The explosion of artificial intelligence has sparked a race to build the cloud infrastructure that powers AI models. Two emerging competitors, Nebius Group (NBIS, Financial) and CoreWeave (CRWV, Financial), have stepped onto this stage as alternatives to the traditional hyperscalers, Amazon AWS (AMZN, Financial), Microsfot (MSFT, Financial) Azure, Google (GOOG, Financial) (GOOGL, Financial). Both focus on GPU-centric cloud services built for AI, but they differ sharply in strategy, geography, and financial structure.

In this article, I will break down Nebius's business model, technology stack, and expansion roadmap and evaluates how it stacks up against CoreWeave and the broader GPUaaS market.

Business Model

Nebius Group emerged in 2024 following the breakup of Russia's Yandex. The Amsterdam-based firm was rebranded from Yandex N.V. (Yandex's Dutch parent) after it divested its Russian operations for $5.4 billion. At its core, Nebius delivers cloud-based graphics processing unit (GPU) compute, primarily using Nvidia (NVDA, Financial) hardware, for AI model training and inference. In essence, Nebius offers Nvidia GPU clusters on demand, along with a suite of AI development tools.

But Nebius isn't just about raw compute. The company offers an integrated AI platform, supporting Nvidia's AI Enterprise suite and machine learning operations (MLOps) tooling. In June 2025, it partnered with Saturn Cloud to launch an “AI-first MLOps cloud,” reinforcing its ambition to build an end-to-end ecosystem.

While 75% of revenue stems from its core AI cloud, Nebius has three other segments: Toloka, a data-labeling platform backed by Bezos Expeditions, TripleTen, an edtech coding bootcamp, and Avride, an autonomous driving software effort. These units may not yet move the financial needle, but they provide vertical integration and long-term optionality.

Source: Nebius

The company targets AI-native startups, labs, and enterprises seeking scalable infrastructure, similar to rivals like Lambda and CoreWeave. Nebius has highlighted partnerships and pilot customers rather than naming large established clients so far. Saturn Cloud is a good example: it bundles Nebius's compute with user-friendly dev tools. Another example is Nebius's Toloka platform, which secured an investment from Bezos Expeditions (the personal venture fund of Jeff Bezos) in May 2025 a signal that Nebius's ecosystem (in this case, an AI data labeling service) has caught the attention of major tech investors.

That approach also gives Nebius a very different customer mix than CoreWeave. While CoreWeave generated 77% of its 2024 revenue from just two clients (Microsoft alone accounted for 62%), Nebius has no such concentration. Its customer base is broader.

On the technology stack, Nebius is betting heavily on Nvidia hardware and associated software. Nebius offers Nvidia's top-tier GPUs such as the A100 and H100 today, and it has announced early access to next-generation chips like the Nvidia H200 and “Blackwell” GPUs in its pipeline. In fact, Nebius claimed to be the first in Europe to deliver general availability of Nvidia's Blackwell architecture, signaling a close partnership with Nvidia. The cloud platform itself includes compute instances with Nvidia GPUs, high-performance networking, and distributed storage.

Why is this Business Model so hot?

The explosion in generative models like ChatGPT, Stable Diffusion, and enterprise LLMs has triggered a global GPU arms race. Training or fine-tuning these models requires thousands of high-end GPUs, like the H100 or Blackwell, each priced upwards of $30,000. But GPUs alone aren't enough. Running them efficiently requires liquid cooling, low-latency networking, high-throughput storage, and technical talent that most organisations can't afford.

Enter GPU-as-a-Service (GPUaaS). This model allows companies to lease high-performance GPUs on demand, avoiding the upfront capital and infrastructure burden. GPUaaS enables startups, researchers, and enterprises to scale workloads elastically, whether training models from scratch, fine-tuning open-source architectures, or running inference at scale. It's cost-efficient, scalable, and strategic.

But why use Nebius or CoreWeave instead of simply calling OpenAI or Anthropic? The answer lies in control, cost, and flexibility. Using OpenAI's APIs locks users into a black box (limited customisation, no control over latency or inference logic, and costs that scale poorly at volume). GPUaaS providers like Nebius unlock a different path. One where companies can train or fine-tune their own models, maintain full IP ownership, and optimise performance for their specific workloads. It's the difference between renting pre-packaged software versus owning your tech stack. For AI-native startups, research institutions, and even mid-size enterprises with proprietary data, that control is critical.

Expansion Plans in Europe and Beyond

Since mid-2024, Nebius has been executing on ambitious expansion plans, especially across Europe. The company announced it will invest over $1 billion by mid-2025 to build out AI infrastructure in Europe.

Nebius already operates a major data center in Finland (a holdover from Yandex Cloud). In 2024, Nebius began expanding the Finland facility to increase capacity. It also signed letters of intent for two additional data centers in Europe. One concrete project is a new GPU cluster in Paris, France, which Nebius says will offer Nvidia GPUs to European customers. This Paris region gives Nebius a presence in Western Europe's heart, likely aiming to attract EU-based AI startups and enterprises that prefer their data to stay within EU borders.

In June 2025, after Nvidia CEO Jensen Huang said the United Kingdom (UK) is in a Goldilocks circumstance and he would invest there. After that, Nebius officially launched its services in the UK, marking its entry into the British market. As part of the UK launch, Nebius highlighted that it's bringing the latest Nvidia Blackwell Ultra GPU systems to British soil. By offering the newest GPUs (even before many competitors), Nebius is courting the UK's AI developers who need top-tier compute. The UK expansion also likely involves setting up a local office and partnerships to integrate into Britain's AI ecosystem. With Nvidia announcing several commitments and increasing awareness about the UK AI infrastructure, and helping Nebius to deploy new facilities in the country will be a great opportunity for Nebius.

Nebius isn't stopping at Europe. The company has quietly begun expanding into North America as well. In early 2025, it opened its first U.S. GPU cluster in Kansas City to meet surging AI compute demand. Nebius will compete in the home market of CoreWeave, AWS, and other American providers, leveraging perhaps a cost advantage or available capacity to carve out a niche.

Its Finnish data centre is being scaled to support up to 60,000 GPUs, while its new U.S. site in Kansas City could host up to 35,000 more. These clusters are optimised for AI compute, using NVIDIA H100s, H200s, and the upcoming Blackwell line.

Of course, expansion at this scale requires immense capital. Nebius's spending plans include $600 million to $1.5 billion of capital expenditures in the near term to increase capacity in Finland, build out France, and also move into North America. To fund this, Nebius has tapped its cash reserves and raised new funds. In late 2024, Nebius secured $700 million in a private placement from investors including Nvidia and Accel, issuing new shares at $21 each. And in June 2025, it raised an additional $1 billion via convertible notes to bolster its liquidity.

Nebius vs. CoreWeave

Nebius's most direct rival is CoreWeave, the U.S.-based cloud provider that has carved out a strong position as a specialist in GPU computing. Both companies are GPU-native infrastructure providers built from the ground up to serve AI developers, and both emerged outside the hyperscaler ecosystem. But beyond that shared DNA, their strategies diverge sharply.

CoreWeave started in 2017 and pivoted from cryptocurrency mining to GPU cloud services. Since then, it has ridden the generative AI wave to astonishing heights. Revenue surged from just $25 million in 2022 to $229 million in 2023 and $1.92 billion in 2024. This 8× annual growth in 2024 reflects CoreWeave's key role in supplying compute to AI firms. For instance, it reportedly secured a major contract to provide infrastructure to OpenAI in 2023, stepping in to fill capacity gaps left by the hyperscalers. At one point, spot-market GPU prices reached $8/hour for top chips. CoreWeave capitalised on that spike by procuring hardware at any cost, including used GPUs from crypto miners and massive forward orders of H100s.

Nebius, by contrast, entered the market in 2024 but with unique advantages. Rather than building from scratch, it inherited Yandex's cloud platform, engineering talent, and a sizable capital base following the $5.4 billion divestment of its Russian business. While CoreWeave raised over $2.3 billion in equity and secured more than $10 billion in debt facilities to finance growth, Nebius has a stronger balance sheet from the outset. With that being said, Nebius can afford to undercut on pricing and invest heavily without immediate pressure to generate profits, whereas CoreWeave, while also growth-focused, carries significant debt and will eventually need higher margins to service it.

Both companies rent out GPU-heavy compute instances via the cloud, but Nebius's model is slightly more holistic. In addition to bare compute, it offers a full stack of cloud services (storage, networking, managed databases) and is building partnerships to make its platform more attractive to AI teams. CoreWeave, by contrast, emphasises flexibility and raw performance. It offers virtual machines or containers with GPUs, targeting advanced AI labs that require massive training clusters and already have in-house infrastructure expertise.

Source: Nebius

In GPUaaS, where infrastructure is increasingly commoditised, price and availability matter. Nebius recently slashed on-demand rates. Its H100 dropped from $4.85/hour to $2.95/hour pay-as-you-go, with further discounts for long-term commitments. CoreWeave's on-demand H100 pricing remains around $4.25/hour, although it offers up to 60% discounts for committed usage. This 45% discount is meaningful, and such a price difference can be very compelling for cost-sensitive customers like AI startups burning investor cash to train models.

Source: CoreWeave

Interestingly, both companies count NVIDIA as a strategic backer. NVIDIA invested $350 million for a 7% stake in CoreWeave and also participated in Nebius's funding round. This dual alignment signals NVIDIA's intention to support multiple GPU-native cloud providers and maintain high downstream demand for its hardware. That backing also gives both firms early access to next-gen chips—an edge in securing performance-sensitive workloads.

Neither Nebius nor CoreWeave builds its own chips (unlike AWS with Trainium or Google with TPUs), but both are innovating in deployment. Nebius, for instance, secured early access to Blackwell GPUs in Europe, likely aided by its speed of execution and strategic positioning. Similarly, CoreWeave was among the first to offer H100s at scale in 2023. Both aim to stay on the bleeding edge of NVIDIA's roadmap—and attract customers accordingly.

Unit Economics and Technology Stack

A critical piece of the analysis for any cloud provider is unit economics. How profitable (or not) each unit of service is, and what the cost drivers are. In the context of Nebius and its peers, the “unit” is often one GPU hour of computation.

On the revenue side of a GPU-hour, as noted, Nebius cut its on-demand H100 GPU price to $2.95/hour or even lower with long-term commitments. For comparison, AWS charges around $12.29/hour for an on-demand H100 80GB (depending on region).

Now, on the cost side, providing one hour of H100 time involves several components: depreciation or lease costs for the GPU hardware, electricity, cooling, infrastructure overhead, and some share of staffing and maintenance. Assuming a $30,000 price tag for an Nvidia H100 80GB and a 3-year capital recovery period, that's $833 per month. At 730 hours/month (24×7 uptime), that equates to $1.14 per hour just to break even on the hardware.

In practice, utilisation isn't 100% (there will be some idle or reserved time), and there are significant power/cooling costs. An H100 can draw 300 to 400 watts under load. At $0.05–$0.10 per kWh, electricity costs range from $0.15 to $0.30/hour. Add cooling and networking overhead ($0.10/hour), and total operational costs sit between $0.20 and $0.40/hour. Server CPUs, memory, and networking per GPU add further costs, though relatively minor on a per-hour basis.

Bringing this together: if Nebius averages $2.20/hour per H100 (blended between on-demand and committed use) and runs at 80% utilisation, it generates $1.76/hour. With $0.32 in variable opex and $1.14 in depreciation, that leaves $0.30 of gross margin per available hour. Over a year (8,760 hours), this adds up to $15,400 in revenue and $12,600 in gross cash flow, enough to recoup the GPU investment in just under 2.5 years, excluding staff and overhead.

At CoreWeave's $4.25/hour, the math looks different. At 80% utilisation, it earns $3.40/hour, about $22,500 annually per GPU. That allows for a 1.3-year payback, assuming similar cost assumptions. But I would expect the higher price tag may reduce utilisation.

From a capex efficiency standpoint, Nebius may have structural advantages. First, its balance sheet: it holds $1.4 billion in net cash, while CoreWeave carries $6.2 billion in net debt. CoreWeave's debt-servicing costs add to its effective GPU-hour cost. Second, Nebius can build where costs are lowest. For example, its Finnish data centre benefits from ambient air cooling and access to low-cost nuclear and hydro power. Finland's industrial electricity rates (€0.05 to 0.07/kWh) are highly competitive, even versus many U.S. regions. Further Nordic or Eastern European expansion could enhance this edge. Crusoe, for example, takes this to the extreme, placing data centres at oil wells to tap into otherwise flared gas.

Another factor in unit economics is utilization rate. One challenge for Nebius is to attract enough workloads to keep tens of thousands of new GPUs busy. However, the company's latest disclosures show extremely strong uptake. By April 2025, Nebius's Annualized Run-Rate Revenue (ARR) reached $310 million, up from $249 million in March. This implies that newly added GPUs were getting filled with work rapidly. Nebius's ARR was growing 684% year-on-year (YoY) as of Q1 2025, indicating that as soon as Nebius deploys more capacity (e.g., a new cluster online), it's seeing customers take it up, likely attracted by the combination of availability and price.

GPU mix also plays a role. While H100s are the flagship, Nebius offers older and mid-range GPUs too, like A100s, L40s, or A10s. Many of these were acquired second-hand from miners. A GPU like the A10, which costs $3,500, might rent at $0.75/hour. At 80% utilisation, it brings in $5,200/year, allowing the investment to be recouped in less than 12 months. These legacy GPUs offer higher ROI and subsidise thinner margins on cutting-edge chips.

CoreWeave and Lambda Labs use similar strategies. Older GPUs like the A6000 or RTX 3090 have long since paid for themselves, but still rent out at attractive rates. These high-margin units improve blended economics and cushion the break-even curve on H100s or Blackwells.

Financials

Nebius's revenues are starting from a relatively small base but are growing at an extraordinary pace. In Q1 2025, the company reported revenue of $55.3 million, up 385% year over year, while ARR surged nearly 700% YoY, reaching $310 million by April 2025.

This growth, however, comes at a cost: Nebius remains in investment mode and is still loss-making as it builds out infrastructure. In Q1 2025, it posted an adjusted EBITDA loss of $62.6 million and a net loss from continuing operations of $113.6 million. That's wider than the $80.5 million loss in the same period last year, driven by increased operating expenses and depreciation. Still, there are early signs of operating leverage: opex as a percentage of revenue improved from an unsustainable 827% in Q1 2024 to 334% in Q1 2025, an indication that Nebius is gaining scale efficiency even as absolute costs rise.

Cash burn remains high. Operating cash outflow reached $197.8 million in Q1 2025 alone. But Nebius's balance sheet is strong. It holds $1.5 billion in cash and cash equivalents against just $6.2 million in debt, providing a solid funding runway.

Management has guided for full-year 2025 revenue of $500 to $700 million, potentially 5× higher than 2024 levels, with ARR approaching $1 billion. Longer term, the company targets a 30% adjusted EBIT margin, with room for additional upside as utilisation and gross margins improve.

Valuation

When Nebius re-listed on Nasdaq in October 2024, shares initially traded around $20, giving the company a market capitalization of roughly $4 billion—about where Yandex N.V. had been pre-suspension. At that price, with 2024 revenue at $118 million, Nebius was valued at an eye-watering 34× trailing price-to-sales (P/S). Even on a forward basis—say, $600 million in projected 2025 revenue, that implied a 7× forward P/S. A modest multiple for a company growing at 400%+.

Investors, drawn by the company's hypergrowth and AI focus, quickly bid up the stock in 2025. Today, Nebius trades at $47, giving it a market cap of $11 billion. At this price, the forward P/S climbs to 18×. The EV/Revenue multiple is slightly lower, thanks to Nebius' $1.5 billion in net cash, which brings enterprise value down to $9.7 billion, equating to 16× 2025E revenue.

Another lens is ARR. At $310 million in April 2025, Nebius trades at 35× ARR. This multiple reflects the market's expectation of sustained exponential growth, justified, to some extent, by its 684% YoY ARR growth at last disclosure.

CoreWeave, by comparison, has $2.7 billion in trailing revenue and is valued at around $71 billion, implying a trailing P/S of 26× and a forward P/S closer to 14×.

GPU count provides another useful valuation reference. Nebius currently has 30,000 GPUs and plans to add up to 60,000 in Finland and another 35,000 in Kansas City. At today's EV, that works out to $324,000 per GPU. CoreWeave, with an enterprise value of $76.9 billion and 250,000 Blackwell GPUs allocated, trades at $307,000 per GPU.

On a per-GPU basis, Nebius looks slightly more expensive. However, that spread may be justified by Nebius's capital structure. Unlike CoreWeave, which is heavily levered, Nebius is deploying GPUs primarily through cash, implying less financial risk and potentially better long-term margins. In that context, investors could rationally pay more per GPU for Nebius.

It's also worth noting: this valuation assumes Nebius's market cap reflects only its AI Studio business. If you assign any value to Toloka, TripleTen, or Avride, Nebius's implied EV per GPU drops below CoreWeave's. In that sense, the headline multiple overstates how richly Nebius is valued for its core cloud operations alone.

Finally, Nebius isn't just attracting retail momentum—several institutional investors and well-known funds are getting behind the story. In the most recent quarter, John Hussman (Trades, Portfolio), Jefferies Group (Trades, Portfolio), and Paul Tudor Jones (Trades, Portfolio) all increased their stakes in the company, suggesting rising institutional confidence in Nebius's long-term positioning.

Risks

While Nebius offers a compelling avenue to participate in the AI compute boom, the path from promise to profit is far from guaranteed. The first hurdle is execution. Scaling the fleet from ~30,000 GPUs to more than 60,000 across Finland, the U.K., and a new U.S. cluster will require flawless logistics, uninterrupted chip supply, and enterprise-grade reliability.

That leads into a second, more fundamental concern: utilisation risk. Nebius's economic model depends on sustaining high GPU usage—typically around 80%—to make the unit economics viable. But that level of utilisation isn't assured. CoreWeave, AWS, Google Cloud, and niche players like Lambda Labs are all aggressively competing for GPU workloads. Pricing pressure, customer churn, or simple overbuild could impair Nebius's ability to monetise its fleet efficiently.

Finally, Investors should also be aware that many view Nebius and its peers as “AI compute landlords”, by essentially renting GPUs like digital REITs. However, there's a key distinction. While REITs benefit from physical assets that depreciate slowly over decades, GPU hardware becomes obsolete within two to three years. The capital turnover is faster, the risk of stranded assets is higher, and the need to reinvest in cutting-edge chips like Blackwell or H200 is constant.

Conclusion

Nebius isn't just another AI infrastructure play. It's a true challenger in a market long dominated by hyperscalers. Its explosive growth, capital-light balance sheet, and global-first strategy set it apart from competitors like CoreWeave and Lambda. It also offers some diversification through its other business segments. While its GPU fleet is smaller for now, Nebius's cost advantages and platform breadth provide a credible path to scale. If management executes and demand for AI compute continues to accelerate, today's $12 billion valuation could prove to be a stepping stone, not a ceiling.

Ultimately, the AI gold rush won't reward every miner. The winners will be infrastructure providers that combine scale, capital efficiency, and customer-focused design. Nebius is building toward that vision, and it's already secured a seat at the table. I believe Nebius could be a market-beater over the long term.