Dutch Nebius is the Phoenix of the AI cloud

Stocks
, ,
287

Attached is my own updated Nebius Valuation Model v7, free for use. Below is my relatively realistic base case scenario, which I expect Nebius to reach with multiples suitable for this market. There might still be errors in the code; I didn’t have time for a more extensive check. You are free to copy and utilize the model however you like, and of course, feel free to post your own modeling estimates.

image
image504×1215 87.1 KB

Nebius Valuation Model v7 (Eigen AI acquisition added in anticipation (likely to go through), share dilutions/convertible bonds with new share counts accounted for cumulatively, Q1 and 2026 forecasts at the top, GWs updated, two models: EV/ARR and “capacity model” i.e., GW by 2030, etc.).
LINK:
https://claude.ai/public/artifacts/a209db26-593c-44c4-878a-eb917f0f8621

16 Likes
288

I have significantly improved Nebius Valuation Model v8 from version v7. The basic assumption in the model is, of course, that Nebius will grow. There is still much room for improvement, but the new version includes, in my opinion, the following very useful features:

Improvements in v8:

  • Hybrid valuation (EV/ARR and EBIT): Ability to toggle on an annual basis (2026–2030) between the EV/ARR multiple and the EBIT × P/E multiple. For example, EV/ARR is very sensible during the hypergrowth and investment phase in 2026–2027, and EBIT and P/E during the presumably more mature phase of 2028–2030.

  • Annual revenue per MW ($/MW): Each year now has its own slider (range $5–30M, $30M might be a bit of a stretch, but recent growth has shown that compute demand and price can rise drastically). The default value is set at $10M (2026, Hopper/Blackwell) → $15-20M (2027 including Vera Rubin and beyond), because next-generation GPUs, such as Vera Rubin, generate better yields. Meta’s Vera Rubin contract may have been made at a fixed price, but Nebius has the option to choose whom it sells the excess capacity to.

  • An important update is Net Dollar Revenue Retention (NDR), taken from Daniel Koss’s very bullish X post from earlier this week. It’s a very useful and more accurate tool for evaluating and modeling valuation, as long as we get interim data along the way:

    • Split into two segments: Hyperscalers/contract-based (95–160% lower fixed longer contract prices for Meta, MSFT) and Enterprise/consumption-based (100–250%, more valuable new/smaller customers and e.g., Token Factory).
    • Adjustable share of hyperscaler contract revenue (20–90%), i.e., how much of the revenue comes from, for example, Meta’s or Microsoft’s 5-year contracts. The higher the weight on small customers, the better the prices. This is Nebius’s strategy and represents higher quality revenue. Megacaps act as growth financiers.
    • The model calculates the combined NDR and shows the breakdown in the table (base revenue + required new sales). Useful along the way if, for example, Nebius discloses its customer base in reports or even in passing during earnings calls.
    • If NDR alone exceeds the target (i.e., negative new sales), a red warning appears. Daniel explains NDR in his long bullish text with simple examples. I’ll include the link below.

  • EBIT margins: Sliders for all years 2026-2030 (-10% – 50%, a wide range, just in case). Estimated realistic rise from negative/low towards 15–30%+ from 2028 onwards, up until 2030. Ability to alternate between EV/ARR and EBIT P/E on an annual basis.

  • Scenarios and presets Bear, Base, Bull, Ultra Bull: Several ready-made scenarios that set exemplary growth outlooks; the intention is to use the model as new data becomes available. Otherwise, the default settings are indicative; I entered values mainly for testing purposes ($/MW, NDR, and margin curves).

  • Other features:
    ◦ Current share price default at $154 (the close on May 1st was approx. $154.50).
    ◦ The range for multiples has been expanded (EV/ARR 3–25x, P/E 15–60x, I should have set P/E to 10-60x).
    ◦ All annual panels clearly show revenue, capacity, and EBIT data. Added an NDR sanity check at the end to prevent errors.

I am satisfied with this model; the previous ones were light versions, and I’ll leave the rest of the improvements for the next version once we get more data. Please report any potential errors or inconsistencies.

Free to use, copy, and modify.
EDIT: updated slightly, nothing major. Added a link to the forum thread and my usernames; v8 remains the same and the old link still works.
https://claude.ai/public/artifacts/f70abb2b-740d-4407-8125-238c859056b4

EDIT2: Updated version v8: NDR affects price targets.
https://claude.ai/public/artifacts/505a3a1a-e1ad-43db-a1f6-77826a2f63d3

Daniel Koss’s bullish evaluation - Nebius to a trillion by 2030, part 4
https://x.com/daniel_koss/status/2048730040871915936?s=20

20 Likes
289

Nebius’s F-20 report confirmed that Microsoft’s approximately $17 billion prepayment was 40%, amounting to $6.96 billion. According to Morgan Stanley, this is significantly higher than what analysts had expected (e.g., CoreWeave’s [CRWV] typical 15-25% prepayments).

During the last Q4/25 earnings call, Nebius management confirmed that the new contracts at that time were made with 60% prepayments, and in some cases, the payment was already 100%. The price of compute and tokens has only risen sharply since that call, so the assumption is that NBIS has an excellent opportunity to negotiate high prepayments in all contracts.

Prepayments significantly ease the funding of the massive $16-20 billion 2026 capex, and in my opinion, NBIS has shown an exceptional ability to avoid diluting current shareholders’ ownership. CEO Arkady Volozh, who owns a significant number of shares, has interests fully aligned with other owners, which is one reason why the stock has been easy to hold and add to despite the winter’s declines. Tranches 1 and 2 of the MSFT contract have been delivered on time, which increases confidence.

The report also reveals that shorter contracts have been made with other customers, which could have a significant positive +$550 million ARR impact. Nebius should thus be able to benefit quickly from the risen compute prices. Additionally, the report validates that Nebius is seen as a reliable partner in the eyes of the largest Megacaps. Positive confirmation for last year’s prepayment speculations was essentially obtained from the F-20 form.

On X, there has also been speculation that Nebius might soon sign a deal with Anthropic, which may have contributed to the stock’s rise on Friday alongside the Eigen AI deal. According to a check by JPM, OpenAI is now a customer of Clickhouse, which is partially owned by Nebius; the same applies to Anthropic.

https://x.com/zen_tropy/status/2050945242384277876?s=20

image
image387×747 118 KB

20 Likes
290

I would also like to ask @Thiebault, did you somehow manage to factor the competitive advantage brought by Eigen into your valuation model? With the market’s most efficient optimization (Eigen AI), there can be significant beneficial multiplier effects to curb the upward price trajectory of total GPU solutions, which is currently manifesting, for example, as price overheating in semiconductor memories and fiber optic networks. And indeed, optimization saves energy or, conversely, enables larger computational models with the same infrastructure costs. This is exactly why I like the Eigen deal, as it somehow tempers the already overheated component market for AI infrastructure.

5 Likes
291

On Friday, I thought about adding an “Eigen AI efficiency factor” to improve the $/MW already found in the model. However, I thought the matter needed some more mulling over, especially as I ran into an issue with the NDR update. I consider it almost certain that the benefit of Eigen AI will be real and significant. I am also very satisfied with the deal, the more I’ve looked into it.

Other options I considered regarding AI to account for Eigen in the model could be an “NDR boost”, meaning Nebius would be seen as a better provider, where customers get more for their dollars and therefore expand their usage.

A third could be, for example, the win rate per new customer, and it could increase the model’s “New Logo Revenue” if/when Nebius becomes more attractive, especially among startups/customers who do a lot of benchmarking.

Finally, I thought: the more efficient Nebius inference offered by Eigen, which leads to lower service costs, leading to higher EBIT margins over time. So, a kind of “Eigen margin” slider on top of the base EBIT margins. For the next few years, however, we are still going by ARR.

All of these could very well be added after the earnings report, once we get new data and ideas. And indeed, my model can be modified, used, and anything else, but I hope, of course, that you’ll post it here (or at least send it to me) when you make it better.

I did indeed update the model again today as the NDR wasn’t linking to the target prices. Now it hopefully works. All previous versions are also still available.

10 Likes
292

Attached is the updated Nebius Valuation Model v9, which includes the Eigen AI enhancers mentioned by @karhulalainen (4 units). The multiplier effects over the years can be absolutely massive, and the “compound interest” phenomenon can grow surprisingly large in hyper-growth companies. I should also mention that humans often find it very difficult to perceive exponential growth several years into the future, and the purpose of this model is to visualize that potential to the best of its ability. Personally, I’ve fallen into this trap with the likes of Palantir and Tesla and sold too early.

There may be (and probably are) errors in the model and things I know should be implemented slightly differently. In broad strokes, however, it is a good tool for testing and observing different scenarios.

Nebius Valuation Model v9

https://claude.ai/public/artifacts/2eb805eb-f899-41c8-8bff-1bebee8540b4

18 Likes
293

Thanks @Thiebault for the v9 calculations. Nebius is a truly difficult company to grasp in terms of its potential reach. But since the optimization is done via software, the margin is better than optimization done with “iron” (hardware). The Nebius Valuation Model is challenging to understand in itself; but no matter, any model is better than no model at all. I think Nebius could secure the best deals in the near future. It might even succeed in building a significantly better software ecosystem for AI/AI edge computing than its competitors. In that case, Nebius could become a dominant platform, and we would reach the Nebius Valuation Model bull scenarios.

Iren bought Mirantis at the price of Eigen AI. In my opinion, this is more “bulk” compared to Eigen. Other Neoclouds are now in a panic because of Nebius’s Eigen deal. That’s why this deal happened now too.

image
image710×612 96.5 KB

7 Likes
294

This is an absolutely nonsensical claim. This has been in the works for a long and arduous time. Iren’s vision is very long-term, as can be seen from their powered land portfolio, which they started building up last decade.

In no universe would the brothers start making acquisitions in a week, and even less so because of another company’s acquisitions.

10 Likes
295

Yeah, nothing happens in the M&A market in a week or even a month. By default, the discussions with Mirantis were likely initiated (as usual) in the spring/summer of 2024. For a listed company to reach the point of signing a binding purchase agreement—even if the price is right—it takes at least six months for DD (Due Diligence). You’re only kidding yourself if you think M&A processes are fast and dynamic; it isn’t quick, even when there is plenty of desire and will.

5 Likes
296

An absolutely excellent interview regarding data centers. Daniel Koss interviews Charles-Antoine Beyney (DataOne), who is a developer and partner for Nebius’s data centers.

The interview provides a lot of detailed information about the current situation in the data center market. Additionally, Beyney mentions that the Vineland project is well underway, although the coldest winter in recent history caused some delays. Furthermore, the mention of the French national grid being 100x better than the U.S. grid is a positive point, and in my opinion, Nebius’s data center locations around the world are a good strategy that mitigates risk to some extent. The interview touches on all the essential matters that are significant during the design and construction phase of data centers. Cooling solutions, the best forms of energy, water usage, and related factors are also explored in sufficient depth.

image
image440×709 22.5 KB

https://x.com/daniel_koss/status/2054183497279881636?s=20

P.S. The company that Charles-Antoine Beyney (Owner/CEO of DataOne) mentions at the 10:23 mark is Syklea.

1 Like