Deep Dive 5: Nebius Group (NBIS) – The Technologist
Today: Nebius Group (NBIS)
Next Week: Lightmatter (private company)
Week After Next: Palantir (PLTR)
Lede
Today we do a deep dive into public company Nebius Group (NBIS) - one of the rare (if only) pure plays into AI data center infrastructure that is not based on a real-estate investment, but rather data centers for generative AI workloads.
Nebius went public on October 20, 2024.
One Weird Thing
We note that the company owns some other assets (included in the write up) which make little sense to us, distract from the pure play infra investment, and feel like a poorly aligned management with the ultimate goal of infrastructure.
Sometimes this is nothing more than a weird “eh, never mind.” Other times, in truth, it has been an early symptom of a company that is… not at all able to do what they claim they can.
Tell it to Me Like I’m 11-years Old
Imagine a giant playground where computers work together to solve puzzles and learn new tricks, called "the cloud."
Nebius Group builds a special kind of playground (cloud) for computers that is different from the huge ones run by big cloud companies called hyperscalers.
They create super-efficient computer buildings where thousands of powerful machines work together without wasting energy.
They even design their own computer parts to make everything run perfectly, just like building a custom LEGO set.
Instead of being as enormous as the giants, Nebius focuses on a smaller, specialized cloud that fits unique needs better.
Their data centers use smart tricks to stay cool and even reuse extra heat to warm nearby homes.
With AI and cloud infrastructure taking over the world and driving huge demand, Nebius rides this massive tailwind to help companies train computers to do amazing things while saving energy and money.
In short, Nebius Group offers a fun and efficient computer playground that is a great alternative to the giant cloud playgrounds run by hyperscalers.
Idea
Before we dive into NBIS, there is an idea floating around in the depths of tech and infra:
The major cloud providers (“hyperscalers”) posted slower-than-expected growth across the board—AWS at 13% in Q4 2023 (19% in Q3 2024), Microsoft Azure at 19% (20%), and Google Cloud at 26% (35%).
This all while their capital expenditures (CapEx) have reached astronomical levels and while top $250B in 20205 alone.
Now here’s the thought…
… It could be the case that enterprises are increasingly repatriating workloads from public clouds due to hidden costs and subpar AI returns, while the rise of specialized AI providers, microclouds, and edge computing signals a shift toward tailored, hybrid, or private cloud solutions that better meet specific needs.
Specialized AI providers and microclouds, like NBIS and Paperspace (owned by DigitiaOcean (DOCN)) are emerging as attractive alternatives because they offer tailored, high-performance solutions for machine learning and deep learning workloads.
As a result, the cloud market is evolving into a more diverse and specialized landscape.
This is where NBIS fits in and this is why Nvidia invested $700M into the company - to diversify the dominance of the cloud away from the hyperscalers, if even a little.
Preface
Nebius Group is a Europe-based technology company that provides AI infrastructure and cloud services to “AI builders” around the world (advfn.com).
In its investor presentation, Nebius outlines a strategy centered on delivering an AI-native cloud platform powered by large GPU clusters.
Nebius Group’s formation is a remarkable case of corporate transformation under geopolitical pressure.
Born from Yandex’s desire to safeguard its advanced tech projects from the fallout of the Russia-Ukraine conflict, Nebius went from an idea in late 2022 to a fully independent company by mid-2024.
Before we get to the company, we will need to do a deep (ish) dive on the history of this firm because it is quite unconventional.
Formally, the company introduces itself like this:
Nebius is a technology company building full-stack infrastructure to service explosive growth of the global AI industry, including large-scale GPU clusters, cloud platforms, and tools and services for AI developers.
Headquartered in Amsterdam and listed on Nasdaq, the company has a global footprint with R&D hubs and offices across Europe, North America and Israel.
Nebius' core business is an AI-centric cloud platform built for intensive AI/ML workloads.
With proprietary cloud software architecture and hardware designed in-house (including servers, racks and data centre design), Nebius provides AI and ML practitioners the compute, storage, managed services and tools they need to build, tune and run their models.
The group also operates additional three businesses under their own distinctive brands:
Toloka AI – data partner for all stages of AI development from training to evaluation: humans training machines;
Avride - one of the world’s most experienced self-driving teams: driverless cars and delivery robots;
TripleTen - a leading edtech platform specialising in reskilling individuals for successful careers in tech.
Those last three are the ‘weird’ businesses that do add significant risk in my opinion.
Uniquely Efficient
Before we get into the nuts and bolts, there is a section I’d like to put first and it surrounds the unique efficiency that Nebius Group claims to deliver:
Nebius Group operates one of the world’s most energy-efficient data centers, leveraging in-house hardware design, advanced cooling, and heat recovery systems to achieve sustainability and performance on par with industry giants like Google and Microsoft.
Power Usage Effectiveness (PUE):
Nebius' data center in Finland has achieved a PUE of 1.13 under maximum loads, indicating high energy efficiency. cdn.prod.website-files.com
Here’s a chart from the investor presentation:
In-House Designed Hardware:
The company designs and assembles its servers and racks in-house, optimizing them for intensive workloads and enhancing energy efficiency. nebius.com
Advanced Cooling and Heat Recovery Systems:
The data center employs free cooling techniques, utilizing external ambient air to cool servers, and operates efficiently at temperatures up to 40°C. This design reduces the need for subcooling and enhances energy efficiency.
Additionally, waste heat from servers is repurposed to heat nearby residential buildings, further improving energy utilization.
These features collectively position Nebius' data center among the most energy-efficient globally, comparable to facilities operated by tech giants like Google and Microsoft.
Background
Nebius Group N.V. is a technology company that emerged in 2024 as a spin-off from Yandex, Russia’s largest internet tech firm.
It was created when Yandex’s Dutch holding company separated from its Russian operations amid geopolitical challenges, rebranding itself as Nebius and refocusing on artificial intelligence (AI) infrastructure (en.wikipedia.org) (datacenterdynamics.com).
Nvidia and Accel Ventures recently invested $700 million in Nebius to help the company expand its AI infrastructure (Source).
As of September 30, 2024, the group employed around 1,300 people.
Below is a comprehensive overview of the key events leading to Nebius Group’s formation and the leadership team that spearheaded its establishment.
Timeline of Nebius Group’s Founding as a Yandex Spin-off
•1989: Arkady Volozh, a pioneering Russian technologist, establishes the company that would later become Yandex N.V., originally as a holding entity for Yandex’s businesses (en.wikipedia.org).
(Yandex would go on to become known as “Russia’s Google,” dominating the country’s search and online services.)
• May 2011: Yandex N.V. raises $1.3 billion in an initial public offering on the NASDAQ stock exchange – the biggest IPO for an Internet company since Google’s 2004 debut (en.wikipedia.org).
• This listing solidified Yandex’s international presence and provided capital for its expansion.
• February 2022: Russia’s invasion of Ukraine triggers international sanctions that profoundly affect Yandex’s business structure (en.wikipedia.org).
• By March 2022, trading of Yandex N.V.’s shares on NASDAQ was suspended due to the sanctions-related volatility (en.wikipedia.org).
• June 2022: Arkady Volozh – Yandex’s principal founder and long-time CEO – resigns from all positions at Yandex after being placed on the European Union’s sanctions list (marketscreener.com).
He called the EU’s decision “misguided,” but stepped down to shield the company (marketscreener.com).
Following his exit, Yandex’s board (led by Chairman John Boynton) began exploring strategic options for the company’s future.
• November 2022: Yandex N.V.’s board announces a major restructuring plan in response to the geopolitical situation (marketscreener.com).
The plan proposed divesting Yandex’s core Russian businesses – including its search, advertising, and ride-hailing services – to local investors, and spinning off the company’s key international projects into separate entities (marketscreener.com) (marketscreener.com).
• Specifically, Yandex identified four units with global potential to be developed independently outside Russia: its self-driving technologies, cloud computing services, data-labeling/crowdsourcing platform, and education tech business (marketscreener.com).
This move was aimed at insulating the high-tech ventures from political risk and potential Kremlin influence, while allowing the Russian arm of Yandex to continue under new ownership.
• December 2022: In an internal farewell message to staff, Arkady Volozh bids goodbye ahead of the impending corporate split (marketscreener.com).
He notes that he hasn’t been involved in the Russian side of the business for some time due to his exit, and expresses hope for the future of the four new international startups born from Yandex’s technology units.
• Volozh offers to advise these ventures – which would later form the core of Nebius Group – as they chart an independent path focused on cloud (AI infrastructure), autonomous vehicles, crowdsourced data (Toloka), and online education
This message underscored a turning point: Yandex as known for decades would effectively split, with its next year (2023) dedicated to executing the separation.
• 2023: Through the first half of 2023, Yandex N.V. negotiates the sale of its Russian assets and reorganizes the identified units into a cohesive new company. By early 2023, the “Nebius” brand began to be used internally for the burgeoning cloud and AI-focused business in preparation for the spin-off (datacenterdynamics.com).
The Amsterdam-based holding firm (Yandex N.V.) works to consolidate the non-Russian operations and ensure it can operate entirely independent of the Russian entity.
In March 2024, Arkady Volozh is removed from the EU sanctions list, a development that frees him to take a direct leadership role in the new company (en.wikipedia.org).
• July 15, 2024: Yandex N.V. completes the sale of all its Russia-based businesses to a consortium of Russian investors, in a deal worth $5.4 billion in cash and shares (au.marketscreener.com) (verdict.co.uk).
This transaction is described as the largest corporate exit from Russia since the war began (albeit at a steep discount to Yandex’s pre-war valuation) (au.marketscreener.com).
• With this deal, all of Yandex’s Russian operations (including the Yandex.ru search engine, e-commerce, and other domestic services) are transferred to the new local owners.
Yandex N.V. retains only the international-focused businesses, namely: its AI-centric cloud computing platform (later branded Nebius.AI), the Toloka AI data-labeling crowdsourcing platform, the self-driving and robotics unit (branded Avride), and the online education platform TripleTen (datacenterdynamics.com) (interfax.com).
Essentially, these units — which had been operating in Europe, Israel, and other global markets — now become the core of the new company moving forward.
Immediately after the sale, Yandex N.V. begins operating under the Nebius name (initially as a brand) and Arkady Volozh formally returns to lead the company as CEO, after two years out of management (interfax.com).
At this point Nebius has over 1,000 employees and a strong financial base (around $2.5 billion in cash on hand) to fuel its new direction (interfax.com).
• August 2024: An annual general meeting of shareholders formally approves the renaming of Yandex N.V. to Nebius Group N.V. and appoints a new Board of Directors for the spin-off (au.marketscreener.com) (interfax.com).
With shareholder sign-off, the company officially adopts the Nebius identity and discontinues the Yandex name for its international business.
Around this time, Arkady Volozh’s appointment as Chief Executive Officer is ratified by shareholders as well.
Nebius Group, headquartered in Amsterdam, is now completely independent from its Russian roots, owning no Russian legal entities and having no business in Russia (group.nebius.com).
By late August, Nebius positions itself publicly as a new enterprise focused on building one of Europe’s leading AI infrastructure platforms, free from the restrictions that had affected Yandex during the sanctions period.
(Note: In the aftermath of its founding, Nebius quickly pursued growth initiatives – for example, announcing plans to invest $1 billion in AI infrastructure and raising $700 million from investors like Nvidia and Accel (verdict.co.uk)).
Leadership and Key Personnel in Nebius’ Establishment
Nebius Group’s creation brought together a leadership team comprising both veteran Yandex figures and new experts, under a structure committed to strong international governance.
The company’s Board of Directors is led by John Boynton as Chairman, and Arkady Volozh as Chief Executive Officer, alongside a team of senior executives who were instrumental in launching Nebius
Below are the key people and their roles, backgrounds, and contributions:
Business Model
Nebius Group’s core business model revolves around AI-Infrastructure-as-a-Service. The company offers on-demand access to high-performance GPU computing so that clients can train and deploy AI models without managing any hardware.
Revenue is generated primarily through cloud services usage fees – essentially selling compute time, storage, and related AI cloud services to customers building AI applications.
Service Offerings: Nebius has built a full-stack AI cloud platform that integrates its own hardware and software. Key elements of the business include:
• Large-Scale GPU Infrastructure: Nebius operates massive GPU clusters and makes them available as a cloud service.
It designs and manages these clusters in-house, which allows for optimized performance and cost.
The platform is built on NVIDIA’s reference architecture and includes thousands of the latest GPUs for intensive AI workloads (group.nebius.com).
• Cloud Platform & Developer Tools: On top of the raw compute, Nebius provides a suite of cloud services and tools tailored for AI developers.
This includes managed Kubernetes for AI (for containerized model training/serving), distributed computing orchestration (e.g. Slurm job scheduling), data processing tools, and other developer services (cdn.prod.website-files.com).
These offerings enable customers to focus on AI development while Nebius handles the underlying infrastructure.
• Own Data Center Operations: A key pillar of Nebius’s model is owning and operating its data centers, which improves unit economics.
By controlling facility design and energy management, Nebius achieves higher energy efficiency and lower costs per unit of compute (cdn.prod.website-files.com).
For example, its primary data center in Finland runs with an industry-leading power usage efficiency (~1.1 PUE, vs ~1.58 global average) – reducing operational costs and allowing more competitive pricing (group.nebius.com).
Market Approach: Nebius is positioning itself to ride the explosive growth in demand for AI infrastructure.
The company cites a large and growing total addressable market for AI cloud services – internal and external estimates put the opportunity at $150+ billion by 2030 for the segments Nebius targets (scribd.com).
Rather than compete as a general-purpose cloud provider, Nebius focuses on the niche of AI workloads.
This focused approach, coupled with owning the full stack (from data center hardware up to software), aims to deliver better performance-per-dollar for AI use cases than traditional clouds.
In summary, Nebius’s business model centers on vertically integrated, AI-optimized cloud services that generate revenue from usage fees and scale with the burgeoning AI market.
AI Infrastructure
At the core of Nebius’s value proposition is its AI infrastructure – the physical and software infrastructure enabling cloud-based AI computing.
We present a detailed breakdown of Nebius’s data centers, GPU clusters, and platform capabilities:
• Data Centers & GPU Clusters: Nebius currently operates a state-of-the-art data center in Mäntsälä, Finland, which it fully owns. This facility is packed with GPU servers – about 14,000 GPUs have been paid for and deployed there as of Q3 2024 (cdn.prod.website-files.com).
Thanks to Finland’s cool climate and Nebius’s engineering, the site achieves exceptional efficiency (PUE ~1.1), well above industry average (group.nebius.com).
Nebius is aggressively expanding capacity: it announced plans to triple the Finnish data center’s capacity to 75 MW, enough to support up to 60,000 GPUs at that location alone (sec.gov).
This scale would make it one of Europe’s largest AI compute hubs.
• New European GPU Cluster (Paris): To extend its footprint and serve Western Europe, Nebius launched a new GPU cluster in Paris, France in late 2024.
This cluster is hosted in an Equinix data center and is part of Nebius’s plan to invest $1B+ in European AI infrastructure by mid-2025 (capacitymedia.com).
Notably, the Paris zone is among the first in Europe to offer NVIDIA’s latest H200 Tensor Core GPUs, giving Nebius’s customers cutting-edge hardware for AI training (group.nebius.com).
The Paris deployment underscores Nebius’s strategy of quickly standing up GPU capacity in key regions to capture demand.
• US Expansion (Planned): Nebius is also expanding into North America. It has announced its first U.S. GPU cluster in Kansas City, scheduled to go live in Q1 2025 (group.nebius.com).
The Kansas City cluster will house thousands of NVIDIA’s newest GPUs (H200 generation) in its initial phase (group.nebius.com).
Along with this, Nebius opened offices in San Francisco, Dallas, and New York to support U.S. clients.
This move into the U.S. market will position Nebius to compete globally and access the deep pool of AI startups and enterprises in North America.
• AI-Optimized Cloud Platform: On the software side, Nebius’s cloud platform is purpose-built for AI workloads.
The infrastructure is configured with InfiniBand networking and high-bandwidth interconnects between GPU nodes, which are critical for distributed training of large AI models (cdn.prod.website-files.com).
Nebius offers a fully managed Kubernetes service tailored to machine learning, as well as support for HPC job schedulers (with a Slurm-based orchestration system) to efficiently allocate jobs across the GPU cluster (cdn.prod.website-files.com).
In practice, this means customers can spin up clusters of tens or hundreds of GPUs with minimized networking bottlenecks – suitable for training large language models and other compute-intensive tasks.
Nebius also provides tools for data management and model deployment (including an AI inference platform), aiming to cover the end-to-end needs of AI developers.
Overall, Nebius’s infrastructure strategy is to build and own cutting-edge GPU supercomputers in multiple regions and expose them to users via a cloud platform.
By controlling everything from data center construction to cluster management software, Nebius seeks to deliver high-performance AI computing at lower cost.
This heavy investment in infrastructure is the foundation for the company’s competitive differentiation in the AI cloud market.
Target Customers
Nebius’s target customers are the organizations and developers driving the AI revolution – a segment it often refers to as “AI builders.” The investor presentation and related analyses highlight a few key client segments:
• AI Startups and SMBs: Nebius explicitly tailors its offering to smaller and mid-sized companies building AI-powered products.
These clients often need significant GPU resources to train models but may find major cloud providers cost-prohibitive. Nebius’s strategy of providing cost-effective solutions for SMBs and AI startups has been a growth driver (stockanalysis.com).
By offering lower prices (thanks to its efficient infrastructure and focus on core compute) and flexible usage terms, Nebius makes advanced AI hardware accessible to startups that cannot afford large capital investments.
This segment values Nebius’s pay-as-you-go model and high-performance hardware to iterate quickly on AI models.
• Enterprise AI Teams and Research Labs: In addition to startups, Nebius also serves larger enterprise teams and research institutions that require dedicated AI compute capacity.
These customers might already use public clouds but are looking to avoid vendor lock-in or to reduce costs for specific AI workloads.
Nebius fills a market gap by providing compute-only GPU access without the ecosystem lock-in that hyperscalers often impose (reddit.com).
For example, an R&D team that wants raw GPU instances for a sensitive AI project can use Nebius and not be forced into a broader proprietary stack (unlike with, say, AWS where ancillary services and architectures create lock-in).
Enterprises with data residency concerns may also prefer Nebius’s European data centers for compliance reasons.
While Nebius is still young, it positions itself as a flexible alternative for any organization building AI systems at scale, promising more transparent pricing and control.
• Global AI Developers and Communities: Because Nebius operates globally via the cloud, it can attract individual developers or AI communities as well.
Anyone developing large-scale AI models – from Kaggle grandmasters to open-source AI project teams – is a potential customer if they need on-demand access to GPU clusters.
Nebius’s documentation and tools are geared toward modern AI/ML workflows, which helps draw in tech-savvy users.
By being cloud-based, Nebius can serve clients in multiple regions who need low-latency, local compute.
The launch of a U.S. cluster, for instance, will make Nebius more appealing to North American customers who previously may have been concerned about latency or data jurisdiction.
Differentiation from Hyperscalers: Across these segments, Nebius’s differentiation lies in its focus and flexibility. Unlike Amazon, Microsoft, or Google Cloud, Nebius is not trying to sell a broad menu of IT services – it is laser-focused on AI infrastructure, which lets it optimize everything for that purpose. This focus yields tangible benefits for customers:
• Lower Cost Structure: Because Nebius avoids the heavy overhead and high profit margins of general-purpose clouds, it can offer GPU computing at lower prices.
Hyperscalers often charge a premium for GPU instances (since they bundle them with other services and enterprise contracts), whereas Nebius’s lean approach passes on savings.
Independent analyses note Nebius’s platform is highly cost-effective for AI workloads, which is especially attractive to startups on a budget (stockanalysis.com).
• No Lock-In, Open Ecosystem: Nebius emphasizes interoperability and freedom.
Clients are not tied into a complex web of proprietary services – they rent GPUs and use open-source or preferred ML tools on Nebius.
This appeals to customers who want more control or the ability to easily port their workloads elsewhere. As one analysis put it, Nebius provides pure GPU cloud access “without the cloud ecosystem lock-in” of the big providers (reddit.com).
In fast-evolving fields like AI, this flexibility is a significant draw.
In summary, Nebius is targeting the innovators in AI – from scrappy startups to advanced enterprise labs – who need large-scale compute on demand.
By offering a specialized, cost-efficient AI cloud, Nebius aims to win these customers away from both do-it-yourself on-prem setups and the expensive, generalist public clouds.
Competitive Positioning
Nebius operates in an increasingly crowded landscape of cloud and AI infrastructure providers. Its investor presentation outlines how the company positions itself against both the hyperscale cloud giants and the emerging specialized players:
• Versus Hyperscalers (AWS, Azure, GCP): Nebius’s main competitors are the big three cloud providers, which all offer GPU instances and AI services.
However, Nebius differentiates through specialization and efficiency.
The company’s single-minded focus on AI workloads allows it to optimize performance and pricing in ways hyperscalers (with their broad offerings and legacy systems) struggle to match.
For instance, Nebius’s custom-built data center has a PUE of ~1.1, far more efficient than typical data centers (avg ~1.58), translating to lower operating costs (group.nebius.com).
• This efficiency edge is passed to customers via lower prices for comparable GPU power.
Moreover, Nebius doesn’t bundle unnecessary services – clients pay for the GPUs and networking they need, not a premium “enterprise package.”
In effect, Nebius is aiming to be the low-cost, high-performance provider for AI computing, undercutting hyperscalers on key AI jobs.
Its AI-native cloud platform is also an advantage: Nebius’s software stack is built from the ground up for machine learning, whereas hyperscalers have to retrofit or layer AI services on general cloud infrastructure.
This means Nebius can sometimes roll out features (like new GPU types or ML workflow tools) faster and with less bloat.
• Versus Specialized AI Cloud Startups: Nebius is not alone in targeting this space. Other “neo-cloud” companies focusing on AI have sprung up, such as CoreWeave, Lambda Labs, Vultr, and others – many of them in the U.S. Nebius is very much part of this cohort of new providers that have collectively raised significant capital to take on the incumbents (techblog.comsoc.org).
• Within this peer group, Nebius’s distinguishing factors include its geographic base and its origin.
It is one of the leading AI cloud players based in Europe, which gives it first-mover advantage in that region (most rivals are US-based).
Being in Europe positions Nebius to capture EU demand and support European initiatives for AI infrastructure sovereignty.
Additionally, Nebius inherited technology and talent from Yandex (the Russian tech giant) when it was spun out, giving it a mature starting point.
The company boasts ~850 top-tier engineers with AI/ML and cloud expertise as of late 2024 (scribd.com) – a substantial human capital advantage for a startup.
This helps Nebius rapidly develop its platform and scale operations to compete with better-known rivals.
Here is a formal list of competitors:
CoreWeave
Services & Infrastructure: CoreWeave is a GPU-centric cloud platform built on a Kubernetes-native architecture for bare-metal performance (ankursnewsletter.com).
It provides a broad range of NVIDIA GPUs (including the latest H100s) with ultra-fast networking (up to 400 Gb/s InfiniBand) to support intensive AI workloads(ankursnewsletter.com).
Instances can be spun up within seconds, allowing rapid scaling on demand (ankursnewsletter.com).
CoreWeave also supports specialized configurations for distributed training, boasting training speeds up to 35× faster than some traditional clouds for large ML models (ankursnewsletter.com).
Target Customers & Use Cases: CoreWeave caters to industries and teams that require massive compute on demand. For example, it's used in visual effects (VFX) rendering and animation (eliminating traditional render queues), large-scale scientific or financial simulations, and generative AI model training where quick scaling to many GPUs is essential (ankursnewsletter.com).
Its fast provisioning and varied GPU options make it appealing to AI startups and researchers who need cloud GPUs on tap for experiments and production.
Key Differentiators: CoreWeave positions itself as a cost-effective alternative to hyperscalers, claiming workloads can be run for up to 80% less expense than on major public clouds (ankursnewsletter.com).
.This dramatic cost advantage comes from usage-based billing and efficient provisioning (important for budget-conscious AI teams).
It also offers bare-metal performance (no virtualization overhead) and a cloud-native toolkit, which gives experienced users fine-grained control.
CoreWeave’s focus on GPUs and high-speed interconnects is purpose-built for AI, making it a “GPU cloud, purpose-built for AI” rather than a general cloud
Its ability to scale quickly (instances in ~5 seconds) and handle diverse GPU types is a game-changer for dynamic AI workloads.
Lambda Labs (Lambda Cloud)
Services & Infrastructure: Lambda Labs offers an AI cloud that provides on-demand GPU clusters and also sells physical AI hardware. Their Lambda Cloud features a one-click cluster deployment, where users can spin up multi-GPU setups (including up to 8× NVIDIA H100 Tensor Core GPUs connected via InfiniBand) in minutes (ankursnewsletter.com).
The environment comes pre-configured with popular ML frameworks (TensorFlow, PyTorch, CUDA, etc.), creating a turnkey solution for training models without complex setup.
For those with on-premise needs, Lambda supplies high-end GPU workstations and servers, bridging cloud and on-prem AI workflows.
Target Customers & Use Cases: Lambda’s cloud is geared toward AI researchers, engineers, and startups that need to train large models (like NLP large language models) or run other intensive deep learning tasks.
Its support for multi-thousand-GPU clusters and high-performance networking makes it ideal for cutting-edge generative AI development, large-scale model fine-tuning, and advanced computer vision or scientific ML projects
Because Lambda does not require long-term contracts for big GPU clusters, it’s popular among startups and labs that want to quickly prototype or scale up experiments cost-effectively (ankursnewsletter.com).
Key Differentiators: Lambda Labs distinguishes itself with developer-friendly simplicity and transparent pricing. The platform’s ease of use (pre-set ML environments and one-click cluster deployment) allows teams to start training on powerful GPUs with minimal DevOps overhead.
In terms of cost, Lambda’s prices are often lower and more straightforward than hyperscalers – for example, on-demand A100 40GB GPU instances start around $1.25/hour, and H100 instances around $2.49/hour
Paperspace (owned by DigitalOcean (DOCN)
Services & Infrastructure: Paperspace is a cloud provider centered on GPU computing for AI and computational workloads.
It offers a range of NVIDIA GPU instances and a user-friendly platform called Gradient, which provides managed Jupyter notebooks and ML pipelines. Developers can launch GPU machines with minimal configuration, or use Paperspace’s hosted notebook environments to write and run code in the browser.
The service supports a wide array of GPU types (from older cards to modern RTX/A100) and includes features like persistent storage and team collaboration tools.
Target Customers & Use Cases: Paperspace primarily targets developers, small teams, and educators in AI who want an easy-to-use, low-friction environment for model development, training, and inference.
Its free-tier options (such as limited free GPU time on Gradient) and one-click templates for popular models and frameworks make it especially appealing to students and startup teams on a budget.
Common use cases include prototyping deep learning models, running tutorials or experiments, and even GPU-accelerated rendering or data processing. Paperspace has a community of over 500,000 users, indicating its popularity among individual AI practitioners and developers.
Key Differentiators: Unlike many cloud providers, Paperspace emphasizes an accessible developer experience. Users can get started with a free account and even access free GPUs in certain instances, lowering the barrier to entry (paperspace.com).
It provides ready-to-use environments (with pre-installed ML libraries) and a catalog of starter templates and projects, which speeds up onboarding for AI workflows
While it similarly offers state-of-the-art GPU hardware on demand (comparable to CoreWeave or Lambda in that sense), Paperspace sets itself apart by focusing on simplicity and community features rather than just raw performance.
This makes it a strong choice for those who want to focus on coding and experimentation without managing complex infrastructure – effectively a vertically integrated alternative to big clouds, but with a developer-first approach.
Cerebras
Services & Infrastructure: Cerebras is unique in offering cloud access to its proprietary Wafer-Scale Engine (WSE) – the largest AI chip ever built. Its flagship hardware, the Cerebras CS-2 system, contains an entire silicon wafer as a single chip, with 2.6 trillion transistors and 850,000 AI-optimized cores on one processor (ankursnewsletter.com).
This massive on-chip compute and 40 GB of on-wafer memory allow it to handle AI tasks that would normally require a large GPU cluster.
Cerebras provides access to CS-2 systems through partnerships (for example, via the Cirrascale cloud services) and its own Cerebras Cloud offerings, often as part of the Cerebras AI Model Studio for training large models (cerebras.ai).
Essentially, clients can rent time on a CS-2 or a cluster of CS-2s to run extremely heavy workloads.
Target Customers & Use Cases: Cerebras is geared towards organizations pushing the limits of AI and HPC. Its architecture shines in high-performance computing tasks like complex physics simulations, climate modeling, and scientific research that require huge parallelism (ankursnewsletter.com).
It’s also very well-suited for natural language processing (NLP) at extreme scale – for instance, training or fine-tuning giant language models that would be impractically slow or expensive on a normal GPU cluster
Early adopters include national labs (for scientific workloads) and cutting-edge AI research groups.
Use cases such as genomics (in medical AI) or real-time weather forecasting have been cited, where the ability to load very large models or datasets into one physical compute unit can dramatically speed up processing.
In summary, Cerebras targets scenarios where a single system with extraordinary compute can replace dozens or hundreds of traditional servers, simplifying the job.
Key Differentiators: The main differentiator is Cerebras’ custom silicon – it’s the most specialized hardware in this space. A single CS-2 offers what would ordinarily require an entire GPU supercomputer: it has 123× more cores than a leading GPU and orders-of-magnitude more on-chip memory and bandwidth.
This specialized design eliminates the need for inter-GPU communication for very large models, which can yield unprecedented performance on certain AI tasks. By condensing a cluster into one chip, Cerebras can also reduce energy usage and administration complexity for those workloads.
However, this comes at high cost and a narrower focus – it’s a “Ferrari” of AI hardware (exceptional on the right track, but not flexible for every task).
For organizations that truly need the absolute peak throughput for massive models, Cerebras provides an option that hyperscalers can’t match with standard GPUs.
Graphcore (IPU Cloud via Graphcloud)
Services & Infrastructure: Graphcore is a semiconductor company that offers an alternative AI processor called the Intelligence Processing Unit (IPU). Through its Graphcloud service, Graphcore provides cloud access to IPU-Pod systems (hosted by Cirrascale) for AI workloads (graphcore.ai).
Users can request IPU instances which come pre-installed with Graphcore’s Poplar software stack and tools for machine learning.
Graphcloud allows integration with existing workflows (secure connectivity to AWS, Azure, GCP or on-prem environments), so users can treat IPUs as a specialized extension of their AI infrastructure.
In practice, a Graphcloud deployment might consist of a pod of many IPU chips working in parallel, available under a straightforward subscription or usage-based model.
Target Customers & Use Cases: Graphcore’s IPUs target innovators in AI research and industry who are exploring models and algorithms that benefit from massive parallelism and fine-grained compute. IPUs have been highlighted for workloads like natural language processing, graph neural networks (GNNs), and financial computing.
For instance, companies doing cutting-edge NLP or recommender systems might test and run their models on IPUs to see performance gains.
Research labs and universities are also a key audience – Graphcore has made IPUs available to academic researchers to experiment with novel neural network architectures.
In general, any AI task that can exploit high parallel core counts and requires fast memory access (e.g., large sparse models, some computer vision tasks) could be a good candidate. IPUs can excel at handling models that don’t map efficiently to GPU architectures, so they cater to those looking for alternatives to GPU-centric approaches.
Key Differentiators: The differentiator is the IPU architecture itself, which is designed from the ground up for AI. A single IPU system can deliver over 1.6 petaFLOPS of AI compute and is optimized for high throughput on complex models.
Graphcore claims performance advantages on certain workloads versus GPU systems, thanks to features like massive parallel cores and very high on-chip memory bandwidth
In practical terms, this means an IPU pod might train or infer some models faster than a comparable GPU setup, especially for models with irregular computation patterns.
The Graphcloud service also touts ease-of-use and integration – IPUs can be accessed on-demand without needing to buy specialized hardware, and the ecosystem (Poplar SDK) is provided to help port models.
While still a nascent alternative, Graphcore’s cloud offering gives AI teams a way to experiment with novel silicon that could unlock better performance for specific AI challenges (like NLP, real-time risk analysis, or GNNs).
SambaNova Systems
Services & Infrastructure: SambaNova is an AI computing company that offers a full-stack AI platform centered on its custom Reconfigurable Dataflow Unit (RDU) chips.
Its cloud-based solution, often referred to as Dataflow-as-a-Service, provides access to SambaNova’s hardware and software for training and deploying AI models. Rather than renting raw VMs or GPUs, clients use SambaNova’s service to run models on SambaNova’s DataScale systems (which contain multiple RDU processors).
The platform is delivered as a managed service with pre-built models (for example, GPT language models, Vision Transformers, etc.) that can be fine-tuned or deployed for enterprise use (hpcwire.com).
In essence, SambaNova offers an AI-as-a-service approach: companies can leverage state-of-the-art models on SambaNova hardware in the cloud (or on-premises appliance) without dealing directly with low-level infrastructure.
Target Customers & Use Cases: SambaNova targets enterprise and government customers looking to quickly implement AI solutions, especially in areas like natural language processing and computer vision.
Its offerings have been tailored for specific industries – for instance, financial services (GPT for banking), public sector analytic needs, and scientific research institutions.
Because SambaNova provides ready-to-use large models (e.g., GPT-style models for text, recommendation models, etc.), a typical use case is a large organization that wants a private, high-performance model for their domain (say, an insurance company building a claims analysis AI or a government agency doing large-scale language translation) without relying on public cloud APIs.
The ability to deploy on-premise if needed (for data privacy or sovereignty) also appeals to customers in regulated industries (sambanova.ai).
Early adopters include national labs (Argonne National Lab deployed SambaNova for scientific AI) and enterprises that need advanced AI with a turnkey approach.
Key Differentiators: SambaNova’s value proposition lies in its integrated hardware-software stack and pre-trained models.
Its RDU chips and Dataflow architecture are highly specialized for AI, often yielding performance improvements over general GPUs.
For example, SambaNova has demonstrated faster training throughput on GPT models compared to equivalent GPU setups.
The platform abstracts away the complexity of distributed training; clients can get running with a state-of-the-art model in “days” rather than investing months in building infrastructure.
Another differentiator is the focus on privacy and flexibility – because the solution can be deployed in SambaNova’s cloud or in the customer’s own data center, organizations are not forced into a public cloud and can keep sensitive data in-house
In summary, SambaNova serves those who need specialized AI hardware performance but packaged in an easy-to-consume service with enterprise support.
It’s less about renting chips and more about delivering an AI outcome (with the custom chips operating behind the scenes).
Back to Nebius…
• Technology & Partnerships: Nebius has strategically aligned itself with key industry partners, most notably NVIDIA. In fact, NVIDIA not only supplies GPUs but has also invested in Nebius (taking part in a recent $700M financing round) (investors.com).
• Nebius is an NVIDIA Preferred Cloud Service Provider and adheres to NVIDIA’s reference designs for AI clusters, which means its hardware and software configurations are validated by NVIDIA for optimal AI performance (nebius.com).
• This partnership ensures Nebius early access to the latest GPU technology and deep technical support – on par with what hyperscalers receive.
By deploying cutting-edge GPUs like the A100, H100, and now H200 at scale, Nebius can offer the same (or better) hardware generation as larger competitors.
Furthermore, Nebius’s focus on open-source tools and standard frameworks makes it easier for customers to adopt its platform without rewriting code, which is a subtle competitive edge over clouds that push proprietary AI toolchains.
• Value Proposition: Summing up its competitive stance, Nebius portrays itself as a “best-of-both-worlds” option: the power and scale of a cloud, but the cost-efficiency and flexibility of a specialized provider.
It addresses a pain point where hyperscalers often charge high margins for AI services – Nebius claims to operate with lower overhead and is willing to sacrifice short-term margins to gain market share.
By filling this gap, Nebius hopes to attract the wave of companies developing generative AI and large language models who might otherwise either overpay big cloud vendors or struggle with limited on-prem resources.
Its recent stock performance reflects some market confidence in this positioning – Nebius Group’s shares more than doubled in the past year, driven by the company’s unique AI-native platform and value offering for startups/SMBs in particular (stockanalysis.com).
In conclusion, Nebius is carving out a spot as a specialist AI infrastructure provider that competes on focus, performance-per-dollar, and partnerships.
It stands against giants by being lean and tailored, and stands out from fellow startups by virtue of its European stronghold and deep engineering roots.
The competitive race in AI cloud is just beginning, but Nebius’s positioning indicates it aims to be a prominent challenger in this domain.
Financials
Nebius Group’s financial profile, as presented to investors, reflects a company in high-growth mode with significant investment plans. Key financial insights include:
Straight from the company:
Depending on the amount of capital available to cover the company’s investment program, we expect our core business to deliver ARR (annualized run-rate revenue) of $500mn to $1bn by the end of 2025.
• Explosive Revenue Growth: Nebius is experiencing rapid revenue acceleration as its AI cloud services gain traction.
In the first nine months of 2024 (Jan–Sept), the company generated $79.6 million in revenue, a 5.6× increase over the same period in 2023 (freedom24.com).
This growth steepened in the most recent quarter – Q3 2024 alone saw revenue of $43.3 million, which was up 1.7× from the prior quarter (Q2’s ~$25M) (freedom24.com) (enterprisetimes.co.uk).
• The annualized run-rate revenue as of Q3 2024 was around $120 million, and Nebius indicated it was on track to reach approximately $170–$190 million in revenue for the full year 2024 (group.nebius.com).
• Such numbers, while still modest in absolute terms, demonstrate an extremely fast growth trajectory for a young infrastructure company.
The surging demand for AI compute (especially after the launch of generative AI applications in 2023) has directly translated into customer uptake for Nebius’s services.
• Profitability and Margins: As expected for a company building out infrastructure, Nebius is not yet profitable.
The heavy investments in data centers, hardware procurement, and engineering talent have meant operating losses in the near term.
The investor presentation didn’t quote current profit margins, but Nebius did share an outlook that it aims to approach breakeven on an adjusted EBITDA basis by 2025 (cdn.prod.website-files.com).
• In other words, management is targeting to scale revenues to a level by next year where the core business can cover its operating costs.
For 2024, one can infer negative EBITDA given the expansion stage, but the path to profitability is envisioned as a function of volume: once utilization of those GPUs grows and revenue scales 3–4×, Nebius expects economies of scale to kick in.
The single sell side analyst that covers NBIS has them reaching positive adjusted EBITDA by Q4 2025.
Gross margins should improve as their data centers fill up capacity (idle servers are sunk cost), and the fixed costs (like engineering R&D) will be spread over a larger revenue base.
Investors are thus told to expect a rapid ramp toward breakeven as early as late 2025, driven by the core Nebius cloud business scaling up (cdn.prod.website-files.com).
• Financial Outlook: Nebius’s management provided bullish guidance for the near future. The company expects revenue to grow 3–4× in 2025, reaching $500–$700 million for the year (cdn.prod.website-files.com) (reuters.com).
• This is an ambitious jump (on the order of half a billion dollars in annual sales) which underscores the demand they foresee in the AI market and the capacity coming online from their expansions.
Achieving the midpoint of that range (~$600M) would imply another ~4× year-over-year growth from 2024 to 2025.
While aggressive, this guidance aligns with the broader trend in the AI infrastructure space – many AI model-training companies are scaling compute spend exponentially.
Nebius believes it can capture a significant slice of that growth. Longer-term, Nebius and analysts have pointed to a multi-billion dollar revenue opportunity by the end of the decade if the company continues executing (the presentation hints at multi-billion annual revenues by 2030 if AI demand continues its trajectory).
For now, the focus is on 2025 as a breakout year, with the company emphasizing that hitting ~$500M+ revenue should coincide with approaching EBITDA breakeven (cdn.prod.website-files.com), setting the stage for potential profitability thereafter.
• Investment and Capital Expenditure Plans: To support its growth, Nebius is undertaking major capital investments.
An investor presentation and subsequent communications highlight plans to invest over $1 billion by mid-2025 to build out its AI infrastructure (reuters.com).
• This capex is going into expanding the Finland data center (tripling capacity), deploying the new clusters in Paris and other European locations, and launching the U.S. data center – as described in the infrastructure section.
Despite the hefty spend, Nebius’s balance sheet is strong enough to cover it: the company was spun off with a cash position of over $2 billion (as of late 2024) (xm.com), giving it a substantial war chest.
Furthermore, Nebius has attracted additional external funding – in December 2024, it announced a $700 million private placement led by top-tier investors including Accel and NVIDIA (investors.com).
This fundraising was oversubscribed, indicating high investor interest in Nebius’s story.
The fresh capital injection will bolster Nebius’s ability to execute its expansion (e.g., purchase thousands more GPUs, build facilities, and possibly enter new markets).
It’s worth noting that NVIDIA’s participation in the funding not only adds cash but also deepens the strategic relationship (a positive signal for Nebius’s access to cutting-edge hardware supply).
• Financial Position and Use of Funds: Nebius emerges from its separation (from Yandex) well-capitalized. With ~$2B cash on hand pre-fundraise and now an added $700M, the company has around $2.7B in capital to deploy.
Management has laid out that this will primarily fund infrastructure growth and R&D – essentially “fuel in the tank” to grab market share in the next couple of years.
The goal of these investments is to scale fast enough to reach that half-billion revenue range by 2025. Nebius is intentionally front-loading capex (building capacity ahead of demand) to be ready for large enterprise deals and surging AI workloads.
This strategy, while pressuring short-term financials, is aimed at establishing Nebius as a leading AI cloud platform in Europe and beyond before too many competitors catch up.
Assuming the growth materializes, Nebius’s finances could rapidly improve after 2025 as high-margin cloud revenue flows into infrastructure that is largely already built.
Strategic Takeaways: From a financial perspective, Nebius is executing the classic playbook of a high-growth infrastructure startup: spend aggressively now to capture a nascent market, with the conviction that this will yield a dominant position and strong financial returns later.
The investor presentation’s projections show confidence in a massive scaling of revenue over the next 1-2 years, which if achieved, would likely make Nebius one of the top pure-play AI infrastructure providers globally.
Investors are drawn to the combination of hyper growth (5–6× annual growth rates) and the clear line of sight to profitability once scale is reached.
However, hitting the 2025 targets will be crucial – it will validate Nebius’s competitive edge and justify the heavy investments.
With substantial cash reserves, Nebius has the luxury of being able to invest for growth without immediate fear of running out of funds, which is a competitive advantage in itself (many smaller rivals may not have this level of funding).
The company’s financial plan reflects an “expand rapidly now, monetize fully later” philosophy, fitting for the current AI gold rush.
If Nebius can continue its current trajectory, it stands to transition from a cash-burning startup into a self-sustaining, highly valuable enterprise at the center of the AI infrastructure ecosystem.
Conclusion
The Nebius Group investor presentation paints a picture of a company at the forefront of the AI infrastructure wave.
Nebius’s business model is built on owning the entire stack – from efficient data centers to specialized cloud software – to deliver AI computing power as a service.
Its AI infrastructure is already significant (with thousands of GPUs online and more being added across Europe and the U.S.), providing a foundation for handling the most demanding AI workloads.
Nebius knows its target customers well: AI-centric startups, scale-ups, and forward-looking enterprises that are hungry for compute and value a cost-effective, flexible alternative to the big clouds. In terms of competitive positioning, Nebius leverages focus and technical excellence to stand out, and enjoys strong backing (including from NVIDIA) to support its ambitions.
The financials reveal a company investing heavily to seize a generational opportunity – rapid revenue growth is being prioritized, with profitability on the horizon once a critical mass is achieved.
Strategically, Nebius is attempting to become a leader in AI infrastructure, particularly in the European market, to complement or challenge the U.S. hyperscalers.
The investor presentation’s contents highlight Nebius’s key strengths: a cutting-edge platform, impressive growth metrics, clear market demand, and ample funding.
Of course, the execution risks are real – scaling an infrastructure business is capital intensive and competitive. But if Nebius delivers on its roadmap, it could very well emerge as a major “AI cloud” provider in the coming years, riding the AI revolution much like early cloud companies rode the internet boom.
The presentation’s insights reinforce that Nebius’s management is keenly aware of both the opportunity and the challenge ahead, focusing on core principles (efficiency, specialization, and aggressive growth) to guide the company’s journey.