Is Nebius the AI landlord of Europe?

Dear WIMM Supporter,

We look today at Nebius - The AI cloud company - with headquarters in Amsterdam.

Let’s do this!

TL;DR

Nebius is building an AI cloud. In plain language, it sells access to high-performance compute for companies training, fine-tuning, deploying, and running AI models. The core business is Nebius AI cloud, which now represents almost the entire group. Around it sit Avride, an autonomous vehicle and delivery robot business, TripleTen, an edtech reskilling platform, and equity stakes in Toloka and ClickHouse. Those side assets matter financially, but the investment case is now overwhelmingly about AI infrastructure.

The business model is simple to describe and hard to execute: secure power, acquire GPUs, build or lease data center capacity, wrap it in cloud software, then sell capacity to AI-native companies, enterprises, model builders, and hyperscalers. The gross profit comes from the spread between the cost of power, chips, real estate, depreciation, financing, and operations on one side, and the price customers pay for scarce compute on the other. The key variable is utilization. Empty GPUs are dead capital; heavily utilized GPUs are money-printing machines with depreciation schedules.

The first moat lever is scale. Nebius now has more than 3.5 GW of contracted capacity, raised guidance to more than 4 GW by year-end, and says owned capacity represents more than 75% of the contracted power base. That matters because the AI cloud market is becoming less like traditional software and more like energy plus semiconductors plus logistics. Power is the new distribution channel. The company that secures land, power, interconnects, permits, NVIDIA supply, and customer commitments has a real advantage before the software demo even begins.

The second moat lever is switching cost. Once a company trains, deploys, and optimizes workloads on a cloud platform, migration becomes painful. Nebius is leaning into this with Aether 3.5, serverless AI, data transfer services, governance tools, Token Factory, and inference optimization. The strategic logic is clear: raw GPU rental is a commodity; production AI workflow is sticky. If Nebius becomes the place where customers train, deploy, optimize, retrieve, infer, monitor, and scale, then the moat moves from capacity to workflow dependence.

The third moat lever is IP and technical depth. The acquisitions of Tavily, Eigen AI, and Clarifai show the company is trying to climb the stack from infrastructure into inference, agentic search, orchestration, token efficiency, and optimization. This is the right move. The real prize is not selling hours of compute; it is lowering cost per useful token for customers. If Nebius can consistently deliver better throughput per GPU, better inference economics, and faster deployment, it can compete on effective cost, not headline price.

The contrarian insight is that the real risk for Nebius is that demand stays enormous but the company becomes a low-margin construction company wearing a cloud multiple. Q1 argues against that risk, at least for now. Revenue exploded, capacity is increasingly pre-sold or strategically financed, and AI cloud adjusted EBITDA margin reached 45%.

The financial message was blunt: Nebius has crossed from concept into execution. Group revenue hit $399 million, up 684% year over year, and core Nebius AI cloud revenue reached $389.7 million, up 841% year over year. ARR reached $1.92 billion, up 674% year over year and 54% from Q4 2025. This is no longer a “future AI infrastructure” pitch; this is current demand meeting scarce supply.

Margins were the surprise. Group adjusted EBITDA reached $129.5 million, while Nebius AI cloud reached a 45% adjusted EBITDA margin, nearly doubling quarter over quarter according to management. That is the key number in the whole report. In AI infrastructure, revenue growth can be bought with capex. Margin expansion during hypergrowth suggests pricing power, utilization, and operating leverage. That is moat evidence, not accounting decoration.

Cash flow needs careful interpretation. Operating cash flow was $2.3 billion positive, but that was heavily supported by deferred revenue, meaning customer prepayments and long-term contracts are financing the buildout. Capex was massive: purchases of property and equipment and intangible assets were $2.47 billion in the quarter. This is the trade: Nebius is becoming more valuable if it can turn capital into contracted, utilized capacity faster than competitors; it becomes dangerous if financing, delivery, or utilization slips.

Management emphasized capacity. Contracted power already exceeds 3.5 GW, and the company raised year-end guidance to more than 4 GW. It announced a second owned gigawatt-scale U.S. site in Pennsylvania with up to 1.2 GW, alongside the Missouri AI factory, plus a large Finland site. The operating message is that Nebius is no longer just renting availability from the market. It is trying to own the bottleneck.

The Meta agreement is the most important strategic signal. A $27 billion five-year deal gives Nebius revenue visibility, financing leverage, and validation from one of the largest AI infrastructure buyers in the world. Structurally, this matters because customers are not merely buying compute; they are helping finance capacity expansion. That is how a challenger cloud survives against hyperscalers: turn demand into funding before the balance sheet breaks.

The NVIDIA relationship is the other pillar. NVIDIA invested $2 billion, Nebius achieved Exemplar Cloud status on GB300 NVL72, and the partnership now includes deeper work on software, AI factory design, inference, and agentic infrastructure. The risk is dependency. The benefit is credibility. In an industry where customers worry about performance, availability, and execution, NVIDIA validation lowers perceived risk.

Segment by segment, Nebius AI cloud is accelerating, Avride is progressing but remains optionality, TripleTen is stable but no longer central, and ClickHouse is a hidden asset with valuation upside. ClickHouse’s reported valuation contributed a large non-cash revaluation gain, which inflated net income. Investors should ignore that for operating quality and remember it for balance sheet optionality.

Concluding, Nebius is trying to become a scarce-capacity platform with software differentiation, not a GPU reseller. The moat is strongest where power, customer prepayments, NVIDIA alignment, and inference software reinforce each other.

Nebius has one job over the next year: bring capacity online without breaking execution quality. The company expects 800 MW to 1 GW of connected power by year-end, while contracted power should exceed 4 GW. This gap between contracted and connected power is the central operating tension. Contracted power is future potential; connected power is monetizable reality.

The product roadmap is moving toward production AI. Aether 3.5 adds serverless AI, data transfer, governance, security, billing, and audit capabilities. Token Factory is positioned as the inference layer for open-source and custom models. Tavily adds agentic search. Eigen AI adds inference and model optimization. Clarifai adds system-level orchestration and talent. The strategy is obvious and correct: sell infrastructure first, then increase software attachment until customers stop viewing Nebius as interchangeable capacity.

Near-term catalysts:

The operator lesson is brutal: in AI infrastructure, sales is now finance, finance is now power procurement, and product is now utilization engineering.

Over the next 1-3 years, Nebius wants to become a full-stack AI cloud for production workloads, not just a place to rent GPUs. The desired end-state is a global AI factory network with owned capacity, pre-committed demand, NVIDIA-aligned hardware roadmaps, and a software stack that makes workloads cheaper and easier to run. If successful, Nebius becomes the alternative AI cloud for companies that do not want to be fully dependent on AWS, Azure, Google Cloud, or Oracle.

The major CAPEX arc is enormous. Pennsylvania, Missouri, Finland, New Jersey, Alabama, Spain, and other locations create a global footprint that looks less like a startup and more like an infrastructure utility. This is capital intensive, but it also creates barriers. Competitors cannot copy gigawatts of power, data center construction, supply agreements, and customer contracts overnight.

Verdict: Buy.

WIMM for companies: If you would like to offer the newsletter as a benefit to other employees in your company, please contact me directly. ([email protected])

Work with me: If you are interested in working with me on your personal portfolio, again, contact me. I’m a certified financial consultant.

Finally, “Where is my moat?” is designed for individual readers, though the occasional forward is absolutely fine.

Thanks for your support & have a wonderful day!