Criticisms on $NBIS : One of my main points in my bull thesis was that NBIS would move away from the commoditized GPU/hour dynamic and build products/move up the stack for higher margin workloads. And controlling the inference/agentic stack would become a moat. But right now I dont see that happening with Nebius. Because: 1) they are themselves heavily leaning on bare-metal deals 2)the software stack is already commoditizing pretty fast, not a long-term moat 3)NVIDIA's platform strategy structurally limits neoclouds' own customization/moat 1st) Bare-metal deals: We all know the Meta/Microsoft deals were for fundraising and not the long term goal. But the famous 4 deals signed last quarter were all bare-metal. Reflection, Cohere, the anon neolabs all have their own software and dont need to pay for Nebius'. The core part of Nebius' bull thesis is that "bare-metal pays well, but with the SaaS layer on top it pays much better". So why hand out scarce capacity for bare metal deals? If the management is talking about its software pricing power and how it can command higher pricing power in every investor conference, why still do bare-metal deals? 2nd) The inference-as-a-service sector is already commoditizing pretty fast and the competition is brutal. New players like $DOCN, $BTDR (yes even an ex-btc miner, 2nd image) are entering the market which is already cutthroat with PaaS companies like TogetherAI and Fireworks. In fact, Nebius introduced serverless inference on March, and right now even Bitdeer has its own serverless inference service. Last week, great news dropped about $NBIS reaching platinum-level on SemiAnalysis ClusterMAX. Which is amazing but how can someone argue "software is a moat" when ironically Nebius itself disproves that point by challenging Coreweave's SaaS capabilities. That tells you that no one is safe by just having "better software" and the gap can be closed. There's been a lot of talk about Nebius' upcoming agent-as-a-service product (which I did a write-up on in May) but it still hasn't been shipped yet. A successful launch here might show that Nebius can outrun commoditization by moving up layers quickly. But it would turn into a "escape the lava" type "climb upward to survive" game which is not a sustainable model at all. 3rd) NVIDIA designs their products so no individual neo gets too powerful: As all investors know, $NBIS designs and builds their own server racks for HGX GPUs. This helps save on margins and become better vertically integrated. On Hopper and Blackwells (HGXs) this worked pretty well. However as compute scales up and out, there is more demand for rack-scale solutions. And NVIDIA pushes neoclouds to buy prebuilt racks from OEMs like DELL, ASUS, Wiwynn etc. This eliminates Nebius' advantage of building own racks. Because NVIDIA pushes the NVL72 architecture on every neocloud, the rack is no longer a Nebius one, its always Nvidia's. This consequently makes compute the same everywhere, a commodity. If no neocloud can differentiate meaningfully, a GB300 rack is substantially the same whether you buy from Coreweave, Nebius, Nscale, Lambda etc. On the 4th image you can see a pretty generic tweet about NVIDIA's connectivity solution being adopted by Nebius. This is a small example but shows how NVIDIA likes to standardize everything, limiting Nebius' ability to differentiate. These were the parts that damaged by original bull thesis quoted below. I still think Nebius is the best neocloud but im not convinced how they will differentiate from other meaningfully and build a long-term platform moat.

$NBIS is down about 13% over the last 3 months. The good news that has come out in those 3 months makes your head spin. Eventually price catches up with news.

JUST IN: China considers permitting ByteDance and Alibaba to buy new Nvidia chips

Wedbush: Nvidia's IR sees elevated growth levels at least through 2030. $NVDA

Nice $NBIS note from BofA

$31.6 TRILLION. That’s PwC’s estimate for global data-centre capex through 2050. $IREN $CIFR $NBIS or $WULF, some might 5-10X and some 20X or more And PwC says POWER will be the decisive factor in where that capital flows. And the wild part? It doesn’t peak. Annual spending is projected to rise from $800B in 2026 to $1.8T/year by 2050. AI infrastructure isn’t a one-time buildout. It’s a multi-decade capex cycle. ⚡️

Total data center & AI investment is projected to reach $10.3 trillion from 2025 to 2032 ⚡️ $IREN $CLSK $CIFR $WULF $HUT $NBIS $CRWV $SLNH $BTDR $WYFI

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/

$NBIS: Although this is well known within the Nebius community, I still see many considering and modeling the Meta deal as standard hyperscaler backlog. Once again, the biggest portion is a backstop, exposed to rising prices and well above what you'd see in bare metal deals.

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! 🆙

Heading to GTC Berlin? Take 20% off your pass with code ref-spo-265118. See you on the show floor. Visit us at Booth 3013. Get it here: nebius.com/events/nvidia-gtc…

$NBIS Token Factory explained in 15 minutes So I decided to write a little explanation of what the main differentiator of $NBIS is, and that is the "Token Factory" they introduced in November last year I will try to simplify this so that every investor can understand it. If you are interested in more technical details, ask me in the comments, and I might be able to explain it. So if someone asked me how I would explain it in one sentence, I would say that Token Factory is an AI platform designed to simplify the deployment, management, and scaling of large language models (LLMs) and other generative AI systems. It is a production-grade inference platform that allows organizations (and it is especially helpful for smaller businesses) to use AI models without the complexity of managing the underlying infrastructure. Token Factory is built around inference - the process where you generate outputs from already trained models (for example, asking questions and getting answers from models like ChatGPT). Whenever you ask ChatGPT a question, request code generation, or any other task you have ever done, the model produces a sequence of tokens that are, at the end, transformed into text/document that you can read. In the earliest stages of AI, we just had models like ChatGPT 3, Claude 3, etc. You paid your subscription of $20, and you were able to prompt infinitely, but lately the scale of these prompts has increased heavily, and enterprise adoption of these models has led OpenAI, Anthropic, and others to shift from simple subscription pricing to price/token, meaning that each prompt and each task is priced differently. The cost of the token is increasing rapidly, with supply not being able to meet demand. This is why $NBIS came with Token Factory, which is basically an optimizer for generating these tokens as efficiently, reliably, and cost-effectively as possible. The name kind of explains itself there. Traditionally, companies that wanted to deploy large AI models had to acquire and manage expensive GPU hardware, configure inference servers, monitor performance, handle traffic spikes, and continuously try to optimize their deployments. This process required companies to have experts in mainly these two fields: 1) Cloud infrastructure 2) Machine learning operations (MLOps) It is quite difficult to obtain a skillful team in these areas, so Nebius decided to go and remove majority of this complexity by providing a managed service that handles not only infrastructure, but also scaling, monitoring, and deployment via Token Factory. Developers can now simply connect to the platform through an API and immediately begin using advanced AI models. So the key strength is the exposure for smaller enterprises to open-source foundation models without acquiring a whole team of experts. Organizations can access and deploy models from families such as Llama, Qwen, DeepSeek, and from the latest announcement also NVIDIA Nemotron. The platform has interfaces that are compatible with widely used AI APIs, making migration and intergration relatively straightforward for development teams. What I did not understand initially, was that Token Factory goes beyond basic inference, it supports the whole lifecycle of AI applications. Users can tune their models on proprietary data to create domain-specific assistants for many industries like finance, healthcare, law and many others. This opens new possibilities like "parameter-efficient fine-tuning", "post-training optimization" that enable companies to customize models without the cost of training it from scratch. There are other fancy applications like Retrieval-Augmented Generation (RAG), where you combine LLMs with external knowledge sources like documents. But I don’t want to bore you to death as I understand majority of investors reading this are not machine learnings experts, so let’s skip this technical part. However, one last major advantage that you should be able to understand about Token Factory is the ability to scale "automatically". When you create an application and demand starts increasing, you usually start running into high latency and capacity problems. Instead of you having to allocate new compute to your application, which takes time and it might cause some downtime for your servers which are costly, Token Factory platform dynamically allocates additional computing resources to maintain both low latency and high throughout. The important thing is that this works the opposite way as well. When demand decreases, resources are released, helping companies optimize costs. This elastic scaling allows Token Factory to attract both small pilot projects to large-scale production deployments serving thousands of users and more. Now that I finished this paragraph, I realize that I completely forgot about one more thing and that is what we call in business "Enterprise governance and security". Token Factory includes features such as role-based access control, team management, authentication integration, usage monitoring, centralized billing and many other things that help companies maintaining control over AI deployments while meeting operational and compliance requirements. To somehow summarize everything, think of Token Factory as the "AWS of AI" or more precisely "AWS of AI inference". Companies bring their applications, Nebius provides the infrastructure and models, and charges for the AI output generated. The more AI is used, the more valuable Token Factory becomes. It is really that simple. I spent more time than I initially wanted on researching Token Factory and its use cases, but it really helped me to understand that this is something that gives $NBIS an unfair advantage against others in the sector. You should really understand this part of their business if you are an investor, so I will gladly answer your questions. If you found this a valuable read, follow me for more. Thanks! (picture is from ChatGTP)