Nebius at the UBS conference today

Investment Analysis Clubs Saul’s Investing Discussions
21

It’s also worth pointing out that Cursor does not run exclusively on Nebius’s instrastructure. According to Real-world engineering challenges: building Cursor they quote Cursor’s co-founder as saying:

“We are very much a ‘cloud shop.’ We mostly rely on AWS and then Azure for inference. We also use several other, newer GPU clouds as well.” Most of the CPU infra runs on AWS. They also operate tens of thousands of NVIDIA H100 GPUs. A good part of GPUs run within Azure.

As for AI platform, all that’s mentioned is:

Terraform is what Cursor uses to manage infrastructure such as GPUs and virtual machines, like EC2 instances.

Nebius does supply Terraform but then so do most AI Cloud Service providers.

4 Likes
22

OK, fine. Here’s what Marc Boroditsky, Chief Revenue Officer for Nebius said in that event about the Microsoft deal:

It’s a very scaled single instance, massive cluster that they can be utilizing for any aspect of their business…And then they have the Azure business, which competes with us in being able to deliver compute in order to service end customers.

“competes with us” – How did Bert miss that?

Neil Doshi, VP & Head of Investor Relations, later says:

We are not just building a GPU-as-a-Service bare metal product, we are building a cloud business on top of that. That cloud enables us to actually drive better customer retention. We can charge more of a premium price for a premium product.

And all I’m adding is that the “GPU-as-a-Service bare metal product” business is 90% of their revenue. That makes it harder for them to focus on the higher margin software infrastucture business, which is no longer anywhere near close to their primary business by dollar volume.

13 Likes
23

Although, of course, after initial figuring out, the GPU-as-a-service business could be quite predictable and not require much attention compared to the revenue.

8 Likes
24

Yup, and it’s fine if that’s the business Nebius (and CoreWeave, etc.) want to be in - no problem with a safe and steady profit.

But, then we should rerate and value them as such.

11 Likes
25

I’ve found this discussion to be quite interesting – from my point of view (I have a position in Nebius, but not in any of the other data center companies) it has highlighted many of the risks to this investment that I had been kind of hoping away.

One of Nebius’ principal claims is that it is a more than a bare metal provider, that it has and continues to build out tools that make it much easier for companies to take advantage of AI. I’ve subscribed to their announcements and I get regular notices of such things as their token factory, etc., which suggest they have resources to devote to what they claim is their main purpose (I.e., not to be a simple supplier of GPUs).

I’m curious if any of the Nebius bears (not limited to @Smorgasbord1 ) have any thoughts as to what results in Nebius’ next quarterly report would lead them to turn positive (or at least neutral) on the company?

14 Likes
26

Cold mountain,

That’s an amazing question.

For me I think that depreciation is a cost of revenue for these hyper scalers. They are selling run time on a GPU and their hardware. So life time of the unit is being used to generate revenue.

Because that’s how I value these companies, I add 99 million dollars to their COGS which brings it around 142 million. With a revenue at 146 million gives a gross margin of 2.9%

For me to think of this company positively I would like their revenue to grow significantly faster than their depreciation and other costs of revenue. Because 2.9% gross margin leaves little room to grow the bottom.

Drew

18 Likes
27

Nebius depreciates GPUs over a 4-year period, however, in 2 to 3 years NVDA will release a new chip. Revenue from those GPUs will then drop off a cliff and will not be enough to cover the depreciation.

NBIS has the shortest depreciation period among AI providers, so that’s a positive at least.

3 Likes
28

The main thing to me would be getting clarity on how they’ll fund the capex and the cost. These neoclouds are not asset heavy but asset extreme businesses. And I feel there are more easier ways of making money, so many things have to go right for these investments to be successful. Saul’s board has found great success focusing on the exact opposite end - asset light models.

The amount of capex needed (not just Nebius) means it has to be primarily funded by debt. Equity markets simply can’t bridge the gap. The one advantage Nebius has vs. others is they can choose to sell the subsidiaries (AVRide might be the only one that commands a premium price).

Oracle is the company to watch to see how the whole industry evolves. Their CDS blowing up recently is a worrying sign. That says the debt markets might not be willing to fund without a high rate (which will almost certainly doom the business model).

The other option I guess is private credit, but they’re likely to be more expensive than bond markets. In short, if they can show a credible path to funding their capex needs for 2026 and beyond then I would be willing to look at them again.

12 Likes
29

I may have already written too much on Nebius, but I still think the market is misunderstanding what Nebius is actually trying to do.

The market pressure these past couple of weeks, which have been painful for me with both Nebius and Iren, shows me that these businesses are hard for the market to value.

Nebius isn’t a business trying to look pretty on near-term margins or quarterly EBITDA at the moment. They’re knowingly taking on risks that the big cloud players don’t want to carry themselves — mainly the risk that GPUs lose value fast and need constant replacement….depreciation.

The Microsoft partnership is a big thing. If owning and running this capacity made sense for them, they’d do it. Instead, they’re happy to let Nebius handle that headache.

The key thing people miss is that Nebius isn’t trying to win by being the cheapest GPU shop. That’s just the entry point. The real aim is to become the place companies go when building AI infrastructure internally is unrealistic. Everything from raw compute to networking, storage, training pipelines, and workload management, all handled in one stack. For most enterprises, governments, and even large tech platforms, doing this themselves is expensive, slow, and messy. Nebius is offering a way to outsource the entire problem.

Cheap pricing right now is part of the setup. They want their data centers full, their customers committed, and workloads deeply integrated early on. Once a customer is scaled and running production workloads, pricing power shifts away from GPU rates toward longer contracts, bundled services, and operational dependence. We’re already seeing this with Microsoft and Meta, and once those workloads are embedded, switching becomes very painful, like Bert showed with Cursor in his excellent article.

What underpins the whole story is the balance sheet. Nebius, unlike so called Debtweave, has the cash and manageable debt to keep spending through hardware write-downs while others are forced to slow down. That lets them continuously upgrade GPUs and keep adding capacity.

If they even come close to hitting their power and capacity targets, the current valuation doesn’t reflect the business they’re building.

They are wanting to position themselves as core AI infrastructure for anyone who doesn’t want to build and operate it themselves.

It seems Ive been a bit of a lonely voice recently in terms of the bull case, and I may be totally wrong of course. I am far from an expert, but even though it has dropped a lot recently (though rising a bit today so far) it is my number 2 holding, having fallen just behind APP in my condensed portfolio.

And even after all of the falls of the last 6 weeks or so I am still up 106% YTD. Much of that growth is down to Nebius, APP and Celestica.

Jonathan

24 Likes
30

I am far from an expert, but even though it has dropped a lot recently (though rising a bit today so far) it is my number 2 holding, having fallen just behind APP in my condensed portfolio.

I’m with you on this, for the record. It is not in my largest group of holdings, but it is substantial. I have found solid arguments from the bears, but I find nebius’ plans and execution so far to be very promising. Of course, only time will tell. I try to keep in mind that there are many ways to be a successful investor. Good luck to all!

9 Likes
31

The real aim is to become the place companies go when building AI infrastructure internally is unrealistic. Everything from raw compute to networking, storage, training pipelines, and workload management, all handled in one stack. For most enterprises, governments, and even large tech platforms, doing this themselves is expensive, slow, and messy. Nebius is offering a way to outsource the entire problem.

I’ve got feet on both sides of the fence reading this whole thread, but this resonates with me quite a bit. I don’t think anyone would argue that we’re in VERY early innings as far as new products/solutions powered by AI (from existing companies) let alone new companies altogether coming from a new wave of entrepreneurs that are super well versed in the technology. If Nebius is able to land a good chunk of those businesses to come, and even take a stake in some…that all seems very interesting.

9 Likes
32

Financing the Burn

Because FCF is negative, Nebius is filling the gap through:

  1. Equity Dilution: Implementing an At-The-Market (ATM) program for up to 25 million shares.
  2. Asset-Backed Debt: Using the high credit rating of its customers (Microsoft/Meta) to secure lower-cost loans against its future receivables.

My reading is that looking at Just 2026, we have $5B of debt on a NBIS EV of $20B (at today’s ~$80/share). And we have additionally about 25% dilution of stock.

I see the huge NBIS negative FCF as the cost of a “land grab” in the AI infrastructure space. We deal with these kinds of stocks every day on Saul’s board. NBIS reaffirmed their mid-term EBIT margin guidance of 20~30% and clarified that the Microsoft/Meta deals would be structured in that range as well. If NBIS can execute, this will be a strong validation of their business model. So, a key data point in 2026 is when does revenue with MSFT and META start to be realized? The switch from construction to operational is the key to start revenue recognition.

Go-Live Timeline

  • First Phase: The first capacity was scheduled to go live in the summer of 2025.
  • Microsoft Service Start: Under the terms of the $17.4 billion agreement, dedicated services for Microsoft are scheduled to begin in late 2025 (Q4).
  • Staged Rollout: The partnership is designed to scale in stages throughout 2026 and 2027. While the facility “goes live” this year, it will take until late 2026 to reach the first major power milestone (approx. 100 MW of the 300 MW total design capacity).

One question that many have is why are the hyper scalars going to NBIS, CRWV, IREN, etc.? Why not just do it themselves like they always have? Is it time to market? Or is it a less attractive, low margin business with a lot of debt? This has Wall Street and some of us on Saul’s board brain twisted. Clearly with Oracle, they do not have the cash flow to fund all of it. The margins are low with Platform as a Service (Oracle calls this ‘hard metal’) so the market is killing their stock right now. But the others such as MSFT, META, AMZN do have the massive cash flow and cash reserves. Microsoft has been clear it is speed to market to satisfy the demand. This is not like the Internet bust where dark fiber and data centers were built for expected demand that was slow to develop. The AI demand is real and it is here now. We have an arms race. To meet Microsoft’s urgent demand for GPU capacity, NBIS utilizes a modular construction approach in partnership with DataOne, aiming for a record-breaking 20-week construction cycle for the initial halls. Again:

Blockquote

NBIS reaffirmed their mid-term EBIT margin guidance of 20~30% and clarified that the Microsoft/Meta deals would be structured in that range as well.

Blockquote

Confidence with the NBIS executive and engineering teams has been high based upon historical performance. I think they can do it again. I am holding my NBIS stock tight but with a bulging eye.

-zane

10 Likes
33

Hi all, this will be my first post in many years. Thanks to everyone that posts!

We as investors are constantly assessing the risk of an investment and have our criteria and tools (quantitative, qualitative, portfolio goals etc) to decide where to put our money.

My 2 cents here is that when I review the Nebius thread

(1) Many of the posts seem to give no credit to management or board ability to assess risks and I believe ignoring or dismissing their ability severely weakens our own decision making. We seem to be saying that a management group that grew Dutch-registered Yandex into the largest independent tech company operating in Russia and then survived a necessary split due to Russian invasion of Ukraine (forcing Yandex to sell most of the company to a Russian consortium with the remaining portion becoming Nebius) hasn’t learned a few things about risk planning and financial assessments. Additionally, specific to MSFT and META deals: the level of large deal scrutiny in most organizations I’ve been with is substantial with each contract required to pass all kinds of hurdles and stress tests with many macro, market, legal and counterparty scenarios. Additionally, with debt needed for capex involved, lenders will also be scrutinizing and running their own assessments. There is a ton of execution risk and every company has black swan risk but this is a team that isn’t new to challenges or assessing and planning for them.

If you’re interested to learn more on what led to NBIS formation here is a link to a query I made of what happened to Yandex: https://www.google.com/search?q=what+happened+yandex&rlz=1C5CHFA_enCA818CA818&oq=what+happened+yandex&gs_lcrp=EgZjaHJvbWUyBggAEEUYOTIICAEQABgHGB4yCAgCEAAYBxgeMggIAxAAGAcYHjIICAQQABgHGB4yCAgFEAAYBxgeMggIBhAAGAcYHjIICAcQABgHGB4yCAgIEAAYBxgeMggICRAAGAcYHtIBCTE0NTgwajBqN6gCALACAA&sourceid=chrome&ie=UTF-8

(2) Concerns regarding valuation and whether Nebius is a value-added cloud platform or basic AI bare iron compute provider based on the revenue from the MSFT and META deals (which average out to a combined amount of 4B/year of their 5 year contracts) seemed to look backwards and not at forward revenue projections which are: ARR by end of 2026: 8-9B from company guidance, and analyst estimates of 2027 revenue to be 10-12B and 2028 at 20B. Concerns about NBIS becoming only a bare iron metal provider to MSFT and META fade as the percentage of revenue from those contracts diminishes over time. I give a team that’s already “been-there and done that” with Yandex a reasonable vote of confidence they can pull this off.

(3) As for margins: with shortages of AI compute capacity a limiting factor, why would a company that is fully contracted with no excess capacity sign up for a low margin contract? Traditional justification for low margin contracts (entering new markets, launching new products, or need to keep the lights on in weak market/economic conditions) just don’t seem to apply here. Conceivably, low margins might have also deterred lenders of billions of dollars. I did a quick search on interest rates on NBIS debt and see 2 tranches worth 3B of convertible debt with rates of 1% on 2030 notes and a $138.75 conversion price and 2.75% on 2032 notes with an effective conversion price of $159.96 according to a September 10, 2025 article I found in Investing News Network. I did not bother to search beyond this.

(4) A final note on depreciation and returns from chips. Relative depreciation of a chip’s compute power is not new although it may be accelerated and have different characteristics with AI - but again NBIS’s leader (Arkady Volozh co-founded Yandex in 1997) has been in closely related businesses for almost 30 years and likely has about as much or more insight into the variables and likely outcomes as anyone. Maybe others can correct this but from what I’ve read, chips that are several generations beyond their prime still generate revenue and because they are fully depreciated the margins can be quite high – although this may be different with AI.

Nebius is my 5th largest tech holding.

32 Likes
34

I think the market knows exactly what Nebius is trying to do - grab some big, low-margin deals supplying bare metal to their AI Cloud competitors, using that scale and profits to fund their own AI Cloud offerings (NeoCloud).

The market is currently worried about the debt Nebius will have to take on to service Microsoft and Meta. Mr Market may not have faith that Nebius can be successful with their NeoCloud business, but there’s no confusion over what they’re trying to do.

No, Microsoft is leasing GPU bare metal, and so they can easily and quickly move workloads around. The ironic thing is that Nebius is enabling Microsoft to provide value-added AI Cloud services, and so switching away from Microsoft will be harder down the road. This hurts Nebius in the medium to long term. So, it’s a gamble that Nebius is taking now.

While Nebius’s current Debt To Equity Ratio is better than CoreWeave, Nebius still took on serious convertible debt and diluted shareholders after the MS deal, to the tune of $3.75B total,:
Cresco Capital.

$2.75 billion in convertible notes…and…$1 billion in new shares sold at $92.50 apiece

According to company filings, capital expenditures tied to the Microsoft deal will be financed through this raise and debt secured against the contract itself. Nebius is also said to be evaluating additional financing options to sustain even faster-than-planned growth.

And this was after a previous $1B unsecured convertible debt raise just 3 months earlier.

14 Likes
35

This is maybe where the big disconnect is. What are your assumptions on how much capital is needed to keep upgrading GPUs? I think it’s in the $10s of billions of dollars, certainly not easy for Nebius to manage. They have nowhere near the capital to do this.

Nebius current position: $800M net cash ($4.8B cash and $4b debt, down from $2.4B net cash at end of ‘24).

2026 CapEx:

I asked Gemini to estimate based on Nebius stated connected power (EO’26 800MW-1G from 220MW EO’25, 680MW increase at the midpoint). It came to $30B at $45M per MW. $30B also matches what someone else had posted in the board earlier. Is $45M in the ballpark?

" A cost of $45 million per megawatt (MW) is a highly reasonable, if not conservative, estimate for a state-of-the-art AI data center buildout when including the cost of high-end GPUs like NVIDIA’s Blackwell (GB200)."

Do you agree with these numbers or do you think we need a lower capex? Also, since 80% of capex is GPUs, what would you estimate the capital need in 2027?

Nebius has stated that they plan to use Microsoft’s credit to borrow which is a good thing “issuance of debt secured against the contract… at terms that reflect the credit quality of the counterparty [Microsoft].” This is also where it might be useful to reference Oracle, the overall funding market has become really skeptical in the last couple of months since the earnings release.

Even if they use Microsoft and get a near 5% rate (I don’t think this is possible), spending $30B to get $8B revenue per year over 5 years gives you a NPV of -$8B!

Just some thoughts for the longs to consider. Good luck to everyone.

9 Likes
36

This is what is under-discussed here and over at SeekingAlpha imo.

What happens to $NBIS’ business in 2-5 years if/when GPUs are suddenly in VERY short supply? imo here is what happens: their hyperscaler customers have the rights to usage of almost all of the GPUs in NBIS’ datacenters, on a percentage basis, at low/fixed prices, with the right to extend their contracts by several years e.g. well into the useful life of said GPUs.

…Leaving $NBIS with not enough GPUs for their larger aspirations, and unable to capitalize on the supply/demand driven increase in price for access to existing GPUs that would happen in a GPU shortage scernario.

I think that long term, $NBIS’ business model assumes availability of large quantities of GPUs 2-5 years from now, and I think that availability is of questionable likelihood.

11 Likes
37

“Many of the posts seem to give no credit to management or board ability to assess risks and I believe ignoring or dismissing their ability severely weakens our own decision making.”

Yes, I’ve been saying for much of this year, that one of the top reasons I’m so heavily invested in Nebius is because of Arkady and the 1000 plus engineers and team he brought with him from Yandex. We do need to trust them more. They’ve proved themselves so far, and I think they can continue to do it again.

Let’s also take fully into account their other businesses, AV Ride, Triple Ten, Toloka and their 28% stake in Clickhouse. With the AV ride partnership with Uber I see this part of their business has much further to grow.

**Do you agree with these numbers or do you think we need a lower capex?** Also, since 80% of capex is GPUs, what would you estimate the capital need in 2027?

In terms of capex, I am looking at smaller numbers than the ones pointed out above, but I guess it depends on the growth opportunities they continue you see.

Depreciation is a non cash expense. And, like many companies in hyper growth, they can literally chose when they become free cash flow positive. I guess it could be as early as 2027-28.

Basically, Nebius can become free-cash-flow positive once it gets through its main build-out phase and slows how fast it adds new data centers and GPUs. Right now it is spending heavily, but most of that spending is upfront and doesn’t last forever.

To reach its goals, I am guessing that the company might spend about $12–15 billion between 2025 and 2027, mainly on data centers and GPUs. I too asked AI about how much 1GW of data centre power costs. It said “About $12–15B capex. This comes from Nebius’ plan to reach about 1 GW of data center power, multiplied by typical AI data center costs of roughly $13–15 million per megawatt for buildings, power, cooling, and GPUs. After that, yearly spending should drop to around $1–1.5 billion just to maintain what it has.”

With long-term contracts already signed (about $8 billion a year in expected ARR by 2026) and high usage levels of around 75–80%, the existing equipment can keep bringing in cash even as new spending slows.

Because GPU depreciation is an accounting cost and not real cash going out the door, Nebius can turn free-cash-flow positive once the cash it generates each year is higher than what it needs to spend to maintain the business.

Jonathan

21 Likes
38

I find that really cheap because IREN spent 5.8B for 200MW of GPU or 1/5 of the compute your talking about. IREN total cost for 1GW based on all their numbers is close to 30B with 29B+ from GPUs alone. Which is double your high end.

I agree that depreciation is just an accounting cost and not real cash going out the door but Capex is not an accounting cost and is real cash going out the door. When Capex falls below depreciation for a heavy asset company, it signals that the company doesn’t have much growth. You either get growth (bad cash flow) or good cash flow (stagnation/decline). I don’t think you can get growth and good cash flow.

Nebius could potentially turn the corner in the future with higher cash flow and reduced spend with growth but that’s a story and not supported by their current financials. Here is a good quote from Saul on story stocks like Nebius.

Overall, I find this company very interesting, just the financials of this company are what’s holding me back. I would recommend caution or smaller position sizes because the company is priced for its success already and success would not move the needle that much.

Drew

14 Likes
39

I was listening to one of the interviews (I can’t remember which one at this point otherwise I would reference it) and the CRO Marc Boroditsky was asked how in the world they would compete with the hyper scalers in a cloud offering for AI. His response was that he sees it similar to cloudflare who was able to build a SAAS cloud and reasonably compete with the hyperscalers in their offerings. Take this for what it is worth, but there are companies in niche businesses that absolutely compete with the hyper scalers, and even collaborate with them. Take this info for what it’s worth. Long $NBIS

14 Likes
40

One of the recurring doubts Nebius bears have had is that the huge contract with MS will prevent them from devoting resources to their stated aim of becoming a full-stack provider (developing further their in-house tools, etc.). As an hypothesis it has some merit. After thinking about it for some time, I have come up with the following as ways to reject/not reject it.

Nebius already has a very robust set of tools, certainly more than IREN (though in fairness IREN has a different model). Just recently Nebius announced its Token Factory to complement its AI Cloud (with its various services), so this lends some credence to its claims.

As a proxy for the companies willingness to further fund R&D, build tools, etc., I looked at their recruitment page (both IREN and NBIS), and found it very instructive.

IREN had 257 employees at the start of 2025 (according to Seeking Alpha). It currently has 79 open positions, broken down as follows:

Commercial (M&A, etc): 5

Construction (electrical, mechanical, network, supervisor, etc): 18

Engineering (civil, HPC electrical/high voltage, civil, CMMS): 7

Finance: 4

HPC (HVAC maintenance, logistics coordinator): 2

IT (4 networking, then 7 other categories): 11

Marketing: 1

Operations (grab bag of jobs, about half are DC techs of some kind): 28

Procurement: 2

Risk and Compliance: 1

So, very focused on building and running DCs, with no need for other skills.

Nebius had 1371 employees at the start of 2025 (according to Seeking Alpha). It currently has 174 open positions, broken down as follows:

Communcations (PR, branding, events): 5

Customer Experience (Account executives 2, Ai/AI-ML/Cloud solutions architects 8): 10

Product Designer: 1

Finance: 11

Go-To-Market: 4

HR: 15

Hardware Infrastructure (DC tech, engineer, manager, Logistics, GPU architect, etc): 31

Legal: 6

Marketing: 7

Operations (this is not a DC ops category): 5

Product (solutions architect, product managers, etc): 14

Sales: 15

Security: 6

Technology (reliability, network, ML, software, HPC clusters etc etc): 43

To my eye there are enormous differences in the two companies as reflected in the roles they are hiring for. It is a direct consequence of their corporate strategies, and there looks to me to be no slowing down in NBIS’ work to build a full stack cloud.

10 Likes