Started buying $NBIS in Dec '2024. Naysayers will say this is "Inspect Element." Some people started thrown around PT's for $NBIS. It's too early to put a price on this gem. Agentic AI is lifting off. The Physical AI tsunami is next. HPC demand will achieve escape velocity.

Most of you know me as the buff Arkady who knows $NBIS better than most people know their own family. Well… meet the version with glasses: @nebulizerlabs. Yesterday I took the first step and launched my Substack. The goal is simple: go deeper 😏 Genuinely, thank y'all for your parasocial support. For cheering on my obsessions and autistic deep dives. You've helped me rediscover how much I love writing, researching, and teaching myself new things. In particular, I want to thank these individuals for sharing my work and encouraging me to keep putting myself out there. @babyfolio @TheBigBerbowski @AndersInvests @EndicottInvests @FrancisCashFlow @RayT168 @mvcinvesting @geokoutalidis @reelyphishy Nothing on X is going behind a wall! Substack just has better UI/UX scaffolding for longer form content. Expect plenty of free research. I hope it becomes a place for those who want to understand, not just trade the headlines. With that being said: Part 1 of Nebulizer’s Guide to Nebius is live ⚡️ My most in-depth $NBIS writeup yet. Completely free. ദ്ദി(˵ •̀ ᴗ - ˵ ) ✧ Time to grow some fresh cortical folds. smasimho.substack.com/

BNP Paribas is forecasting $NBIS Q4'27 ARR @ $22bn Wells Fargo is forecasting 2GW of active capacity by EOY '27. Implies a blended rate of $11MM/MW/YR. This is a hyperconservative estimate even relative to where existing bare metal contracts are at. $22bn is FAR TOO LOW! 🆙

Do you know how much capital Nscale still needs to raise to actually deliver on a near 100% bare-metal backlog to Microsoft and Anthropic? ~ 1.1GW of future capacity is locked in at today's deal prices. Nscale currently has just ~55MW active. nitter.net/IPONewsroom_/status/21…

I find the “power is the constraint” argument increasingly parochial and myopic. I’ve noticed that this rhetoric is primarily pedaled by smaller BTC-miner outfits that don’t have the upper stack of higher margin managed inference and agentic solutions. Something that both $NBIS and $CRWV do. For one, the US is one of the most expensive places to build datacenters. For another, this increasingly hackneyed, nativist argument assumes that inference demand exists only in the US. It ignores the reality that countries around the world will consume AI services locally, driven by latency requirements, data sovereignty considerations, and the strategic desire to maintain domestic AI capabilities. Why compete in a small pond constrained by power and politics and you can swim in the ocean? nitter.net/RepzdaCanes/status/210…

Yo but can $META build a social networking component into Muse so that Nebio can go on a playdate with Peepeepoopoo? She can learn about $NBIS and he can learn about flipping burgers. I'll check his Soul.MD after to see if she's a good influence. @alexandr_wang @strictrope

The headlines are getting it all wrong. They leave out that Harvey then pivoted toward open-weight models, post-trained for specific legal workloads, and then build an economical production workflow $NBIS Token Factory was built around the expectation that this is where enterprise AI would go: specialized open models, optimized and served economically in production. Enterprise adoption of open models isn’t coming. IT'S ALREADY HERE. nitter.net/theredlinehq/status/21…

The writing is on the wall (all over my X feed today). $NBIS is going pedal-to-the-metal on the Agentic Stack. Remember, this layer sits even higher than @nebiustf and carries even richer margin economics. Only a few weeks ago during Goldman Fireside, Mark Boroditsky said: "We're going to be adding to our agentic capabilities. Right now we're delivering blueprints but we'll be building out more and more solutions to support agentic requirements." Just today we saw WEKA for shared KV cache / long-context agent memory, and now Comet Opik for agent observability, tracing and evals. The pieces are starting to snap together. And if I know one thing, it’s that Nebius is relentless about taking customer feedback, iterating quickly, and bringing new products to market to continuously improve its stack. AI Cloud → Token Factory → Agentic AI Stack nitter.net/Cometml/status/2103209…

$NBIS Update 🚨 Nebius is now pushing KV cache beyond GPU HBM into shared infrastructure. Together with WEKA, $NBIS benchmarked a distributed KV cache on HGX B300s that lets agents reuse context even after it gets kicked out of GPU memory. Results are pretty 🥜 1. 93% cache hit rate vs 40% on HBM alone - More requests can reuse existing context instead of recomputing it. 2. 3.7x more output tokens / GPU-hour - Each GPU gets significantly more useful work done. 3. ~5x lower mean Time To First Token (TTFT) - The model starts responding much faster. This matters a ton for agentic workloads. Agents keep dragging around their massive, ever-growing swollen nutsacks of context. Without a shared cache, once that context gets kicked out of HBM, you are back to recomputing the same shit all over again. Less recompute. More throughput. Better economics from the same GPUs. nitter.net/nebiusai/status/210319…

You shouldn't treat ClusterMAX 3.0 as the Magna Carta. But it's equally important not to minimize the rigorous technical benchmarks underpinning it. $NBIS’s ClusterMAX 1.0 → 3.0 arc is more than just a rank 🆙 story. The real story is the engineering-first DNA that can't be bought with capital, power, or silicon. $NBIS is an organization that repeatedly and proactively absorbs product and engineering feedback from SemiAnalysis, and by extension its customers, ships against it, and compounds those improvements. This is rarer than it sounds in an industry increasingly crowded with finance-first operators and repurposed bitcoin mining outfits. ClusterMAX 1.0 (🧈 Gold): Praised $NBIS’s robust technical capabilities. ODM hardware and fully managed Kubernetes offering. 1. No fully automated managed Slurm - SemiAnalysis called this a “significant gap.” 2. Health checks and monitoring lagged the leaders - More active testing, better dashboards, and stronger automated lifecycle management still needed. 3. Weak UI/UX Product still complex and unintuitive, with adoption lagging despite aggressive pricing. ClusterMAX 2.0 (🧈 Gold): 1. Managed Slurm matured - Soperator shipped, taking $NBIS from a “significant gap” to a fully functioning self-service Slurm product with expected NCCL/PyTorch performance. 2. Health + monitoring improved - Major progress in checks, dashboards, and storage, though visibility/notifications still needed work 3. Architecture validated - KubeVirt stack delivers bare-metal-class performance, while storage architecture materially improved provisioning and startup times. ClusterMAX 3.0 (💠 Platinum): 1. Soperator becomes differentiator - Weakness in 1.0 has turned into production Slurm-on-Kubernetes stack adopted by other cloud providers including Voltage Park and GCORE. 2. Health + monitoring become strengths - SemiAnalysis deliberately induced a GB300 failure. $NBIS detected immediately and automatically returned the node to service. - Dashboards and visibility criticized in 1.0 had matured into what SemiAnalysis called an “excellent suite of dashboards.” 3. Storage feedback loop plays out in real time - SemiAnalysis found pathological storage behavior. $NBIS immediately took feedback, retuned the filesystem, and the retest eliminated the errors while pushing performance above median. 4. Commercial validation arrives - SemiAnalysis describes $NBIS as having “strong offerings in every category” and the ability to “command a significant price premium,” while serving many happy short-term customers.

Mood after $NBIS tops out ClusterMAX 3.0 and is the only green in a sea of red today Disclaimer: Any resemblance between myself and $NBIS CTO Danila Shtan exists solely within my own delusions of grandeur. Photo Creds: @itsalasdairmann

🚨 $NBIS Update BNP Paribas just delivered another wicked PT increase: $260 → $399 👻 This comes off the back of ClusterMAX 3.0, another round of GPU price hikes, MLPerf wins, dynamic spot pricing, and the Palantir partnership. Quite the stretch of validation for $NBIS.

🔊PSA for anyone looking at the Nscale IPO S-1 markets $103.4B backlog as a huge selling point. DO NOT BE FOOLED❗️❗️❗️ If you flex your dendrites a little, you’ll see that this backlog is less a growth pipeline and more a cosmic-scale ceiling on future monetization. 97.5% remains undelivered. Pricing is largely fixed at 2025–2026 bare metal rates of roughly $13M/MW/YR. Average contract value spans 5.7 years. $43.8B is tied to Microsoft through 2033. Another $44.6B is tied to Anthropic at Monarch. The first Rubin tranches do not arrive until late 2027. In other words, much of the next five years is already spoken for. The monetization path has largely been written in advance. This is a perpetual bare metal wholesale infrastructure story. Value it accordingly.

“IREN continues to come up in our conversations with neocloud customers, and usually for the wrong reasons.” — SemiAnalysis, ClusterMAX 3.0 What's interesting though is $IREN recently signed compute deals in a similar range to $NBIS and $CRWV. The implication is that scarcity is doing the heavy lifting. Customers want the GPUs badly enough that they are willing to be optimistic about infrastructure reliability, tolerate a weaker managed cluster layer (AI Cloud), and trust their own engineers to brute-force through the gaps. 1/ Realistically, SemiAnalysis points out that $IREN’s physical infrastructure is improving materially at newer builds like Childress and Sweetwater. 2/ Additionally, $NVDA’s DSX/NCP reference architecture should help $IREN close the datacenter-design gap over time while raising the technical quality of its AI Cloud stack. Remember, the scope of ClusterMAX’s rankings is managed clusters. It is grading the AI Cloud layer above bare metal: cluster provisioning, Kubernetes/Slurm orchestration, scheduling, networking, storage, monitoring, health management, support, and related managed services. So an Underperforming rating does not mean $IREN’s GPUs are uncompetitive or that customers do not want the compute. It means the managed product wrapped around that compute is still materially weaker than the leading players. The real question is whether scarcity lasts longer than it takes $IREN to close the gap. As long as customers are fighting for access to GPUs, the compute can still clear at strong prices despite a weaker managed layer.

I will allow this. $NBIS finally joins $CRWV in its rightful position as the only other AI cloud in ClusterMAX 3.0’s Platinum Tier.