Nebius Part 2: Cash, Optionalities, Challenges
In “Nebius Part 1: the Smooth (Developer) Operator”, I shared my point of view on why Nebius is the most compelling pure-play AI neocloud with its smooth developer experience, its history, and the pros and cons of its Yandex and Russian roots.
In this Part 2 installment, I explore the other, even more intriguing elements of this company – its cash position, its other divisions and optionalities, and the immense challenges it is facing in this crowded market.
Cash Hoard
The hardest, but also the most exciting, aspect of investing in a company like Nebius is putting a price on it.
It’s exciting because, until last Friday, few on Wall Street had heard of the company. No brand name brokerage has put a sell-side analyst to cover the company, though I suspect that will change soon. That’s music to the ears of an investor like me, because I get to evaluate the business I can understand well without any noise, as I wait for others to catch up and join the chorus with sell-side price targets.
It’s hard because there is no direct comparison to other existing companies. Nebius compares itself to CoreWeave, which as I’ve laid out in Part I, is not a great comparison. But it is the best we have to work with. CoreWeave is also a private company, so until its rumored IPO later this year becomes reality, doing an apples-to-apples valuation comparison is also difficult.
The most useful and reliable numbers to think about Nebius’s value right now are:
1. The company’s 2025 annual revenue guidance of $750 million to $1 billion in ARR, which is aggressive and impressive if it can pull it off;
2. Its roughly $3 billion in cash on the balance sheet with no debt – $2.3 billion from the Yandex divestment plus $700 million from Nvidia, Accel and Orbis.
The cash hoard is especially important because it separates Nebius from the rest of the neocloud crowd, almost all of which have taken on a lot of debt to build its AI infrastructure.
CoreWeave has more than $10 billion in debt. Lambda Labs has more than $500 million in debt. Crusoe has more than $200 million in debt. Even the vaunted Stargate project will likely need hundreds of billions in debt to fulfill its $500 billion investment promise.
There is nothing wrong with using debt, per se, to finance AI data center buildouts. These buildouts are more akin to real estate and energy projects than technology projects, and debt is very common in those industries. However, taking on debt gives you much less flexibility to appeal to developers, who are fickle sometimes but ultimately the most important audience you need to win over if you want to build a sustainably successful cloud business.
Nebius, with its strong cash position, will outright own all of its data center footprint, from Finland to Paris to various locations in the United States. Even if it uses debt to expand in the future, it is starting off from a strong equity and revenue base. This strong financial advantage gives Nebius an equally strong staying power, to keep honing its product as a relatively new player on the scene.
Other Optionalities and Add-Ons
The other intriguing part of Nebius is its collection of assets from the Yandex divestment: Toloka, Avride, Tripleten, and Clickhouse.
Most people I know who have invested in Nebius, or at least find Nebius as an interesting investment idea, think of this collection of assets as free call options. The company has telegraphed as much, stating that they are actively looking for strategic investment partners for each of these divisions to effectively “sell them off”. While this approach makes sense for some assets, it may not make sense for all of them.
Toloka: This division is the most intriguing for me. It is essentially a training, reinforcement learning, and evaluation data provider but for domains that require higher level expertise, e.g. math, coding, other scientific fields or knowledge work industries. Thus, it is a high-end boutique data provider targeting generative AI needs. Think of it as a smaller Scale.ai. As models like DeepSeek’s R1, OpenAI’s O3, and other reasoning models have shown, post-training with reinforcement learning is a direction that still has a lot of runway to improve foundation model capabilities. Therefore, data providers like Toloka hold strategic value in domains where human expertise is still valuable and not easily replaceable (for now) by synthetic data. Instead of selling this property off, folding Toloka’s capabilities into Nebius’s larger cloud portfolio could become a real differentiator versus a sea of lookalike neoclouds. It could even help Nebius punch above its weight versus the hyperscalers. Keeping and leveraging Toloka may make more strategic sense than selling it.
Avride: This division was Yandex’s autonomous vehicle and mobility division. It has been based in Austin, Texas for a long time, so it has the deepest American roots of all the subsidiaries. Impressively, it already has products deployed in the wild – autonomous delivery robots. Uber is already a partner. More recently, Grubhub has also signed up as a partner to offer robot delivery services on US college campuses, starting with Ohio State. Thus, Avride is compelling in its own right, and I can see it thrive both as an independent company with outside investors or as part of the Nebius umbrella.
The most important technology synergy is to build a symbiotic relationship between the Avride robots and the Nebius cloud, so the cloud platform can be strengthened and honed by autonomous vehicles deployed in the real world. The ability to serve this type of workload will further differentiate Nebius from the pack. Only Google Cloud with Waymo has this relationship. Tesla certainly has many more autonomous vehicles in the wild, but its computing infrastructure is not a separate business line service for 3rd parties. Nebius could become a legitimate cloud option for autonomous vehicle and robotics companies, with Avride as its first lighthouse “customer”.
TripleTen: This is a technology training bootcamp business for new engineers, with market penetration in the US and Latin America. It is probably the most “random” of all the assets, made especially so as AI coding products improve dramatically and the entire field of software engineering is in flux. I currently don’t have a strong opinion on whether TripleTen is better sold or kept. Selling it off to the team can focus on its cloud business is a very reasonable path. Leveraging it to develop a Nebius-specific training and certificate program is also a viable option and part and parcel to building a large cloud business (see AWS’s massive certification program).
Clickhouse: The last, but certainly not the least, of the assets is Clickhouse, an open source data warehouse that is one of the most performant options in the market and well-liked by developers. Its adoption is deep, from large enterprises, to hot Silicon Valley startups, to the AI upstart that took the world by storm, DeepSeek! Nebius's 28% stake in Clickhouse stems from its Yandex past, because Clickhouse was originally developed inside Yandex before being spun out as an independent startup. Its cap table is a who’s who of leading enterprise software VCs – Benchmark, Index, Redpoint, etc. Its last valuation was $2 billion in 2021, so likely a bit overvalued given how frothy that market was at the time. Nevertheless, Clickhouse’s product and open source traction is legitimate, so whether it pulls an IPO one day or becomes a first-class cloud product unique to Nebius, this 28% stake is a valuable asset with immeasurable upside.
Immense Challenges
Whether it's the cash, optionalities, or smooth developer experience, Nebius has a lot of going for it. As the Nvidia investment revelation gets absorbed by the market, it is a name that won’t fly under the radar for much longer. However, it still faces immense challenges as it tries to stand out in a crowded AI infrastructure market with many ambitious players.
One of its immediate challenges is making sense of all of its subsidiaries, so strategic value is maximized and distraction is limited. As I laid out in the previous section, it is not clear whether and which of the subsidiaries should be sold or kept. Some could enhance and differentiate the Nebius story; others are better off being managed and operated by someone else. Coming to terms with them sooner rather than later would be important in the immediate future for Nebius to take off.
Another perennial challenge for all new cloud players is its positioning versus the hyperscalers. Nebius is no exception. Startups are smart at playing off a new player for cheaper compute versus one of the hyperscalers, who have always been generous with free credits. Being able to appeal to developers from new startups and larger enterprises at the same time is a difficult game to win. The best case scenario is for a hot generative AI startup built natively on Nebius to take off and stay with Nebius for the long-haul, much in the same way that Netflix’s early adoption of AWS put the cloud juggernaut on the map. Will Nebius attract and keep the Netflix of generative AI with either its smooth developer experience or a special AI-native sauce crafted with, for example, Toloka, its in-house high-end data service, is the intriguing unanswered question.
A third challenge is the prioritization of its geographical expansion plan. It is known that the company has ambitions in both the US and Europe. As I noted, it also has the cash to execute. But the two geographies are quite different markets, and drifting increasingly apart geopolitically when it comes to AI. The US is obviously the more vibrant and lucrative market, and thus also the most crowded. The EU is lagging in all sorts of ways when it comes to AI development, but a growing rift with the US on all dimensions of AI’s future, epitomized in Vice President JD Vance’s speech at the Paris AI Action Summit, could mean Europe will be less willing to use American technology. Mistral’s CEO is already calling for Europe to build, own and operate its own AI infrastructure. This bodes well for Nebius as it has a head start in Europe with its main data center in Finland, expansion plan in Paris underway, and most of its core engineering force based in Amsterdam and Israel. If Europe becomes serious about building AI on its own terms, Nebius may become the go to choice, possibly owning the EU market at the American hyperscaler’s expense.
Whether that future pans out or not, of course, has less to do with the merits of technology and all to do with the fickleness of geopolitical wind. And if there is one tech founder that is all too familiar with that fickleness, it is Arkady Volozh. At the ripe young age of 61, Volozh is already a billionaire many times over from his multi-decade success of building Yandex since 1997. No one would blame him for phoning it in and enjoying the good life after completing the difficult divestiture of his entrepreneurial baby in Yandex. Instead, he is saddling up once again as the CEO of the reborn Nebius to prove that he can do it again, despite all the immense challenges in front of his team.
There are many talented, driven, and resilient tech founders in Silicon Valley, but none of their drive and resilience have been forged by a literal war, making Arkady Volozh and the Nebius crew hard to bet against.
If you missed Part 1 of my writeup on Nebius, I highly encourage you to give it a read!
Thanks for both Part 1 and 2. Excellent information. I am an investor in NBIS since January and looking forward to being a long-term part of its growth.
Any insight into their revenue makeup ?