Stephen, Every Tree Counts

I manage a Facebook group for Nebius Investor, facebook.com/groups/2355224… feel free to join, we have 750 members as of Jan-2026

@waqasmakhdum @nebiusai @mastra @nebiustf @calcsam @SherryYanJiang @swyx @aiDotEngineer Singapore is the best place to deploy food delivery from NBIS's AVride bot. Also Robot Taxi will work best. I want to see Nebius pushes that direction.

@nebiusai will be at AIE Singapore, and yes, we’re bringing robots! also, will show how @mastra + @nebiustf enable agent builder use cases in @calcsam's talk Thanks to @SherryYanJiang for the opportunity & shoutout to @swyx for taking @aiDotEngineer global See you there 🇸🇬!

@daniel_koss Didn't you sold out a lot from 75% allocation to 25% allocation? I this was someone else? Sell low Buy high? Sorry if this is not you.

Added a new 30% position today! Been obsessed with this company the entire weekend. Listened to earnings calls and industry reports. Think it can 10x easily. Haven't been this excited about the risk reward of a stock since I found $NBIS at $40. Dropped the ticker for Subs, but will share the full thesis for free tomorrow. I think it's another name that you'll soon see EVERYONE on X talking about. Just like nobody owned $OUST and now everyone loves it (rightfully so!).

@TopTechStocksUS We need to add EiGen.ai now, thanks.

$NBIS One pager for all the X users who still have not discovered this beauty

@EndicottInvests Where is the beef, no news, why up so much? AI21 acquisition perhaps?

$NBIS ALL TIME HIGH

@EndicottInvests Clickhouse competitor DataDog is $50B market cap.

$NBIS is only a 40B company that will be doing 8B+ in revenue in 2027 Their Clickhouse stake will be worth 6-10B too

@HyperTechInvest Jun-11-2025 youtube.com/watch?v=yIVz0N… I missed this somehow.

OpenAI is transitioning away from Datadog and into ClickHouse services, where $NBIS holds a 25% stake

@EndicottInvests Is it an anecdote, or it was announced?

Clickhouse might hit a $7B ARR before $NBIS at this rate And $NBIS owns a 1/4 of it

@MB_Hogan @daniel_koss Are you talking about Ai21.com?

$NBIS THOUGHTS: I've been pondering what Roman Chernin just days ago told @daniel_koss in an interview, an anecdote that Nebius was in talks with a company that is highly competent at running inference and Nebius shared a blueprint for how to replicate their own superior results and this company still couldn't do it and ended up opting to become a Nebius customer instead. My question is why on earth would Nebius give a competitor a blueprint or even the slightest hint at how to achieve superior margin on inference, the very thing they are trying to build into a moat-backed asset? My best guess here is that Nebius wanted to teach this company a lesson, that even if they were given the instructions for how to do what Nebius does, they would never be able to do it without the level of technical capabilities and engineering that Nebius has developed in-house. In essence the difficulty of achieving what Nebius has achieved must be so great that even a custom roadmap only goes to illustrate the futility of trying to compete. Thoughts?

@AlexfromBabylon @jiahanjimliu Alex, do you mean Fireworks-AI can not compete with Eigen-ai + TF ? or the other way around, thanks, not very clear the way I read yours.

@jiahanjimliu At the pace Nebius TF is moving I think it will be hard to compete for Fireworks AI longer term, especially if they also integrate AI21labs.

Credit Where Deserve: $NBIS acquires Eigen AI Eigen AI, founded by 3 MIT PhDs with strong background in ML Inference, is strong acquisition for $NBIS at 643M. Whereas, I didn't see how Tavily was worth 275M, I see Eigen AI being worth 1B. Eigen AI is competitive with FireworksAI and Hyperscaler for Open Source models - differences in benchmark cut-off time vs releases. Close sourced models are a different story as no one is going to out optimize Anthropic/OpenAI inference stack for Claude/GPT even if they were open sourced. Eigen AI is a much better acquisition than Tavily because inference stack is a full product offering while agentic search is an feature. There's alot of misunderstanding" Tavily does not search the web and does not replace Google search. Given a set of websites, Tavily extracts the page into a form efficiently parsed by AI. Eigen Factory: Talent for Token Factory Alot of $NBIS investors did not like how I said Token Factory was not competitive, but it's clear today that $NBIS management would not drop 643M on a inference optimization if Token Factory was within organic improvement distance to Eigen AI given that Eigen AI was already available on Nebius Cloud through partnership. The reason why $NBIS acquired Eigen AI was so that they can have the team in house to work on development of Token Factory. What does this mean for other Neoclouds like $IREN? Eigen AI was founded in 2025 and one year later it's already being acquired for 643m. AI software is moving fast and will only move faster - prices may compress and it's uncertain how many leading enterprises will be able to move inference stack in house. For $NBIS who is betting on open sourced and becoming a full AI cloud, Eigen AI is a strong synergy to their strategy. For $IREN who is betting on closed sourced frontier models with in-house inference stacks and leading enterprises moving to capture inference margins as software by itself becomes commoditized, Eigen AI would be a pivot. Would $IREN need to pivot in 1-2 years? Possibly. I'm comfortable betting on closed sourced frontier models and leading enterprises wanting to combine inference + software because software alone is no longer a moat.

@daniel_koss You have been lucky, not that you actually knew.

Seems like my timing of selling $NBIS at $168 and buying it back at $136 was pretty good? Funny, because when I announced exactly what I was gonna do and my reasoning, everyone told me I'm super stupid and trading never works 🙄 That being said, this is obviously the right stock to just hold for 3-5 years! The constant flow of great news is not good marketing, it's just the result if great talents doing what they do best: execution.

@SmallCapSnipa Birmingham residents rally is very soft and nice, not like those in our state CA. Go Nebius, already approved within the laws.

$NBIS: Birmingham residents rallied against Nebius’ AI data center project again this week The city passed a 6-month moratorium on new data center projects, but Nebius will proceed uninterrupted as their permits were filed before the moratorium It looks like the company will get their project despite the heavy pushback

@romanchernin @Eigen_AI_Labs OMG it is here youtu.be/jw_o0xr8MWU?t=…

That's how we made a decision to acquire @Eigen_AI_Labs (kidding, but...:)

@EndicottInvests Nothing happening, this is normal.

Uhhhh $NBIS What’s happening lol

@EndicottInvests Now you have to speak up why you will be quiet? Please.

I’ll be quiet today

@EndicottInvests My guess is, there are hidden news and developments leaked to some minority exclusively. When will the ER be, I am too anxious to listen to the ER call.

$NBIS WANTS $170 🔥🔥

@EndicottInvests 25%, please take notes.

Every month I realize how much of $NBIS's subsidiary value is still not being priced in. ClickHouse alone, where Nebius holds ~28%,is tracking toward a $10B+ valuation in the next 18 months. Then you've got Avride. That's another multi-billion dollar asset sitting inside the same ticker. Can't get more bullish. Long $NBIS

@RuebenJay @ilzmcfly Too much supporting infrastructure to get connected to internet and such. That is whey they selected cities instead of forest.

@ilzmcfly It’s almost like building data centres is hard. Even more so when you’re building them in residential neighbourhoods. It’s also going to get harder as more communities stand up to prevent them being next door. Give me 1.4GW in the desert any day!

$NBIS might be COOKED, there is a high probability that the NJ Facility will be delayed for 12-24 months due to air permit issues. (All public Information) There is plenty of concrete evidence that the facility is not operational, let's go through the evidence below. But before we start, $NBIS has never confirmed that the first tranche to Microsoft was delivered from the NJ facility, and in the contract it says that Nebius can provide alternate capacity. The Exact Quote From the "COMMERCIAL AGREEMENT WITH MICROSOFT" section of the Form 6-K: "In the event that, following a grace period, Company [Nebius] FAILS to To meet agreed delivery dates for a GPU Service AND the Company CANNOT PROVIDE alternative capacity, Microsoft has the right to terminate that GPU Service." So Nebius can provide alternate capacity to meet deadlines. I also checked through the Q3 ER and Q4 ER + Reponse to FT with Grok to make sure that Nebius has never said that they delivered the first tranche from the New Jersey Facility. Nebius has not made a single explicit, verbatim public statement like "The first tranche was delivered from our Vineland, New Jersey facility on [date]." Here is the link to the grok chat ---> x.com/i/grok/share/6… So why would we think $NBIS will provide capacity from the Finland Datacenter and not the NJ Facility. Let's start with Finland; recently the Finland datacenter hads been expanded from 50MW to 75MW earlier this year, Setting up a cluster in an existing, power-connected hall takes weeks. Enough time and space to deliver the first tranche without missing the first tranch deadline. With the ability to provide capacity to from other sites then this is possible. I believe this is how they were able to deliver the first Tranche 1 to Microsoft. Now lets move on to why they had to deliver the first tranch from Finland --- >The NJ Facility is not operational and delayed They have not received Air Permits for their gas engines that were filed as pre construction speed process on 12/16/2025 Can they operate the datacenter without these generators? ------------------------------------------------------ Nebius cannot operate the facility at the scale required by the Microsoft contract using only the local power grid. According to Gemini they can use the local NJ grid to power 15% of the facility (45MW) which is where they most likely delivered the other part of the first tranch. But that leaves 255MW sitting empty and delayed. It seems they are using the local grid in NJ to keep the building alive and run small tests, while they use Finland to satisfy the actual heavy-duty compute requirements of the first tranche until those NJ air permits are signed. But this would not check out with the recent Thermal Image analysis of the nebius data centre that show the cooling units installed are not remitting heat. The NJDEP who is the gatekeeper of these permits has issued TWO TDC Warnings in 2026 regarding its filing permit. What are those warnings and how serious is this issue. ------------------------------------------------------ Warning 1: On 01/06/2026, the DEP issued the first Notice of Technical Deficiency. DataOne responded on 01/20/2026, but it wasn't enough. Warning 2: On 01/30/2026, the DEP issued a second Notice of Technical Deficiency. This is the "Red Alert" because it shows the DEP rejected the company's first attempt to fix their data. The New Administrative Failure: Most recently, on 03/30/2026, the DEP issued a new Notice of Administrative Deficiency. This suggests that not only is the "science" (technical) wrong, but the legal/paperwork side of the application is now falling apart as well. Why did they issue these warnings? ------------------------------------------------------ The DEP is questioning if the 32 gas engines and the Data Center itself must be treated as a Single Stationary Source. If they are ruled a Single Source, the facility likely exceeds "Minor Source" limits and must apply for a Title V Major Source permit, which involves federal EPA oversight and years of delay. ------------------------------------------------------ 1. Rejection of Theoretical Data (Science Warning) Regulators found that DataOne’s application relied on theoretical manufacturer models for the 32 Bergen engines rather than real-world, NJ-specific data. - The Problem: The DEP characterized the massive 300MW configuration as a "proof of concept" that hasn't been tested at this scale. - The Demand: The Jan 6 and Jan 30 warnings required DataOne to provide a Reconciliation Report proving the engines would actually meet emissions limits under local conditions. 2. The "Single Source" Inquiry (Regulatory Warning) A major reason for the Jan 30 Technical Deficiency was the DEP's investigation into "Common Control". - The Conflict: The DEP is questioning if the 32 gas engines and the Data Center itself must be treated as a Single Stationary Source. - The Stakes: If they are ruled a Single Source, the facility likely exceeds "Minor Source" limits and must apply for a Title V Major Source permit, which involves federal EPA oversight and years of delay. 3. Community & Environmental Justice (Legal Warning) The March 30 Administrative Deficiency and previous technical stops were influenced by new regulatory standards and heavy local pushback. - Noise and Air Blight: Residents organized formal complaints regarding the impact of the "cruise ship engines" on the nearby community and protected Pinelands boundary. - Compliance: The DEP demanded a more rigorous Health Risk Minimization Plan to address cumulative stressors like noise and acrolein emissions. @BagelC47 found something interesting about the gas gen's in the Dataone PCP Document available on the NJDEP website, this information basically shows what are the actual real world implicancy of these gas gens and would they pass the test: ------------------------------------------------------ 1. The "713 Million Cubic Feet" Problem The 713.10 MMft³/yr figure is the estimated maximum amount of fuel burned per engine. - The Total Scale: With 32 engines planned, the facility is looking to burn roughly 22.8 Billion cubic feet of natural gas annually at full capacity. - The Comparison: To put that in perspective, that is enough gas to power roughly 300,000 average New Jersey homes for a year. - The Regulatory Trigger: Burning this volume of gas in a single residential-adjacent location is what triggered the DEP to investigate the facility as a Single Stationary Source. 2. The NOx Threshold: "3x Over the Limit" New Jersey has some of the strictest air quality standards in the US because it is in a "Non-attainment zone" for ozone. - The Limit: To remain a "Minor Source," the entire facility must stay under 25 tons of NOx per year. - The Reality: As shown in your friend's math, even with 90% Selective Catalytic Reduction (SCR) removal, the facility is projected to emit roughly 73 tons of NOx/yr. - The Conflict: If the DEP rules that all 32 engines must be counted together, the project is 300% over the limit to be considered a minor source. 3. The "Title V" Trap The screenshot in shows Section 7 of the permit, which explicitly states that if the facility's Potential-to-Emit (PTE) exceeds thresholds like 25 tons/yr of NOx or 100 tons/yr of CO, the permittee is required to submit a Title V Operating Permit application. - Current Status: DataOne applied for a Preconstruction Permit, but the DEP’s Technical Deficiencies indicate they believe the project already triggers Title V requirements based on the gas volume math. 4. Likelihood of Passing NJDEP Tests The likelihood of passing under the current "Minor Source" application is Low to Very Low for the following reasons: - Scientific Skepticism: The DEP has already issued two Technical Deficiencies. This means they have already looked at the "713 million cubic feet" math and essentially told the company: "We don't believe your scrubbers can keep you under the 25-ton limit at this volume". - The "Proof of Concept" Label: Regulators have called this a "proof of concept". In NJ, regulators rarely approve a "proof of concept" that sits on the "knife's edge" of legal pollution limits. - Environmental Justice (S232): Because the site is near a residential area, the Environmental Justice Law allows the DEP to deny the permit even if the math was "perfect," simply because the absolute volume of residuals exiting 32 stacks is too high for that specific community. The NJDEP is effectively forcing Nebius/DataOne to admit they are a Major Source, which would trigger the 12–24 month Title V process mentioned in the tweets Why did they file for this so late by end december 2025, when they started construction in early 2025? ------------------------------------------------------ 1. The "Grandfathering" Miscalculation The primary reason for the late filing was an attempt to beat the March 2026 regulatory shift. - Legislative Pressure: New Jersey introduced Bills S3379 and SR18 in early 2026, which mandate much stricter energy reporting and lower emissions for data centers. - The Strategy: By filing in December 2025, DataOne hoped to lock in their "Minor Source" status under 2025 rules. - The Failure: Because the NJDEP issued multiple Technical Deficiencies (TDs) starting in January 2026, the application was never finalized. This effectively stripped away any "grandfathered" protection, as the project is now being scrutinized under the updated Environmental Justice Law (S232) standards. 2. Late Finalization of "Islanded Power" Engineering The site’s reliance on 32 Bergen gas engines for 300MW of internal power was a novel "off-grid" approach designed to bypass the multi-year wait times for the PJM regional power grid. - Engineering Lag: Designing a system that uses 32 cruise-ship-sized engines to power high-density AI racks is technically complex. - The "Proof of Concept" Issue: The DEP eventually labeled this a "proof of concept" that hadn't been tested at this scale. DataOne likely spent most of 2025 trying to produce "theoretical math" that they hoped would satisfy regulators before they were forced to file. 3. The "Concrete First, Permits Later" Gamble Management chose to prioritize physical construction speed over regulatory certainty. - Aggressive Build: Construction began in early 2025 "behind concrete walls" to meet a goal of delivering the first phase in just 20 weeks. - Regulatory Avoidance: Filing early in 2025 would have invited DEP inspectors to the site during the most sensitive early phases of construction. They waited until the building was nearly a "shell-complete" asset before filing the air permit (PCP250002) required to actually turn it on. What happens if they cant get approval for the pre-construction air-permit? ------------------------------------------------------ There is 32 Days Remaining – The number of days remaining for DEP to make a decision based on the Days Allotted. If the Department continues to find deficiencies, they can keep the project in this "Stopped" state indefinitely or move to a final denial. DataOne can appeal a denial by requesting, An appeal process in New Jersey typically takes 12 to 24 months to resolve through the court system.. OR If the NJDEP concludes that the 32 gas engines and the data center constitute a Single Stationary Source, the cumulative emissions will likely exceed the "Minor Source" thresholds. They would need to re-file for a Title V permit under the Clean Air Act. Between the new engineering studies, EPA review cycles, and public sessions, this path typically takes 18 to 24 months to complete. I have asked @nebiusai and DataOne CEO @CABeyney to comment multiple times but there is no response. Also the ceo of DataOne has never confirmed the delivery of the first tranche to microsoft from the New Jersey Data One facility, I would expect the CEO to mark the first MAJOR milestone. I think with the evidence we have, we can conclusivly say that the NJ Facility is delayed and not largely not operational... For how long who knows. But its looking like the Permits may not come in hand this year. But this is the problem Nebius has, relying on third party builders and having no control of the buildout. This may have set them back by 1-2 years. Now they have Shells sitting empty waiting for power approval. All this information I got is public and can be searched here: njems.nj.gov/DataMiner Receipts attached nfa

@daniel_koss it will naturally take a little breather, and continue the rise beyond real execution success. We do not know the line of reality vs perception, so I will not bet on timing. Hence no option, only invest in long term value bullish thesis according to Nebius's layout plans.

Will finally have time to continue my $NBIS deep dive this week! Please don't get confused: Yes, the stock had a massive run recently. No, it's not overvalued. Yes, there's still massive upside from here. I see $OUST as one of the stocks having the highest short-term (next 30-50 days) upside, but in my current models $NBIS actually has higher long-term upside (next 200 days). Reason for that is that I think Ouster is underdiscoverd (yes FinX now knows it, but it's still a tiny company and it feels like money will only start flowing in bigger volumes now...) and was very oversold. Still is too cheap for the coming wave of physical AI, which is MASSIVE. But the underlying trend of the entire world running into an AI compute shortage that Nebius benefits from is just so much bigger and more aggressive right NOW. That creates a monster demand wave that also increases pricing power and directly multiplies not just slightly improves margins. That means re-rating time in a big way. In the next few episodes I'll share EXTENSIVE data and analysis on: Why the stock is starting to run (spoiler: it's NOT just the $META deal and $NVDA investment. Much more data is coming in that confirms the thesis is playing out exactly like many smart AI bulls predicted). Revenue and margin targets. Price targets across bearish to bullish scenarios, plus my personal base case that I use to plan, invest, and size potential LEAPs.

@EndicottInvests You should seriously consider becoming a equity analyst, focus in what you know best, the high tech and AI etc. Good wish to your continuous successes.

Officially put in my two weeks Have a sweet new job lined up in a few weeks Life is good!

@daniel_koss Will pure baremetal make any profit, see SA article on NBIS hold discussion

Update on my $IREN view Changed: - I'm way more certain that a pure Infrastructure business "baremetall" can actually become insanely attractive for 5+ years. I do think that the view that pure infrastructure will become a commodity is WRONG. Talent, GPUs, capital, energy, etc. arr just too scarce for this bottleneck to go away in the near term future. The resulting pricing power means calling the physical infrastructure a commodity is just wrong if you define commodity as sth that purely competes on price. Same as before: I totally get why $IREN investors are so excited about its position. The assets and strategic position in market look freaking perfect. Here are my 2 concerns: 1) I don't think they can execute nearly at the level of $NBIS meaning I think their AI factories have worse uptime, lower efficiency etc. How can we KNOW instead of debate if this will turn out to be true or not? Simply measure those metrics. 2) I also think they have significantly less qualified talent across all the 4 other layers of the "AI cake". Meaning they will offer a less vertically integrated service and have much lower margins. Now the important part: We will see if that's true or not. They should deliver AI factories at meaningful scale this year. If they prove me wrong and get comparable margins to $NBIS on the same amount of capacity I will publicly share that I was wrong and even happily take a position at a higher price! Even if they can NOT match Nebius, the raw compute business alone is turning out to be so much more interesting than anyone thought that IREN could still crush it. I just see so much upside in Nebius over next 3-5 years that I don't want to invest in IREN until they falsify all my skeptical points very clearly so that I KNOW my money is invested with the idea risk reward, not just high upside IF they perform.