their response in this area may not necessarily be fast enough to secure a decisive advantage. Now I can introduce the central focus of this article: @IREN, the company that carries the core argument of this piece. In the field of high-performance computing and next-generation data center infrastructure, @IREN exhibits several distinctive characteristics: 1.Accurate foresight about the future, coupled with strong execution 2.In-house expertise as data center specialists, including engineering design and construction 3.Experience accumulated as the only true winner in Bitcoin mining 4.Adherence to sound operating principles and a strong sense of risk management 5.Most importantly, it is currently the only company to have achieved vertical integration from land and power, through data centers, all the way to GPU chips Based on these five characteristics, and in combination with the previously discussed elements of AI industry development, I will analyze whether @IREN is the best-prepared candidate to build high-end AI smart factories. Foresight and Preparation From the very beginning of its public listing, @IRENacquired a data center startup and assembled a team of professionals with extensive experience in infrastructure engineering and construction. The founding brothers possessed both financial structuring experience and a track record of successfully delivered infrastructure projects. Very early on, they foresaw a future in which explosive growth in digital-world demand would be mismatched with physical capacity. Acting on this vision, they began searching globally for ideal data center sites. From site location, to required attributes, to high-end configurations—especially backbone network access, grid power availability, green energy, and massive land and power reserves—these requirements could only have come from a team with a clear vision of the future. Otherwise, it would be impossible to explain why the standards were so high, the scale so large, and why all assets were fully owned and built from greenfield sites. At the time, none of this made economic sense. IREN went public as a Bitcoin miner, but Bitcoin mining simply does not require such high-end infrastructure or network configurations. Compared with many peers who housed mining rigs in rudimentary containers, this approach was like using a cannon to kill a mosquito. Bitcoin mining, therefore, was a transition and an experience-accumulation phase—a way to wait for the predicted digital-world explosion while gradually building high-performance computing infrastructure through real operations. When the industry they had anticipated finally began to erupt, opportunities naturally favored those who were prepared. The fact that IREN’s Canadian data centers could be brought online with high-end GPUs after only low-cost modifications is the clearest proof. Saying that IREN “transformed from Bitcoin mining to AI” only captures the surface. The depth of preparation by the IREN team allowed even the three major hyperscale cloud providers to fall behind. This ability—to see the future—is the rarest and most valuable capability of all. IREN’s foresight is also reflected in its data center design philosophy: optionality, diversity, adaptability, and the ability for different chips to operate collaboratively at the same site. When contrasted with the earlier-outlined contradiction between exponential GPU efficiency growth and mismatched physical infrastructure, this design philosophy is clearly aimed at addressing precisely that issue. There are many similar principles embedded in their designs, all of which reflect a deep understanding of the AI industry’s structural evolution. True Data Center Experts IREN’s core team consists of innovators and entrepreneurs in the data center field. Prior to joining IREN, they participated in the design and construction of multiple advanced data centers and had long collaborated with hyperscale enterprises. They possess a deep understanding of the past, present, and future of data center development, along with their own distinct innovations. However, during the CPU-dominated era, many of their ideas were not adopted by the industry. Performance was “good enough,” and projects with massive capital expenditure were avoided whenever possible—even by hyperscalers. As a result, they went through a difficult period. Some team members left to work at hyperscale companies, only to quickly encounter their conservatism and large-enterprise inertia. This experience helped them see both the constraints and the vast opportunity space in the sector. After the AI boom began, IREN—having accumulated long-term experience—earned trust and recognition during negotiations with hyperscale firms. In its collaboration with Microsoft, Microsoft’s assessment was direct: because we couldn’t do it ourselves. IREN’s innovation and forward-looking configurations in data centers have always been the aspect I personally pay the most attention to. I come from an engineering background but was initially unfamiliar with data centers. Fortunately, today’s large models are excellent teachers. I spent more than a month fully understanding the technical level of the data centers designed and built by IREN. This cannot be fully presented here—it is simply too extensive. Investors in IREN should do this homework themselves to truly appreciate it. In short: the future-oriented design and configuration of IREN’s data centers are extremely advanced. There is a qualitative difference between outputs produced by a team constantly innovating at the frontier of data center technology and those built by more traditional teams focused on maintaining the status quo. Practice and Experience Accumulation Bitcoin mining is not favored by Wall Street due to multiple uncertainties and therefore commands low valuations. Yet it is a classic high-performance computing business, and operating it well is extremely difficult. Among all Bitcoin miners, only IREN achieved sustained profitability with high margins. This success was the result of a complex, systemic engineering effort. The core began years earlier with data center site selection: choosing grids that avoid competition with residential demand, securing peak-load interaction subsidies, sourcing sustainable green energy, and controlling costs across multiple stages. Second—and most critical—was the precise planning of the compute growth path. Every upgrade in compute capacity was accurately positioned and delivered exactly as promised, without delays. Behind this lies orchestration capability, engineering experience, and the power of vertical integration working together. Third was the ability to foresee the optimal efficiency curve, wait patiently, and then act decisively. Early on, IREN had little presence in Bitcoin hash rate rankings, which were dominated by three major miners. But when mining hardware efficiency reached a certain threshold, IREN suddenly accelerated. In just over a year, it became a major miner. Had it not redirected its focus entirely to AI, it would likely have become the industry leader. After becoming a major miner, IREN was also the only mining stock with the strongest profitability. What impressed me most during this phase was IREN’s grasp of hardware efficiency. Once mining hardware efficiency and IREN’s comprehensive cost control reached the optimal curve, it acted immediately—and had the ability to execute flawlessly. Mining rigs are chips; GPUs are more complex, but the nature of efficiency improvement is the same. GPUs, too, will reach their optimal curve—the best input–output ratio. I am convinced that the future Sweetwater megasite was designed with this in mind, and that similar preparations have already been made at the current Childress site. Bitcoin mining not only provided IREN with strong operating cash flow; more importantly, it delivered irreplaceable hands-on experience in high-performance computing operations. IREN even records key operational parameters of its data centers on a daily basis. This has laid a solid foundation for future stability and energy-efficiency control in high-performance data center operations. This is the kind of work that is easy to talk about but extremely difficult to execute—and in this regard, IREN is unique globally. As discussed earlier, GPU efficiency improvements converge on an optimal balance among efficiency, cost, and performance, and AI development advances under combined constraints of resources, energy, and rational returns. @IREN, having already validated this through Bitcoin mining with strong profitability, is likely unique in its awareness, experience, and execution capability. More importantly, identifying and capturing the optimal curve is a systemic endeavor—it cannot be done simply by “understanding” it intellectually. It requires long preparation, internal and external coordination, capital planning, and accurate judgment of when GPU chips reach their optimal point. The integrated difficulty is extremely high. @brianfry01 @FransBakker9812 @jiahanjimliu @TheKamaHsutra

THERE IS NO SUCH THING AS RISKY STOCKS, JUST RISKY PRICES What we are watching right now is not a market rejecting innovation, nor is it a market suddenly deciding that artificial intelligence no longer matters. It is something more subtle and more common. The market is re-pricing certainty. When expectations become concentrated, when narratives become dominant, and when leadership narrows, risk quietly migrates from the business to the price. That is usually only obvious after prices move. Over the past several trading sessions, equity performance has not looked like a broad risk-off liquidation. Indices have been mixed, breadth has often been better than headlines suggest, and leadership has rotated rather than collapsed. Consumer-facing and idiosyncratic names have held up or even performed well on days when AI-linked stocks struggled. That pattern matters because it tells you the market is not indiscriminately selling equities. It is selectively reassessing where expectations had become the most compressed and the most certain. The AI complex sits squarely at the center of that reassessment. For much of the past two years, the dominant assumption embedded in prices was that AI adoption would be rapid, returns would accrue disproportionately to a small group of leaders, and margins would remain structurally elevated even as capital spending surged. That combination justified premium multiples and sustained concentration at the index level. What is changing now is not belief in the technology, but confidence in the smoothness of the payoff. One underappreciated driver of this shift is competition at the very top of the AI stack. As OpenAI, Google, and other platform leaders push more aggressively against one another, the narrative subtly changes from scarcity to rivalry. Rivalry does not eliminate growth, but it does complicate margins. When leaders compete intensely, they spend more, price more aggressively, and pull forward investment. Markets, which discount future cash flows rather than technical breakthroughs, respond by widening the range of plausible outcomes. When that happens, multiples compress even if revenue trajectories remain intact. This matters because the largest AI beneficiaries have also been the heaviest weights in the major indices. When the market begins to question whether today’s dominant platforms will extract monopoly-like economics or instead fight capital-intensive battles for share, it naturally feeds through to index-level earnings expectations. That dynamic alone can drive volatility, independent of any deterioration in near-term fundamentals. It is not a referendum on whether AI works. It is a debate over who captures the economics and at what return on capital. That debate spills outward into the broader AI ecosystem, including companies tied to compute, power, and infrastructure. Names like $NBIS, $IREN, and $CRWV have sold off sharply at times, despite announcing large, long-duration contracts that imply meaningful revenue visibility and, in some cases, upfront payments. On the surface, that looks irrational. In practice, it reflects the market shifting its focus from demand certainty to financing, execution, and timing. Large contracts reduce business risk. They do not eliminate equity risk. When projects require massive upfront capital, multi-year buildouts, and complex customer relationships, the equity outcome becomes highly sensitive to assumptions about cost of capital, dilution, and schedule. A contract that looks extraordinarily valuable on paper can still coexist with a volatile stock if the price already assumes flawless execution and favorable financing conditions. That is the distinction many investors struggle with during transitions like this. $NBIS is a good example of that tension. The company has positioned itself alongside hyperscaler demand and secured high-profile counterparties. At the same time, the scale of those commitments necessitates capital raises and balance sheet decisions that matter enormously to equity holders. When the market becomes more sensitive to dilution and funding structure, the stock can fall even as the underlying business opportunity grows. Both things can be true simultaneously. $IREN presents a similar profile. Long-term agreements, including prepayments, meaningfully de-risk demand and cash flow visibility. Yet the equity remains exposed to the pace of buildout, the magnitude of capital expenditures, and the conditions under which future capacity is financed. In an environment where investors are reassessing the return profile of AI infrastructure spending as a whole, those variables receive more scrutiny. The result is volatility that reflects price risk rather than a sudden loss of confidence in the business model. $CRWV underscores the same principle from a different angle. Multi-year agreements with blue-chip counterparties lend credibility and stability to the revenue base. But they also highlight how infrastructure businesses increasingly resemble long-duration financial instruments. Small changes in discount rates, construction assumptions, or counterparty risk perceptions can have outsized effects on equity valuation. When markets move from enthusiasm to selectivity, these sensitivities become more visible. What is important is that this dynamic is not confined to data center or infrastructure names. Memory suppliers like $SNDK and $MU illustrate the same lesson in a different part of the AI value chain. The long-term story around AI-driven memory demand and margin expansion remains compelling. Yet these stocks, too, have experienced periods where prices moved sharply on concerns about cyclicality, pricing power, or the timing of margin recovery. In each case, the volatility reflected a reassessment of expectations rather than a repudiation of the secular trend. This is why framing the current environment as “the AI trade is broken” misses the point. The more accurate description is that the market is sorting through second-order effects. It is asking whether AI spending translates into durable profits or merely accelerates competition. It is questioning whether scale leads to operating leverage or to an arms race. And it is adjusting prices to reflect a wider distribution of outcomes than was previously assumed. From an institutional perspective, this is a healthy process. Markets that fail to revisit assumptions tend to overshoot. Markets that do revisit assumptions often feel uncomfortable in the moment but ultimately create better entry points and clearer signals. Volatility, in this context, is not evidence of chaos. It is evidence of price discovery. For investors, the practical implication is straightforward but not easy. It requires separating belief in a business from discipline around valuation. Contracts, growth rates, and secular narratives all matter, but they do not immunize a stock from risk if the price embeds overly optimistic assumptions about execution, financing, or competitive dynamics. Conversely, a stock can look volatile and uncomfortable while becoming less risky if the price adjusts faster than the fundamentals deteriorate. This is why experienced investors tend to focus less on whether a company is “good” or “bad” and more on whether expectations are reasonable. Risk lives in the gap between what must go right to justify today’s price and what can realistically go wrong along the way. When that gap narrows, stocks become safer even if headlines look worse. When that gap widens, stocks become riskier even if the story sounds flawless. The current market feels like a moment when that gap is being re-examined across AI and beyond. Leadership is rotating, assumptions are being challenged, and prices are doing the work of absorbing uncertainty. That process can be uncomfortable, particularly for crowded trades, but it is also how long-term opportunities are ultimately formed. There is no such thing as a risky stock in the abstract. There are only prices that leave no room for error and prices that already reflect doubt. Understanding which is which matters far more than predicting the next headline. - RJC

Hi @franklee6924T. IREN’s data center design is evolving very fast and in partnership with NVIDIA. Just like chip design is using AI so is data center design. Rather than just provide high level answers to your questions I decided to use Super Grok to go deep. It delivered some fascinating results. Let me know what you think. ### Question 1: Is the 750-mile internal fiber's scalability advantage only feasible for massive single-site facilities like Horizon? How does this contrast with mid-sized or distributed setups like CoreWeave's, in terms of latency, costs, and AI training efficiency? Yes, the 750-mile internal fiber optic network in IREN's Horizon project—primarily at the 750MW Childress campus in Texas—is a scalability feature that's realistically only feasible for massive, single-site "giga-campus" facilities like Horizon. This design allows for campus-wide interconnectivity across expansive layouts (e.g., 420-576 acres at Childress), enabling seamless integration of multiple data halls into unified GPU superclusters without external network dependencies. The fiber supports multi-terabit throughput via dense wavelength-division multiplexing (DWDM), minimizing signal loss and enabling low-latency data shuffling critical for distributed AI training, where even microseconds matter for gradient synchronization in models like large language models (LLMs). In contrast, mid-sized or distributed setups like CoreWeave's (which operates ~30 facilities globally, often 10-50MW each, spread across urban/regional hubs) face inherent trade-offs: - **Latency**: Internal fiber keeps Horizon's intra-campus latency under 1-2μs per km (effective <1ms campus-wide), ideal for tight-knit superclusters. CoreWeave relies on external dark fiber or leased lines between sites, adding 50-200μs latency per hop—cumulatively 5-20ms across distributed clusters—degrading efficiency in all-to-all communication patterns for AI training. - **Costs**: Horizon's owned fiber avoids recurring carrier fees (e.g., $0.50-$2/Mbps/month for leased dark fiber), with capex amortized over 2GW+ scale yielding ~20-30% lower TCO. CoreWeave's distributed model incurs 2-3x higher opex from multi-site leasing, redundancy, and failover, plus elevated power draw from extra switching gear. - **AI Training Efficiency**: Horizon enables "campus-scale" superclusters (e.g., two 100MW zones for 100k+ NVIDIA Blackwell GPUs), supporting exaFLOP-scale training with 95%+ utilization via RDMA over Converged Ethernet (RoCE). CoreWeave's setup suits inference or smaller jobs but bottlenecks at scale, requiring software workarounds like sharding that reduce throughput by 10-25% and extend training times from weeks to months. This single-site moat positions Horizon for hyperscaler-grade AI workloads, where distributed alternatives like CoreWeave excel in flexibility but lag in raw performance for frontier models. ### Question 2: [Note: Based on context, this appears to address complementary aspects like power integration or cooling synergies with the fiber—e.g., how does the fiber enable hybrid air/liquid-cooled deployments at varying densities?] How does the internal fiber integrate with Horizon's flexible cooling (air/liquid) and rack densities to future-proof for evolving AI hardware? The 750-mile fiber acts as the "nervous system" for Horizon's flexible infrastructure, integrating seamlessly with hybrid cooling and variable rack densities (130-200kW/rack) to support a 5-10 year hardware evolution curve. By providing redundant, high-bandwidth pathways (400Gbps+ per link, scalable to 25Tbps), the fiber decouples compute placement from cooling constraints—GPUs can be clustered in liquid-cooled zones for high-density (e.g., NVIDIA GB300 NVL72 racks at 200kW) while air-cooled legacy hardware runs in adjacent halls, all synchronized via sub-ms latency interconnects. This future-proofs Horizon by: - Enabling dynamic reconfiguration: Fiber allows "hot-swapping" zones for next-gen chips (e.g., Rubin in 2026) without rewiring, maintaining 99.999% uptime. - Optimizing efficiency: Closed-loop liquid cooling (minimal water draw) pairs with fiber's low-joule photonic signaling, reducing overall PUE to ~1.1 and energy costs to $0.028-0.035/kWh. - Scaling for diversity: Supports mixed workloads—e.g., 100MW superclusters for training (fiber-shuffled data) alongside inference pods—yielding 20-40% higher ROI vs. rigid designs. Without this integration, evolving hardware would force costly retrofits; Horizon's approach ensures plug-and-play for AI's chip roadmap, from Blackwell to post-Moore paradigms. ### Question 3: What competitive moats does Horizon's design create for IREN versus other AI infrastructure providers? Horizon's design erects multiple interlocking moats that position IREN as a Tier-1 AI infrastructure player, outpacing commoditized providers like generic colos (e.g., Equinix) or crypto pivots (e.g., Hut 8): - **Scale & Site Control**: 750MW energized today (expanding to 2.75GW by 2027) on 100% owned land (thousands of acres) enables giga-campus economics—$300-350M capex for 75MW Horizon 1 yields $10M/MW ARR potential. Rivals with fragmented 50-200MW sites face 30-50% higher build times (12-36 months vs. IREN's 6-12) and grid bottlenecks. - **Network Fortress**: The internal 750-mile fiber + carrier-neutral meet-me rooms (Zayo, Lumen) deliver <6ms to Dallas IXPs, creating a "low-latency enclave" for AI. This barriers entry for distributed players, who pay 2-4x premiums for equivalent bandwidth. - **Vertical Integration**: 100% renewable power at $0.05/kWh (ERCOT direct-connect), plus ARIN ISP status for in-house DDoS/firewalling, slashes opex by 25-40%. Liquid-cooling readiness for 200kW racks locks in NVIDIA ecosystem partnerships. - **Execution Edge**: Rapid pivots (e.g., ASIC-to-GPU in months) and $1B ATM facility fund 57EH/s+ expansions, outrunning supply-constrained peers. Moat durability: 5-7 years, as replication requires rare combos of land/power/fiber. These create a flywheel: Early AI contracts (e.g., multi-year for 23k GPUs) fund further scale, widening the gap in a market where 80% of new capacity is pre-committed. ### Question 4: How does Horizon's architecture enhance NVIDIA chip performance, especially amid hyperscaler custom chip developments (e.g., Google TPU, AWS Trainium)? Horizon's architecture amplifies NVIDIA's GPU performance by 15-30% in real-world AI training/inference, countering hyperscaler custom silicon (e.g., TPUs v5p, Trainium2) through optimized interconnects and density: - **Interconnect Boost**: 750-mile fiber enables NVLink/NVSwitch fabrics at campus scale, achieving 1.8Tbps/node for Blackwell/GB300—reducing all-reduce times by 20% vs. standard Ethernet. This sustains NVIDIA's CUDA ecosystem edge, where custom chips lag in software maturity (e.g., TPUs require JAX refactoring, adding 10-20% dev overhead). - **Density & Cooling Synergy**: Direct-to-chip liquid cooling at 200kW/rack maximizes TDP utilization (e.g., B200 at 1kW/GPU), yielding 95%+ FLOPS efficiency. Hyperscalers' air-cooled customs cap at 100-150kW, throttling sustained loads. - **Latency/Throughput Edge**: <1ms campus RDMA shuffles datasets for trillion-parameter models, outperforming distributed TPU pods (5-10ms inter-pod). Horizon's fiber also supports hybrid NVIDIA-custom runs, hedging while leveraging CUDA's 90% market share. - **Resilience to Customs**: As hyperscalers internalize (e.g., 40% of Meta's training on MTIA), Horizon attracts NVIDIA-dependent indies/SMBs with $0.05/kWh power + fiber, offering 2x faster time-to-market vs. cloud markups (20-50% premium). Overall, Horizon extends NVIDIA's moat by providing the "physical CUDA"—infrastructure that unlocks full chip potential, making it indispensable even as customs proliferate.

IREN — The only AI high-performance data center startup (intelligent factory) in the U.S. stock market This article is important and worth reading for every @IREN_Ltd investor. Over the past month, AI infrastructure sector stocks have been undergoing a broad adjustment. This cooling-off period is a very suitable time for investors to think: what exactly is the most important part of this industry’s development? At the moment, when infrastructure construction has not even reached an early prototype stage, some discussions are overly enthusiastic about the topic of space data centers. I think this is far too off-track—before we even learn to walk, we are already trying to run, or even fly. This is clearly detached from reality. The next two years will be the period in which AI-era data center standards are created and formed. Compared to the internet era, data center technology cannot simply be copied and reused. The early generation of GPUs could rely on that approach, but the further we go, the higher the technical requirements become. Only companies that can achieve breakthroughs in this area, possess technical leadership strength, and ultimately become the formulators of new industry technical standards—only such companies truly own the future. Before IaaS settles into an industry standard and before scaled service providers emerge, discussing PaaS may be premature, let alone space data centers. Even if AI data centers are really built in space someday, the most qualified one will still be the absolute leader in the field of AI data centers (intelligent factories) on Earth. This is inevitable. It is impossible for a “space data center expert” or “leader” to appear out of thin air without any foundation. Its technical references and standards will definitely come from the largest, best-performing, safest, and most stable one on Earth. Therefore, we should return to Earth, and return to the topic of leading AI data center technologies. This is the most important focus area for the next two years. It is obvious that the recent focus of the IREN community has drifted, to the extent that a post by @brianfry01 did not receive enough attention. As far as I recall, this is the longest post Brian has ever made. If everyone values the fact that he was a member of IREN’s early team and has spent many years working on frontline data-center technology entrepreneurship, then the topics he emphasizes should be taken seriously. Here is the link to that post: x.com/brianfry01/sta… The first theme this article is going to discuss is: “In the AI era, what qualitative changes have occurred in data centers compared to traditional ones? Which factors will become the decisive standards for future AI data centers?” First, data centers in the AI era are not an upgraded or enhanced version of the old model—they are a completely different species. This is manifested in four critical dimensions: ① A qualitative leap in rack density: from 5–10kW per rack → the 200–300kW era Traditional data centers are primarily designed for servers, storage, and light computation, so rack density is typically: 5–10 kW (general servers) 15–30 kW (high-performance HPC racks) But AI GPUs (especially NVIDIA Blackwell / Rubin / post-Blackwell generations) require: 1000W+ per GPU Each NVL72 rack draws 130–200 kW Future densities may reach 300 kW This is a 20–40× generational difference. The resulting qualitative changes include: Cooling must shift from air cooling → cold plate → immersion liquid cooling Power infrastructure must be completely redesigned (substations, lines, redundancy structures) The data hall architecture changes from “server shelves” to “a semiconductor-factory-like power distribution system” This is not an upgrade; it is a rebuild. ② Network structure shifts from “rack interconnects” → “supercomputer-grade interconnect architecture” Traditional data center networks focus on: User → server Internet ingress/egress bandwidth Master-slave structure, load balancing AI data centers focus on: Continuous high-throughput GPU↔GPU communication Inter-cluster memory synchronization, gradient updates, parameter exchange Latency, bandwidth, and jitter matter far more than total bandwidth AI training requirements resemble supercomputers far more than traditional data centers. AI internal bandwidth needs are 100–1000× those of the internet, with near-zero latency tolerance. ③ Data center “scalability” shifts from linear → exponential Traditional DC expansion is simple: Build more buildings Rent more power Pull more fiber AI data center scaling has exponential complexity: GPU scale is not linear—larger clusters are more capable Only above certain scale can giant models (100k+ GPUs) be trained Network topology must scale horizontally without fragmentation The larger the cluster, the higher the utilization and the lower the unit TCO Therefore: Large single-site campuses = explosive economic advantage Distributed small sites = network fragmentation + latency → increasingly unsuitable for ultra-large-model training AI infrastructure is no longer a real-estate business, but: Larger scale → stronger compute → attracts bigger models → attracts more customers → drives more scale → irreversible positive feedback flywheel. ④ Overall system stability must reach “telecom-grade / banking-grade / aviation-grade” Training a giant AI model may require: Continuous operation for 3–12 weeks Any short downtime may directly destroy millions of dollars’ worth of compute progress This requires: N+1 / 2N power redundancy Network redundancy GPU cluster topology redundancy Cooling redundancy Full concurrent maintainability Traditional DC uptime requirements are high, but compared to AI’s requirement for ultra-long uninterrupted training, they are an entirely different magnitude. Second question: Which factors will become the decisive standards for future AI data centers? Future AI data center standards will be defined by 5 key indicators: ① Power scale (GW-level) + extremely low electricity cost ② Single-site scale (single location >250MW, ideally >500MW) ③ Internal networking capability (fiber density, DWDM, multi-path ultra-low-latency RDMA) Key indicators include: DWDM wavelength density 400Gbps → 800Gbps → Tbps upgrade capability Campus-wide latency <1ms GPU-GPU communication jitter must approach zero ④ Cooling capability (200–300kW racks / liquid-cooling dominant) ⑤ Vertical integration (power + fiber + operations + financial supply-demand structure) Traditional data centers follow the real-estate + power-hosting model. Top-tier AI data centers resemble TSMC-like industrial giants. Traditional winning criteria: “who has more buildings, higher occupancy, slightly better PUE” AI-era winning criteria: “who can build larger single sites, higher density, lower latency, stronger interconnect, stronger power + cooling, and the most stable giant GPU clusters.” Everything above is a basic整理, not necessarily exhaustive, but the core points are all covered. For those familiar with the IREN community, none of this is new. Putting them together comprehensively might be the first time. My conclusion is: The “colocation” option that everyone constantly discusses is simply not possible as a standalone choice in a top-tier AI data center. Or to put it differently: IREN providing top-tier data center services is equivalent to an intelligent factory. CIFR providing colocation is equivalent to a traditional data center migrating into the AI era. The gap in technical difficulty and complexity is exponential and cannot be compared. This gap will become increasingly obvious over the next 1–2 years, and the market has not priced this in at all. From the perspective that high-level AI data centers = intelligent factories, CoreWeave and Nebius are simply not on the same dimension, because neither company has treated data center technology itself as a key point from day one. Their investment cannot compare with IREN’s. And one thing the market has still not fully recognized is: IREN is actually a high-performance data center startup. If IREN is considered a high-tech company, where is its high tech reflected? Naturally, in its next-generation high-performance AI data center technology—its “AI intelligent factory” technology. If the market created a category or concept specifically for this field, I believe IREN is currently the only representative. High-performance data centers adapted for AI are an entirely new creation. Traditional data centers cannot compare. While most Wall Street analysts cannot yet see clearly and have not positioned IREN correctly, @IREN_Ltd has actually, through years of early preparation and practical execution, already set the industry standard for AI high-performance data centers—AI intelligent factories. This may sound sudden to first-time readers, but I have researched this topic for half a year, and I can now confidently give this definition. We can observe over the next year whether this positioning becomes increasingly clear. What does this imply? Its overall technical architecture is unique. Its vertical integration is unique. Its scalable, flexible design and practical know-how are unique—especially at non-laboratory, large-scale, commercially replicable levels. Multiple ultra-large single sites are an absolute leadership advantage; combined with other IREN “uniqueness elements”, it becomes uniquely unique. Finally, what I want to emphasize: The newly disclosed Horizon 1–4 projects feature 750 miles of internal fiber. IREN casually mentioned this in a video, but I believe this is explosive information with profound implications. @brianfry01 also called this out specifically—and with his years of experience in frontier data center engineering, his sensitivity is not simple. What does 750 miles of internal fiber mean? • Land scale & continuous layout: Deploying 750 miles of internal fiber requires enormous cable trenches, handholes, conduits, and redundant routes. Only campuses spanning hundreds to thousands of acres make this economical. • Centralized Power Generation and Distribution: Ultra-large clusters require hundreds of MW of IT power delivered centrally; scattered sites cannot aggregate power into a single low-latency domain. • Controlled network domain & latency guarantees: Internal fiber allows end-to-end design, redundant rings, minimal hops—city/ISP links cannot guarantee similar jitter/latency. • Cost/operations & permits: Such a project demands long-term CAPEX, complex permitting, and a dedicated maintenance team—only feasible with major customer pre-commitments or strong financing. • Hyperscalers (Microsoft/Google/AWS) do have huge backbone fiber mileage, but these are cross-region/global totals—not the internal fiber mileage within a single campus. (Microsoft’s “hundreds of thousands of miles” refers to its global backbone, not any single campus.) Conclusion: Based on publicly available data, the only clearly confirmed single-site project with this scale of internal networking is IREN’s Childress (Horizon). Thus, we can infer that such massive internal fiber deployment within a single site currently exists only at IREN. Core value of internal fiber: (1) Lower latency Single-campus, point-to-point, fully controlled Very few hops No ISP routers/intermediate equipment → Reduces GPU supercluster synchronization time by orders of magnitude. (2) Lower jitter Jitter is one of the most dangerous killers for large-scale AI training Small models aren’t sensitive; large models (especially MoE) are extremely sensitive → Internal fiber provides stable optical paths for all-reduce and gradient sync. (3) Higher bandwidth “Bandwidth” here refers to single-link throughput: 400Gbps 800Gbps 1.6Tbps DWDM can stack these. (4) Total / Aggregate bandwidth This is not “bandwidth repeated twice”—it means: “Single-link bandwidth” ≠ “total campus throughput capacity”. They are entirely different. The purpose of doing this is the following: 750 miles of internal fiber = creating “near-theoretical-limit” compute stability for large-scale GPU clusters. It allows high-performance computing power to become controllable, predictable, jitter-free, hop-free, and free of any carrier involvement, because this 750-mile network is a fully private fiber system. Its significance for GPU mega-cluster training is a qualitative leap. It fundamentally solves the biggest killer in distributed training — jitter. This 750-mile internal fiber is the ultimate solution to jitter. 750 miles of internal fiber + single-campus mega-cluster architecture Based on NVIDIA’s recommended practices and HPC supercomputer papers, it is reasonable to expect: ≈ 99.995% – 99.999% (four-and-a-half nines to five nines) training stability. During training, GPU node availability can even reach: Six nines (99.9999%) network stability. @jiahanjimliu — do you think these nines are sexy enough? Up to this point, IREN’s Horizon project clearly reveals its uniqueness — and this is absolutely not trivial. Once built, with performance this high, the per-MW unit price should undoubtedly be the highest. Microsoft paid an upfront commitment, largely because there are many things in high-performance datacenters that Microsoft simply cannot do — but IREN can. Being able to train frontier-scale models in a data center with such ultra-high performance would allow any customer to leap far ahead of competitors. And for IREN, this battle is crucial, because it is a milestone in setting the standard. The exponential growth of AI compute is largely driven by large-model training, not inference. And large-model training is precisely the decisive battleground for top-tier AI players. If IREN can provide this level of ultra-premium stability, then the bargaining power in partnerships will naturally sit entirely on IREN’s side. So here’s the question: Will IREN investors fully recognize IREN’s technical leadership in AI datacenter architecture? Do you all see it now? Things that people keep highlighting — Coreweave’s strong software stack, or Nebius’s thousand experienced engineers — do these advantages still matter as much? Yes, I admit they matter. But even if you twist all those strengths into one rope, they still cannot solve the killer-level interference called jitter during ultra-large-scale GPU compute. IREN’s high-performance physical infrastructure nearly solves this problem at the theoretical limit. This is also a killer-level advantage in AI datacenter competition — because even competitors want to build internal networks of this caliber, but it is almost impossible: They need land that is large enough — and privately owned. The single site must be large enough. A customer must be willing to commit early — without that, the economics simply don’t work for such massive investment. Now, do you see what IREN’s moat actually is? In business competition, survival and long-term success ultimately come from having a unique, unmatched, ultimate product. I believe IREN has more than one such product — more will continue to emerge. We simply need to be patient. @brianfry01 @jiahanjimliu @FransBakker9812 @litigious_dulce @TheKamaHsutra

Brian Fry @brianfry01

6 months ago

@franklee6924T @iren Hi @franklee6924T. IREN’s data center design is evolving very fast and in partnership with NVIDIA. Just like chip design is using AI so is data center design. Rather than just provide high level answers to your questions I decided to use Super Grok to go deep. It delivered some

1 1 13 416 10

$CIFR just performing here.🔷🔷 Broke above the silver resistance, and closed above it. Next up $22.72 as potential resistance. Flow pane with another big, strong increase. Third Blue Diamond must be close for $CIFR here. 🔁 Repost for daily $CIFR updates. Interested in trading Suite yourself? Check the link in my bio! You got this! 🫳

TheGentleTraveler @LeaderInvests

7 months ago

$CIFR DOUBLE BLUE DIAMOND⁉️ On the daily, $CIFR confirmed double blue diamonds, and an increase in whale participation! We also closed above the daily golden trend, made a slight higher high. This is very bullish to see. Take out $18.50, the silver resistance, and $22.70 can

3 8 66 14K 4

@vulturetrades $OPEN

@LEAPTRADER_ pearls

“We’re living through the Fourth Industrial Revolution. Every past industrial revolution was defined by factories and the workers inside them. This time, it’s different. Inside these new AI factories, the “workers” aren’t people on assembly lines – they’re GPUs, clustered together, never clocking off, never tiring, able to scale to millions.” At $IREN, we’re building AI factories across North America, leveraging our power, data centers, and GPUs to meet the accelerating demand for compute. As @danroberts0101 explains on @theCUBE, $IREN is scaling with speed and efficiency – and the potential ahead is enormous. Full interview: youtube.com/watch?v=nmBJjk…