Marbles

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/1n8WcKs…). 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.

@AlexfromBabylon Just echoing what @FransBakker9812 said here. It allows $IREN to service a longer-tail of enterprises who want the datacenter abstracted away with Kubernetes / orchestration / observability. Orchestration / kubernetes / observability isn't where the moat in software is.

@AlexfromBabylon It's most likely a fit for demand, not a platform that is trying to be a jack of all trades like $NBIS is doing with their "core business" (which is < 5% of the revenue and MW lol).

$IREN's Kent Draper: "We are working together closely with $NVDA for the future of our software platform" 👀

@BitcoinAIGuy Based.

when $IREN was $69-$75 everyone was a long term investor in it for $1,000+ now $IREN's $41, the business fundamental's are improving and there are zero signs market demand for compute is slowing down...but the tourists left after the stock price tanked... this serenity clown's posts on twitter are clearly signs you'd see at or near the bottom; I don't think $IREN goes under $36 again... if it does, I'll buy calls aggressively. i bought more shares this week and plan on buying more shares every week going forward. I may adjust my share accumulation strategy by the $100s... could $IREN correct from $120-150 to $69-$75? Sure, but that'd be another great time to accumulate before you see $200+ volatility is vitality

@Agrippa_Inv This needed to be said out loud. It's so obvious, but I've noticed many investors still can't tell the difference.

$IREN: Fake vs Real Power Something that appears to be overlooked by the broader $IREN investor community, and especially by analysts, is the recent set of ERCOT (Texas grid) changes. For a long time, most data center developers could purchase random parcels of land, or secure land through option agreements, and then claim they had a “multi-GW” power pipeline in Texas. Somewhere in the footnotes, they would then clarify: *Pending grid studies, interconnection agreements, and ERCOT approvals.* It used to be difficult to distinguish between what was “real” power and what was, at best, aspirational. That was especially true for projects already halfway through the process, such as those that had partially completed grid studies but had not yet secured the remaining approvals. This led to a tsunami of “fake” power claims from operators across the industry, many of whom used inflated pipeline claims to support their valuation and strengthen their negotiating position. With the new ERCOT changes in place, that is no longer possible. The statewide grid operator has now introduced a new batching system that categorizes projects based on real, tangible progress and the steps they have taken, queuing them accordingly. As of today, it largely comes down to whether your project is in “batch 0, base load”. If you are not in this initial batch, you can effectively kiss your project goodbye until, at the very earliest, 2028. For the first time, counterparties (i.e., customers) and investors alike can distinguish between what is real and what is fake, or at best still many years away. This plays directly into $IREN's hands, as it is one of the few operators in the space that has always been straightforward about its power pipeline, only announcing sites that have gone through all the necessary steps and approval requirements. As such, $IREN's Sweetwater 1 & 2 sites are all but guaranteed to be in the first “base” batch, with additional unannounced sites also likely to be part of it. These new changes undoubtedly strengthen $IREN's negotiating position, as the pool of seemingly near term capacity has now shrunk by a meaningful margin.

@mikealfred Tyler's the man.

I respect the objectivity, not many people can go from $IREN to $NBIS. It seems like you have a high degree of confidence Nebius will have the power + infrastructure ready to service Meta by early 2027. Would love to get your thoughts on how Nebius is going to do this given that they have no clear line of sight to any power in 2026 and their ability to energize power that power at scale in 2027 is unclear.

Buying $NBIS on this dip. First add for me. Yes, they didn't share financial metrics with the 27B announcement - but I am not trying to bet against a Russian, Jewish, Israeli, Mathematician who is a hyper successful serial IT founder in this space.

“Grid connected and contracted power are NOT moats” tell that to CoreWeave. They had $66.8B in backlog and “Leadership + Talent” too. Then ONE contractor delayed 5 sites because they couldn’t get power connected on time. 16% crash. Guidance cut from $5.15-5.35B to $5.05-5.15B. Capex slashed from $20-23B to $12-14B. Class action filed February 2026. The bottleneck wasn’t talent. It was transformer delivery times and utility interconnections. The exact thing you just said doesn’t matter. “This is not Bitcoin mining” correct. Bitcoin mining is easier. You plug in ASICs and hash. AI infrastructure requires liquid cooling, high-density racks, custom networking, utility-grade power interconnections, and delivery timelines coordinated across GPU suppliers, construction contractors, and utility companies simultaneously. It’s HARDER than mining. Which is exactly why power and physical infrastructure matter MORE, not less. “Leadership + Talent” — Arkady Volozh built Yandex. Respect. But Yandex was a search engine. This is building 3+ GW of physical data center infrastructure from 170 MW across 9 new sites in 6 countries in 12 months. Leadership doesn’t pour concrete. Talent doesn’t install transformers. And neither controls NVIDIA’s Vera Rubin production schedule, utility interconnection timelines in France, Israel, Alabama, Missouri, New Jersey, and Oklahoma simultaneously, or third-party construction contractors who answer to their own schedules. Let me tell you exactly where we are in the cycle: PHASE 1 — LAND GRAB (2024-early 2025): Neoclouds rush to sign GPU supply deals with NVIDIA and capacity commitments. NBIS, CRWV, IREN, ORCL all pile in. Backlogs inflate. Stock prices run. ✅ Done. PHASE 2 — HEADLINE DEALS (mid 2025-now): Sign massive hyperscaler contracts to justify capital raises. CRWV signs MSFT, OpenAI, Meta. NBIS signs MSFT $17.4B, Meta $3B, then Meta $27B. Each headline triggers equity offerings, convertible notes, warrant issuances. Every deal announcement is a capital raise in disguise. ✅ We are HERE. PHASE 3 — DELIVERY RECKONING (2026-2027): Contracts hit delivery deadlines. Infrastructure either exists or it doesn’t. Prepayments either arrive on schedule or they don’t. Power either connects or it doesn’t. This is where CRWV already cracked — and NBIS hasn’t even entered this phase yet with an 18x gap between active power (170 MW) and contracted target (3+ GW). ⏳ Coming next. PHASE 4 — SHAKEOUT (2027-2028): Companies that built real infrastructure operate at high utilization and generate cash. Companies that over-promised and under-delivered face contract terminations (MSFT 6-K: “right to terminate”), covenant breaches, and forced dilution. The strong survive. The leveraged don’t. GPU prices compress further. Custom silicon from MSFT, META, GOOG takes share. Only operators with owned infrastructure, secured power, and diversified revenue make it through. PHASE 5 — PLATFORM ERA (2028+): The actual prize. Built infrastructure generates recurring high-margin revenue. Platform layers on top of physical capacity create software-like economics. Unit economics compound instead of dilution compounding. THIS is where “Leadership + Talent” finally matters but only if you survived phases 3 and 4 with infrastructure in the ground, not contracts on paper. Daniel, you’re pricing Phase 5 into a company stuck at Phase 2 with the hardest parts still ahead. “It will only get better from here” assumes flawless execution through the exact phase that destroyed CoreWeave’s stock — and NBIS is starting from a worse position. 170 MW vs CRWV’s 850 MW. 18x capacity gap vs CRWV’s 3.6x. $5.7B capital vs $30B+ capex need. Zero operational track record at scale. $NBIS > rest? The filing says otherwise. Show me the delivered megawatts, not the signed megadollars.

@AlexfromBabylon Based take from an investor of both. People should ask how $NBIS will service their contract with $META without clear line of sight to any substantial power this year and even some of the next.

$NBIS $IREN It's all fun and games between the community here on X, but in terms of investment decision these salty takes are unjustified. Meta just commited to $27B buildout with Nebius far in advance of having datacenters ready. Even if Nebius bare metal orchestration / tooling is more advanced that will be made up with IREN having datacenter capacity ready today. This deal again validated the market demand for compute and it does not take a lot of fantasy how this will be positive for IREN as well. That's why I bought more IREN pre market after this announcement as well. Market for now seems to agree. Remember Nebius and Coreweave have quite a lot of revenue bookings right now for which datacenters are still being built. It does not take a lot of fantasy that if you have Meta / Microsoft level competing needs and Nebius / CoreWeave have their hands full for the coming period that it might be a good idea to secure capacity outside of them as well. That will only benefit IREN.

$IREN Weekly Space 3/15/2026 @ 9PM Join us for another weekly IREN themed space. This week we will talk about everything software, tokens, and use-cases of AI. With: @SmallCapSnipa @jiahanjimliu @marblesats Co-Host: @bitcoinbutcher1 Will be recorded x.com/i/spaces/1AxRn…

@FransBakker9812 @SmallCapSnipa @jiahanjimliu @bitcoinbutcher1 👀👀👀

Fair enough. I have a differing view: - I think power bottleneck will get worse -> people are severely underestimating how better + cheaper models will lead to substantially higher need for compute - market is still highly unsaturated, with only heavier saturation in coding with low saturation in administration, healthcare, law, science, commerce, etc. - I don't think $IREN's moat is only around power. Look at $RIOT + $BTDR. They have substantial amounts of power, but why are they underperforming? - I think that memory / silicon is a top of mind constraint but it's really a chicken and egg problem. At some point you'll need power to plug in increasingly power-hungry chips. At which point bottleneck will flow back into power. - fair on capex / dilution but I like Cloud but this really goes back to your view on demand - if you think supply catches up to token demand in the next 1-2 years that's fair. But I don't see supply catching up to demand in the next 5 years. I think people view demand is relatively static when demand is still exponential (bullet point 2). - if you like colocation and demand / supply imbalance only gets worse, will then there's a world where you locked in megawatts at pico-bottom prices. Opportunity cost is risky in its own way.

it seems like silicon / memory has been somewhat priced in. I struggle with IREN because they are taking the most dilutive / capex heavy route and it's unclear what their returns will be. I also struggle with how opaque they are. I don't think power alone is a moat and the power bottleneck is loosening looking out.

On one hand Semi-analysis is saying power is no longer a constraint. On other other hand everyone agrees we need hundreds of gigawatts by 2030 and Dylan Patel says that's impossible.

Even without those things, everyone's still underestimating how much compute is needed - just for LLM's that are still getting better. There is so much opportunity in the market that is still unsaturated: - agentic health workflows - agentic commerce - law - administration, etc. Right now most of the saturation is going to coding applications. Wait until AI permeates into everything else.

@marblesats @BitcoinAIGuy We havent even taken into account Physical AI, quantum computing and other use cases!

$IREN has 4.5 GW secured… but what if they have another 4.5 GW in development? 👀 $90B ARR POTENTIAL BY 2037 NOT PRICED IN MARKET CAP COULD EXPLODE INTO THE TRILLIONS OVER THE DECADE YOU ARE NOT BULLISH ENOUGH!!!!!

@jiahanjimliu @midlevelcruiser Yeah you've been spot on by this. Many large accounts have also been shouting at this ad nauseam all over twitter - indicative of that the fact that this is substantially more priced in than power.

@marblesats @midlevelcruiser I’ve been calling HBM as the #1 bottleneck but that seems to be priced in or a lot priced in.

Is your view silicon related / memory related exposure as a better investment at this point in time? Do you still view $IREN or other power plays as a good investment? Why or why not? Would love to get your thoughts as I know you made a lot of money over the past year but have since reduced your exposure to $IREN into other plays.

@marblesats no he’s saying that the main bottleneck is shifting. Power is still constrained but no longer the number #1 constraint.

@jiahanjimliu You mentioned some companies that needed handholding will go with Hyperscaler PaaS / Databricks PaaS - I'm not bullish $NBIS myself, but I would love to get your thoughts on why they would not go to Nebius.

$IREN: The Future of Open Source Models is on Open Source Inference Stacks on Bare Metal Open vs Closed Sourced Leading enterprises will either use a frontier closed model like ChatGPT, Anthropic or they will use open source models. The benefits of using an open source model is control and margins. A company using an open source model may use a managed inference service like AWS SageMaker or $NBIS Token Factory as an intermediate step. If their motivation for using open source models is control and margins then they will not get stuck on a closed sourced inference stack charging 80% margins. The progression goes like this: 1. Start with bare metal for pre-scale application inference. 2. Buy managed infrastructure to quickly scale their application. 3. Develop their own inference stack to expand their vertical integration and improve their margins. Previously only top companies like Google, Netflix, Meta, Uber, Palantir, Microsoft, Amazon.com, Coupang, Tesla, Apple, Snapchat, Lyft, etc make it to step 3 but that will change with Claude Code. Netflix has greatly minimize it's dependencies on AWS from the early days. What AWS does for Netflix is physical infrastructure, instance provisioning (GPU equivalent would be Autoscaling Kubernetes), managed Dynamo DB. Fully Open Sourced Stack Fully Open Sourced Inference Stack: Application -> KServe -> vLLM/TensorRT -> CUDA -> GPU. The inference stack would then be orchestrated by Kubernetes across GPU nodes. $CRWV's stack is Application -> KNative (1) -> vLLM/Triton/TensorRT -> CUDA -> GPU orchestrated by Coreweave customized autoscaling Kubernetes (2). In fact, $NBIS own Token Factory is just a vLLM inference stack. Nvidia shows many companies use the TensorRT variant of the fully open sourced stack: American Express, AMP Robotics, Baidu, ByteDance, Criteo, GE Health, JD, SAP, Siemen's Health, SNap, Verizon, Zoox (3). The corollary of this fact is that deploy on Kubernetes + bare metal instead of some wall garden inference runtime. Even if a customer uses AWS SageMaker or $NBIS Token Factory, it will not be hard to migrate off as long as the in-house inference stack is well configured. Any AI developers knows that keeping the same open source model but migrating from one API that serves it to an internal API is not hard. In 2-3 years, with Claude Code and AI Agents, managing and scaling an inference serving stack will be doable with a small team of engineers. Bare Metal Value Bare metal solves the problem of hardware uptime, capital, datacenter design, and power. Nvidia solves the foundational layer of software between the GPU and middleware layer. Previously many software companies were just application on top of AWS services but now Claude Code will commoditize application code to force software companies to do middleware. $NBIS investors think that the same software companies that go to open source models will then obediently pay 80% margins and hand over control of their differentiation to $NBIS. The future of open source models will be on open source inference stacks on auto-scaling Kubernetes + bare metal GPUs. Companies that need hand-holding will be on ChatGPT/Anthropic API or Databricks/Hyperscaler. Companies that need top of the line kernel optimizations will be on FireworksAI. Companies that need auto-scaling Kubernetes + bare metal GPUs will be on $CRWV, $IREN, $NBIS.

@CompoundingX Exactly - I applaud their foresight, even if the original vision didn't fully account for the Chat GPT moment. I think they have the best read on just how much the world will need compute in the future and see this not as a 2-3 arbitrage opportunity, but a multi-decade one.

@marblesats Agreed!! Man, the vision of Roberts brothers on couple things have panned out really well.

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.

@pei_liu3 🙌

💯