Thesis
Nebius Group is the cleanest public-market way to own a vertically integrated AI cloud outside of the U.S. hyperscalers. The company spun out of the old Yandex business, re-listed on Nasdaq in late 2024, and raised around $700 million in a strategic round led by NVIDIA, Accel, and Orbis. Since then it has moved fast: large GPU deployments in Finland, new sites in Kansas City and Paris, and a growing set of customers that want Hopper and Blackwell capacity without waiting in a hyperscaler queue. The reason the story matters is that Nebius is not just reselling GPU time. It designs its own servers and data-center hardware, operates the clusters itself, writes the orchestration and ML tooling on top, and has adjacent holdings in Toloka (data labeling), Avride (autonomous driving), ClickHouse (analytics database), and TripleTen (ed-tech).
The current numbers are still small relative to the hyperscalers but the trajectory is what the market is paying for. Management has guided annualized run-rate revenue to reach the high hundreds of millions exiting 2025, with ambitions to scale meaningfully higher as newly committed capacity comes online. The real question is whether Nebius can turn GPU-as-a-service scarcity into a durable, higher-margin AI cloud platform with sticky workloads, or whether it ends up as a commodity GPU rental shop once supply catches up. If it proves out the platform story, the sum-of-parts plus the core AI cloud can justify a much larger valuation than a generic cloud vendor deserves.
Nebius clearly has the team, the hardware, and the capital to become a credible non-hyperscaler AI cloud. The real question is whether the platform gets sticky enough — through tooling, contracts, and integration — to earn a premium multiple rather than just ride whatever GPU-rental cycle is running.
Valuation and financials
The 4Ps
CEO Arkady Volozh co-founded Yandex and is now running Nebius as a clean, Amsterdam-based, Nasdaq-listed AI infrastructure company. That matters because Nebius is not a newly assembled venture — it is a team that has operated search-scale infrastructure for decades and is now pointing that experience at a very different workload. The main execution risk is speed of build-out, not engineering competence.
Nebius designs its own servers and cooling, deploys NVIDIA Hopper and Blackwell at scale, and layers its own training and inference platform, ML tooling, and managed Kubernetes on top. That vertical integration is the point. Customers who need reliable, predictable, non-hyperscaler GPU capacity get an opinionated platform, not a thin wrapper around bare metal.
The bull case has two layers. First, the core AI cloud becomes a real multi-billion-dollar revenue business as committed capacity comes online in Finland, Kansas City, Paris, and new sites. Second, Toloka, Avride, ClickHouse, and TripleTen remain valuable and occasionally crystallize that value — ClickHouse in particular has raised at multi-billion-dollar valuations in the private market. The combined platform can look much bigger than the reported AI cloud revenue alone suggests.
Committed customer contracts and announced data-center capacity give more visibility than a typical early-stage cloud, and the NVIDIA strategic investment is a strong signal on allocation priority. But this is still a business scaling revenue from a small base, funding large data-center capex, and operating in a market where GPU supply, power availability, and pricing can all move quickly.
Portfolio manager lens
Starting point: Nebius is one of the few public-market ways to own a non-hyperscaler vertically integrated AI cloud alongside a bucket of interesting private holdings.
What is in the stock: a credible team, an NVIDIA strategic backing, announced capacity expansions in Finland, Kansas City, and Paris, a sharp run-rate revenue trajectory, and optionality from ClickHouse, Toloka, Avride, and TripleTen.
What can still surprise upside: contracted capacity filling faster than expected, gross margins improving as new sites stabilize, longer-duration enterprise contracts locking in utilization, and a real monetization event around one of the holdings.
What changes the view: GPU pricing normalizing before the platform becomes sticky, capex running ahead of contracted demand, dilutive financing rounds, or the holdings failing to crystallize value on any reasonable timeline.
Trade framing
Nebius is best thought of as a high-volatility platform call option with a real underlying business. It is already generating revenue, it already has the hardware, the team, and the NVIDIA relationship, and it already has the holdings optionality. The question is whether the next 12 to 24 months turn that into a durable platform or into a cyclical GPU rental story.
The next checkpoints are practical: annualized run-rate revenue updates, data-center build progress in Kansas City and Paris, new customer commitments, gross margin trajectory as capacity stabilizes, and any monetization event around the holdings. If those keep landing, the stock can keep earning a better identity. If not, it stays a sharper-than-average GPU proxy.
What matters now
What matters now is whether Nebius can convert announced capacity into revenue on schedule and sign longer-duration customer contracts that lock in utilization. The checkpoints are annualized run-rate revenue trajectory, gross margin evolution as new sites stabilize, data-center build milestones in Kansas City and Paris, and any monetization events around ClickHouse, Toloka, Avride, or TripleTen.
Key questions
Nebius operates a vertically integrated AI cloud. It designs its own servers and data-center hardware, deploys NVIDIA Hopper and Blackwell GPUs at scale, and sells training, fine-tuning, and inference capacity to AI labs, enterprises, and startups. On top of the raw GPU layer, it offers managed Kubernetes, model training platforms, inference endpoints, and supporting ML tooling.
Beyond the core cloud, the group still holds stakes in Toloka (data labeling and human feedback for AI), Avride (autonomous driving and delivery robotics), ClickHouse (analytics database, one of the fastest-growing OSS data systems), and TripleTen (ed-tech reskilling). That sum-of-parts makes the story more than a pure GPU rental business.
Thesis last reviewed April 5, 2026. Live data updates automatically.