Incredible raised ARR guidance from Nebius today - from 1B ARR (run rate revenue) at the end of 2025 now raised to between 7-9B ARR guidance at the end of 2026! That is a huge increase. They also announced Meta as a new customer (3B over 5 years - and I am sure this will increase). Revenue growth this Q was up 355% YoY. Their CEO, Arkady Volozh, said that it would have been more if capacity had not been sold out.
Record growth: Revenue surged 355% YoY and 39% QoQ to $146.1 million
Adjusted EBITDA loss narrowed sharply to $5.2 million, an 89% improvement YoY
The AI infrastructure segment drove ~90% of group revenue and achieved positive adjusted EBITDA.
Major customer wins: Signed two large AI infrastructure deals — Microsoft ($17–19B) and Meta ($3B over five years)
Run-rate outlook: Targeting $7–9B annualized run-rate revenue by end-2026, up from ~$551M at Q3-end.
Capacity expansion: Plans to reach 2.5GW contracted power and 800MW–1GW connected power by end-2026
Financing progress: Raised $4.3B in convertible notes and equity in September to fund growth
ATM equity program: Launching a 25M Class A share at-the-market program to support future expansion
AI cloud innovation: Released Nebius AI Cloud 3.0 “Aether”, featuring enterprise-grade security (SOC 2, HIPAA, ISO 27001)
New platform: Introduced Nebius Token Factory, an enterprise inference and model-optimization platform
Data center scale-up: Expanded in the U.S., UK, Finland, and Israel; new NVIDIA B200/B300 clusters deployed
Subsidiary progress:
Avride: Uber invested $375M; Dallas robotaxi launch set for Dec 2025; >215K deliveries YTD
Equity stakes: Retains significant holdings in Toloka (AI data solutions) and ClickHouse.
Arkady said “2025 has been a building year as we put in place the infrastructure and framework for future rapid growth. This year, we believe that we have successfully laid the foundations for an outstanding 2026 — a year that should firmly position us among the top AI cloud businesses globally. And at the same time, 2026 is still just the beginning”.
I continue to remain heavily invested in Nebius as my number 1 stock.
I was slightly concerned that they are completely sold out as it seems like they’ve reached a bottleneck. All the big contracts seem nice but with demand outpacing supply I wonder if those contracts will hurt their ability to increase price based on demand.
I also didn’t love the dilution. I’m left wondering how many times they’ll have to dilute to achieve their ambitious plans.
Management pointed out that they will remain dilution sensitive. I take this to mean that they will not release all 25m shares to market at once, but will do so at the appropriate time and as they need more capital.
There are currently 218m Class A shares, so this represents about 11% dilution if all 25m shares hit the market share once, which they won’t.
The cash raised will fund their blistering growth, so I’m not too concerned by this.
CapEx for 2025 will be $5B. Since the target ARR is ~$1B in 2025 and ~$8B in 2026, I’m guessing that CapEx in 2026 will be at the level of $30B,which means a lot more dilution or / and debts.
Nebius lowered their 2025 revenue guidance, because of the same reason as why CoreWeave lowered theirs. As we all know, Nebius’s colocation partner, DataOne, delayed the New Jersey data center project by 2 months. So execution risk of building data center is still high, as expected by the board members here.
Both Nebius and CoreWeave confirmed that the biggest bottleneck is the power shell. This bottleneck is industry-wide. This is bullish for $IREN. And, as an OT, this actually makes me think that $NVDA’s growth may be impacted in the near term as well, because the delay in data center constructions will then delay the delivery of GPUs.
According to Arkady, CapEx has 3 stages: securing land & power (stage 1) is only 1% of CapEx; building the data center (stage 2) is ~18-20% of CapEx, and the remaining 80% is for GPUs (stage 3). This seems to align with the cost structure of the Microsoft deal disclosed by $IREN as well.
Nebius wants to do more self-built data centers moving forward compared to co-location. They said they wanted to optimize margins.
There’s no doubt that the growth is extraordinary as Nebius is targeting 8X ARR in next year. Moving forward, I think the market’s biggest concern will be the heavy CapEx and margins of the business. This concern applies to $IREN as well, since the biggest cost is GPU. This is probably why $IREN only targets ~500MW active power by end of 2026 despite the 1GW+ power build-out. It looks like, compared to CoreWeave and Nebius, $IREN is taking a little more conservative approach.
Is this a viable business model? $30B capex ($24B GPU) for $8B revenue. GPUs can be used for 5 years, so that $24B will get $40B revenue in 5 years. But you’re paying capex upfront, so $24B is actually close to $30-35B using NPV. Add in non-GPU capex of $6B we get $36-41B in capital cost to get $40B revenue and that’s not including the other operating costs!
Also, why are the hyperscalers offloading datacenters when they have the cash to do it themselves. Could the reason be they’re uneconomical?
I have a small Nebius position, but the high capex makes me a bit nervous.
I think it’s important to note that these major customers - by far the lion’s share of Neibus’s revenue - are almost certainly interested in leasing raw metal, not in any software infrastructure that Neibus may provide. Meta and Microsoft both have their own software infrastructure that they’ll want to run and they’ll want the same infrastructure on all their machines, whether Neibus or CoreWeave or Iren or themselves are providing the hardware.
Neibus and CoreWeave and the other NeoClouds are literally doing deals to be suppliers to their competition (MicroSoft & Oracle), where their value-add is not lasting. Can someone explain to me how this works out for them in the medium to long term?
I’m not an investor in either of these companies but these deals really help the NeoClouds for multiple reasons.
They give them recognition for other developers
They provide them with net income to allow them to build up their more profitable full stack solutions.
They raise their share prices, which allows for easier raises in Capital
With locked in revenue it allows them to get better interest rates.
All four of those help in the medium and long term. It still has plenty of issues for the company to execute long term. But other than expanding the executive team focus area, I don’t see any downsides on basically taking large amounts of cheap cash.
I think it is becoming clear that NBIS will have to pay middlemen dearly for land, physical infrastructure, power, etc. It’s true that they have a very deep software bench, but the hyperscalers and more adept want bare metal and would prefer to side-step middlemen profit margins. This is a recipe for a profit squeeze.
Just to back this up from the transcript of the CC, Arkady was asked about the Meta deal (and also the Microsoft deal) and he had this to say:
However, these deals are important, these mega deals. It is important to stress that we will remain focused on developing our own AI cloud, which currently serves not only these big deals, but AI start-ups and enterprises. And ultimately, we believe that these large contracts provide us with great sourcing of financing for us to continue building our core AI cloud business.
These big deals are great - but it is not their ultimate focus.
Yes, this does not pass the smell test. I do not understand how it is possible that CAPEX scales at a 5:1 ratio to ARR. As you point out, the business won’t be viable.
CAPEX typically scales early, to build capacity, and as that capacity is filled it increases the efficiency and operating profits. If CAPEX just has to scale 5x more in dollars than ARR then shut the company down immediately.
Just to weigh in. I have been in $NBIS from around $30 earlier this year. My original reasoning for them is that they have the talent and capital to scale much quicker than all these crypto miners trying to transition. If you looked at them before the MSFT and META deals you saw a company with over 1.8 billion cash and other companies available to turn into capital (selling Clickhouse stake, AV ride, and Toloka, etc.). I went with them over a company like $IREN because at the time that $IREN pivoted they were making $150 million in quarterly revenue (bitcoin) and had like $450 million in cash vs. $NBIS’s warchest of capital (in my mind). I am starting to think that they all have to do the same thing to capitalize these build outs. I think the future is a little shaky until we see the margins they are able to get.
The only positive argument I have seen for $IREN has been the power deals they have established. I have not wanted to take a large position in them either because I think of it as a risk that they can’t perform in the end with the scale they are claiming without significant dilution or debt. Wondering about $NBIS as well regarding this at this point now. Long 10% $NBIS
I think you raise some valid points re the Capex of 30B and whether it is a viable business model or not. $30B in capex sounds crazy next to an ARR guide of $7-9B. But Nebius isn’t trying to be another SaaS company; they’re building the grid for the AI age.The GPUs are productive assets — they earn money 24/7 for years. Each cluster gets sold and re-sold through training, inference, and fine-tuning workloads.
The market is also insanely supply-constrained. Everyone — Meta, Microsoft, OpenAI, Anthropic — needs compute yesterday. Nebius locking up 2.5 GW of power and next-gen NVIDIA clusters basically gives them a scarce, revenue-producing asset base that the hyperscalers themselves are struggling to expand fast enough.
As for “why the hyperscalers don’t just build it,” they can’t move as fast. Power access, permits, and shareholder optics on capex intensity all slow them down. Partnering with Nebius lets them deploy capacity off-balance-sheet while still guaranteeing supply.
NBIS raised $4.3 B in convertibles and equity, plus they’re layering in asset-backed debt tied to those long-term contracts. That means Nebius doesn’t have to eat the whole $30 B from equity today — they can recycle capital as facilities go live.
If they really hit that $8–9 B run-rate with 30–40 % EBITDA margins by 2026, you’re talking $2.5–3.5 B in EBITDA.
TL;DR: yes, it’s capital-heavy — but that’s the moat. Nebius is building the “AI power company.” The question isn’t whether the model works today; it’s whether they can keep GPUs full and contracts flowing for five years. If they can, the payoff could be huge.
To be clear, @monkeydluffy was guessing at a number. We need to be mindful of extrapolating too much math from that guess. All of these companies need to prove there will be some sort of margin and profit reward at the end of the journey. IREN seems to be taking the approach of locking in location/power first and then scaling. Others seems to be inking scaling deals first and then locking in location/power.
In either case, we are a long way from seeing how this plays out. Two things are virtually guaranteed though as we go: 1) these companies will need to dilute further and 2) converting all that ARR to real revenue will be heavily dependent on meeting build out timelines. Given the constraints we are already seeing, I’d expect build out delays to be part of the gig no matter what name(s) you own (including NVDA at the top).
Yes, I was aware that this figure was a guesstimate. The company have guided for 5B Capex so far. But the figure of 30B may not be too far from the end figure based on nearly 1B capex for this past Q, and on their stated plans to have 2.5GW contracted power by the end of 2026. YTD they have spent 2B to date on Capex.
According to ChatGPT…. Analysts and investors often estimate total capex by multiplying:
the target power capacity (2.5 GW), and
a cost per megawatt typical for AI-optimized data centers (between $10 million – $15 million/MW including GPUs, networking, and buildings).
The speed and power efficiency of GPUs are evolving rapidly. No-one wants Nvidia A100s today, for instance, and if you have them, it may not even worth the energy and maintenance cost to run them. Nvidia first started shipping them in mid-2020 and EOL’d them in Jan 2024 - so active for only 3.5 years.
This is an interesting article:
Here is the puzzle: the chips at the heart of the infrastructure buildout have a useful lifespan of one to three years due to rapid technological obsolescence and physical wear, but companies depreciate them over five to six years. In other words, they spread out the cost of their massive capital investments over a longer period than the facts warrant—what The Economist has referred to as the “$4trn accounting puzzle at the heart of the AI cloud.”
–
Wait a minute - the deals with Microsoft & Meta aren’t deals with Nvidia nor with electric companies nor data center builders. This recent article says:
By the end of this year, the company expects to secure 220 megawatts of connected power, either active or ready for GPU deployment. This expansion includes major data centers in New Jersey and Finland. Two additional greenfield data center projects in the US are nearing completion, signaling the company’s commitment to scale aggressively and meet the surge in AI compute demand. Overall, Nebius is on track to secure more than 1 gigawatt of power by the end of 2026.
“Last quarter, we guided for 1GW of contracted power by the end of 2026. We are currently in the process of securing additional sites that would bring our total contracted power (i.e., secured land with power) to approximately 2.5GW by the end of 2026,” Volozh said. “Of this contracted amount, we expect to have 800MW to 1GW of connected power (i.e, power connected to built data centers) by the end of 2026. “
As for my supplying the competition concern:
That’s close to lying spin by the CEO in my view, using the word “serves” in a way that falsely implies that these “mega deals” companies are using Nebius’ software. They’re not. This is all about hardware build-out. And for that, there’s no moat, no lock-in: Microsoft and Meta can - and already are - contracting with other companies to provide them with the AI compute they want. Does anyone here really think Microsoft is going to offer Nebius-specific features within their Azure AI Cloud? Not I (although that would be an actual a big deal if it happened).
What’s the expected profit margin guidance on these deals? What is it costing Nebius to finance and stand these data centers up, to run them with 24/7 SLAs, etc. and what are they earning? I couldn’t find any profit margin guidance on these deals, but I admit I haven’t looked deeply.
Those are nothing in my book. There’s no brand cache among developers and we don’t know the net income from these mega deals. The 5-year length of these contracts means the servers themselves are basically useless at the end of the contract (no-one wants to run A100s today). There are even questions as to whether the data centers built today will handle the server needs of the future. We’re finding out that data centers built just a few years ago need expensive retro-fitting for power and liquid cooling to handle today’s Blackwells.
How does any of this really support Nebius’ build out of their own full stack solutions - the software difference that will, supposedly, make AI customers want to choose a NeoCloud instead of legacy Cloud providers?
This is Nebius (and CoreWeave) literally taking a commission to support their competition. I’m struggling to find a similar previous situation. The closest I can think of is Tesla supplying drivetrains for Toyota and Mercedes. Even there, Tesla wanted to get out of those contracts as quickly as possible since they diverted resources from their own efforts to build their own car, the Model S. And that’s a bigger worry for the NeoClouds here, due to the length and size of these contracts.
Yes, but the operative word there is “contracted,” as the CEO pointed out in the recent ER call:
"As we look to 2026, we expect our contracted power to grow to 2.5 gigawatt contracted. This is up from the 1 gigawatt, which we discussed in our previous earnings call in August. Furthermore, we plan to have power connected to our data centers, which means fully built, of approximately 800 megawatt to 1 gigawatt by the end of 2026, by the end of next year.
[/quote]
I don’t know how to estimate the Capex required on non-fully built out power capacity.
Smorg asks the most important question which is; what is the expected gross margin? I believe OpenAI and Oracle have cast a cloud over the neoclouds. Oracle has pulled back ~25% the last month after a strong run up on the OpenAI deal. Mostly due to the analyst worry of the impact to the ORCL software higher gross margins +70%. Further, " OpenAI’s CFO, Sarah Friar, has floated the idea of the company exploring federal loan guarantees."
In our forecast I think we need to at least bifurcate the NBIS margins and revenue based upon the types of business; hyperscalar GPUaaS versus AIaaS (e.g. IaaS and SaaS). Maybe we get too excited over the exploding net income and ARR and what we have here is a fast growing commodity business. Here is a Gemini estimate for this kind of GM break out at Oracle that we might consider as an analog for NBIS.
Blockquote
Oracle Cloud Infrastructure (OCI) - AI Server Rental
Reported Short-Term Margin (GPU rental business): Internal documents reportedly showed that for the quarter ending August (fiscal Q1 2026), the gross margin for its AI cloud server-rental business (specifically GPU leasing) was around 14% . This low figure is attributed to the high cost of acquiring advanced chips (like those from NVIDIA) and the rapid expansion of data-center expenses before customer contracts go live and fully ramp up.
Oracle’s Projected Long-Term Margin (AI Infrastructure): Oracle’s management has publicly stated that they anticipate the actual gross margin across the entire life of a contract for large-scale AI infrastructure projects will be in the range of 30% to 40% (with some specific examples projecting up to 35% ).
2. Broader Cloud Segments
Cloud & Software Segment Margin: In the first quarter of fiscal year 2026, the Cloud & Software segment (which includes both Infrastructure as a Service - IaaS, and Software as a Service - SaaS) had a reported margin of approximately 60% before corporate allocations.
Traditional Software Business Margin: For contrast, Oracle’s traditional high-margin software business typically operates with an overall gross margin of around 70% .
In summary, the most cited figure for the highly capital-intensive, fast-growing AI infrastructure part of the cloud business has been 14% in the short term, but Oracle is guiding investors to a long-term contract gross margin of 30% to 40%. The broader Cloud and Software segment remains much higher, around 60% .
Blockquote
I believe that most of the NBIS revenue for 2026 will be what is called above as AI Server Rental. 14% margin is rather sad. CEO Arkady says he is targeting the Cloud and Software segment with his full stack. But albeit this business will take much longer to grow organically to become a significant factor in overall revenue and gross margin.
BTW NBIS is using a 4 year depreciation schedule versus CRWV at 6 years.
Agree with many of your points on why this industry looks attractive now. I’m just trying to see at a high level how the business model will work out..
$30B Upfront CapEx for $8B revenue over 5 years feels like an insurmountable problem on the face of it. Can you share your assumptions on how you get to $2.5-3.5B annual cash flow?
The other big issue is Nvidia releasing chips every year, so what is the real useful life of the GPUs - is it 5 years? Every year Neoclouds will be forced to upgrade their GPUs. If they need $30B next year, how much in 2027/28 and how can these small companies keep financing those?
Lastly, selling bare metal is as commoditized as it gets. We know commodity hardware companies have crappy margins, very cyclical and overall low quality businesses. What will make Neoclouds any different?
Just back of the napkin maths. If they achieve 9B revenue in 2 years, and they are expecting to be Ebitda positive by next year (so let’s hope for 40% Ebitda margin by 2027), then that will give Ebitda of 3.6B by then. (PS, I edited my post above to say EBITDA rather than cash flow to make it more clear).
Lots of assumptions here of course.
I really appreciate everyone’s comments and input on this. It is what makes Saul’s board so great. And it really makes me think and consider my holdings carefully.
The other thing worth noting, of course, is that they don’t have to spend the Capex all at once. I guess that will depend on market forces at the time. I trust management to make the right decisions to protect shareholders.
But from everything I’ve read, it’s looking like 2026 will continue to see a significant demand so it makes sense to expand as aggressively as they can now.