Nebius Group (NBIS): Q3 2025 Earnings Review
Today, Nebius Group (NBIS) released its Q3 2025 results.
It remains my largest position, with an average cost of $25.67/share.
In this article, I’ll break down everything you need to know about the Earnings Report.
Financial Highlights
Revenue of $146.1M vs. $155.7M est. (+355% and +39% QoQ)
This miss had nothing to do with demand, it was purely a matter of timing in capacity deployment. New capacity comes online in waves, so short-term fluctuations are inevitable. Overall, the underlying demand picture has never looked stronger.
“Every time we bring capacity online, we sell it.”
September ARR of $551M (+28.1% QoQ)
That’s actually consistent with my expectations. The heavy lifting is happening in Q4, with deployments in the UK and Israel, expansion in Finland, and the start of Phase 1 for the New Jersey site (tied to the Microsoft deal). This means ARR growth should accelerate sharply in Q4 and early 2026 as those new sites go live.
Gross margins of 71% vs. 71% QoQ
I was hoping to see a gradual uptick as the company scales, but stability at this level is still positive, especially when you consider CoreWeave’s margins (similar) at almost 10x Nebius’ scale. Over time, I expect operating leverage to push that figure higher.
Adj. EBITDA of $(5.2M) vs. $(5.8M) est.
EPS of $(0.40) vs. $(0.52) est.
More importantly, Adj. EBITDA margin for the core business reached 19%, improving a lot faster than I anticipated.
Raised CapEx guidance from ~$2B to ~$5B
On the surface, that might look alarming, but it’s actually a bullish signal. The company is simply accelerating capacity expansion to meet overwhelming demand. For context, Nebius spent about $2B in CapEx during the first three quarters of 2025, meaning it plans to spend another $3B in Q4 alone. That level of investment should translate into a meaningful boost in operational capacity, so I wouldn’t be surprised if we get some announcements before year-end.
Even more important: MASSIVE guidance.
Year-end ARR guidance reiterated at $900M-$1.1B
· 2025 revenue guidance narrowed from $450-630M to $500-550M
Year-end 2026 ARR guidance of $7-9B
“More than half of our $7-9B ARR target by the end of 2026 is already booked.”
Year-end connected power guidance reiterated at 220MW
Year-end 2026 connected power guidance of 800MW-1GW
Year-end 2026 contracted power guidance raised from >1GW to >2.5GW
For context, back in October of last year, the company was guiding for just 240+ MW in the midterm, meaning 2027 or 2028.
I expected them to provide an ARR target for 2026, especially after the Microsoft deal gave more visibility into next year’s numbers, but I never thought it would be this high. And that jump in capacity estimates is just insane. I’ll definitely need to update my Valuation Model, so stay tuned.
That said, I’d love to hear more details on how they plan to reach these goals, since we still don’t have specifics about the new U.S. locations announced last quarter. What we do know is that the next phase of the UK expansion is scheduled for Q1 2026, Israel is also expected to scale, and Nebius is in the process of securing several large sites across the U.S. and Europe, some with hundreds of MWs of capacity. There’s also an option to add hundreds of additional MWs at the New Jersey site in the future. But honestly, I feel like management could have given us a bit more color on the new locations yet to be announced.
Another Hyperscaler Joins the List
Another great piece of news: after signing a $17.4–19.4B deal with Microsoft, which management said is “fully on track” with execution, Nebius announced a new agreement with Meta valued at roughly $3B over five years.
This marks Nebius’ second hyperscaler contract in just two months.
“Over the next three months, we plan to deploy the capacity needed to service the agreement. In fact, demand for this capacity was overwhelming, and the size of the contract was limited to the amount of capacity that we had available.”
These deals are obviously fantastic for long-term visibility, but they also come with a side effect: Nebius is running completely sold out on capacity. That means it’s currently having to turn away some potential enterprise/AI start-up customers.
“I’ve learned a new skill, one I don’t think many go-to-market professionals have ever had to experience, and that’s learning to say no to customers as we routinely sell out and have to actually let them down lightly and try to convince them to purchase in the future.”
— Marc Boroditsky, Chief Revenue Officer
Still, this isn’t a bad problem to have. Full utilization rates are critical for an aggressive scale-up phase like this, especially when it comes to securing cheaper capital for future growth.
Essentially, both the Microsoft and Meta partnerships act as powerful commercial and financial accelerators for Nebius’ core AI cloud business.
Funding Needs
The company ended Q3 with $4.9B in cash, yet as mentioned earlier, CapEx in Q4 alone is expected to reach about $3B. Given the updated capacity targets and expansion plans, it’s clear Nebius will need additional funding, and management was upfront about that.
The company will rely on three financing sources: corporate debt, asset-backed financing, and equity. The company expects to secure asset-backed debt on attractive terms, supported by the strong credit profiles of its largest customers, while maintaining a disciplined capital structure.
On the equity front, Nebius plans to establish an at-the-market (ATM) equity program of up to 25MClass A shares, with a prospectus supplement expected to be filed on November 12, 2025. This program will give the company flexibility to raise capital efficiently and incrementally, depending on market conditions and capital needs. Dilution won’t happen right away.
“We will remain dilution-sensitive as we seek to finance future growth opportunities.”
Now, I know many investors will immediately think: “Ugh, dilution!” But in Nebius’ case, I’m not concerned. In fact, this was more than expected. If the company maintains healthy capital efficiency, dilution will ultimately enhance shareholder value over time by fueling capacity growth that compounds revenue and profits faster than the share count expands.
Time will tell, but I don’t think we have reasons to be concerned.
Other Highlights
Product Update
Nebius continues to evolve beyond bare-metal infrastructure, building a comprehensive AI cloud stack designed for both enterprises and AI-native startups. In Q3, the company made substantial progress across its hardware, software, and platform layers, reinforcing its competitive advantages and deepening its technology lead.
Nebius AI Cloud 3.0 “Aether”
Launched AI Cloud 3.0 “Aether”, the company’s most significant platform upgrade to date. This enterprise-grade release delivers new capabilities in security, compliance, governance, and developer productivity, making it ideal for highly regulated industries.
Aether introduces major certifications, SOC 2 Type II, HIPAA, and ISO 27001, validated by independent auditors, and aligns with NIS2, DORA, ISO 27032, 27701, and 27799 security frameworks. The update adds an embedded secrets manager for API key and credential protection, advanced identity and access management (IAM) features, granular policy controls, and a unified observability suite that consolidates metrics, logs, and audits across all workloads.
Nebius also rolled out a redesigned developer interface with preloaded AI/ML applications for faster workload deployment, simplified GPU and CPU quota management, and expanded integrations with SkyPilot Server and Anyscale to streamline distributed AI training and inference.
Performance-wise, the company continues to achieve industry-leading scores.
Nebius Token Factory
A major milestone this quarter was the launch of Nebius Token Factory, the company’s Inference-as-a-Service platform designed for the next phase of AI adoption.
As Roman Chernin explained, the first wave of AI growth was driven by foundational model builders. The next wave, however, is being led by enterprises and independent software vendors (ISVs) applying AI to real-world products and workflows. Token Factory was built to serve this expanding market.
The platform unifies fine-tuning, model optimization, and large-scale inference under a single governed environment. It enables organizations to transform open-source models, like Llama, Qwen, DeepSeek, and Nematron, into optimized, production-ready systems with guaranteed performance, sub-second latency, and 99.9% uptime.
Token Factory gives enterprises and vertical AI builders the tools to industrialize the post-training lifecycle, taking open-source checkpoints and turning them into production-grade models with transparent cost per token. Built-in fine-tuning and distillation pipelines help customers adapt large models to proprietary data, cutting inference costs dramatically (one major client reported a 26x reduction).
“We believe that all companies will invest in inference to productize AI.”
With Token Factory, Nebius is positioning itself to capture the coming boom in enterprise inference workloads, which are expected to consume even more compute than training over the long term.
Doubling Down on AI Startups and Enterprise Focus
Nebius continues to reiterate its focus on AI startups and enterprise customers, aiming to be the long-term backbone of AI-driven businesses. Management emphasized that its goal is not just to expand capacity, but to build a sustainable, defensible model defined by recurring revenue, software-driven margins, and deep technological differentiation.
“We have a large pipeline of new software and services that we are continuing to build, which will differentiate us from other cloud companies.”
This approach is already paying off:
Deepened relationships with scaled players like Shopify and World Labs.
Onboarded emerging innovators such as Cursor and Black Forest Labs.
Expanded within healthcare and life sciences, where partners like Basecamp Research, Prima Mente, and Sword Health are training models to address major medical challenges.
Began targeting new sectors, including physical AI, financial services, media, and retail, where enterprise-grade compliance and reliability are becoming prerequisites.
“We want to build partnerships with these customers and help them to meet their capacity requirements in the future, especially with enterprises, because they do not want to actually have a multitude of vendors. They prefer to align with a strategic partner.”
“As busy as we are with these mega deals, our main focus is still to build our own core AI cloud business. The economics and the cash flow of mega deals are attractive in their own right, but they also enable us to build our core AI cloud business faster. This is our real future opportunity.”
Ecosystem Expansion
Nebius expanded its partner ecosystem with leading AI infrastructure and software providers:
TD Synnex named Nebius its preferred AI Infrastructure-as-a-Service partner in North America, opening access to thousands of resellers and enterprise buyers.
Baseten, Anyscale, and SkyPilot launched new integrations for scalable AI training and inference.
Qdrant, dstack, and Union.ai joined the ecosystem to extend Nebius’ reach into retrieval-augmented generation (RAG), workflow orchestration, and GPU marketplaces.
Partnership with Accenture, aimed at helping enterprises and public-sector organizations deploy AI faster and more efficiently.
These partnerships significantly extend Nebius’ market reach and accelerate time-to-value for customers, reinforcing its position as a vertically integrated, sovereign, enterprise-grade AI cloud. While their importance might seem limited for now, given Nebius’ consistently “sold-out” status, these partnerships are crucial distribution channels that will help driving adoption as the enterprise wave of customers continues to accelerate.
The UK Opportunity
One of the most exciting developments this quarter is Nebius’ early progress in the UK.
The company officially launched its UK data center last week, with capacity set to come online within days. What’s even more impressive is that since first announcing the project back in June, Nebius has nearly doubled the amount of capacity it plans to bring online, a direct response to extremely strong demand.
In fact, before the facility even goes live, most of that capacity is already sold out. That’s becoming a recurring theme with Nebius: every time new capacity is added, it’s immediately absorbed by customers eager for AI compute.
Management emphasized that the UK is one of the most vibrant AI markets in the world, rivaled only by the U.S. and China. The country is home to a thriving ecosystem of AI startups, backed by a robust venture capital environment and growing support from the government. Major tech companies are also expanding their regional R&D presence there, a strong sign of structural, long-term demand for compute infrastructure.
Nebius plans to reach peak capacity at the current UK site by January, but that’s only the beginning. The company sees significant room for further expansion across the country as AI adoption continues to accelerate.
The Israel Opportunity
Alongside the UK launch, Nebius also celebrated the successful start of its Israel data center, which is already fully operational. According to management, the entire initial capacity was effectively pre-sold before going live.
The company sees Israel as a strategic and fast-growing market for AI infrastructure. The country’s strong tech ecosystem, deep engineering talent, and growing AI startup scene are driving intense demand for compute resources. Nebius believes there’s significant room to expand capacity over time, as local AI innovation continues to accelerate.
Interestingly, the Israeli government is also playing an active role in stimulating AI development, offering subsidies to startups and research institutions to help them access compute resources. This type of public support not only boosts near-term demand but could also create a more sustainable AI ecosystem in the region.
Management even suggested that Israel’s model could become a template for other countries looking to foster domestic AI industries, combining private-sector investment with targeted government incentives to build local compute capacity.
Will Nebius Pursue M&A Deals?
When asked whether Nebius might partner with or acquire companies that already have secured power or land, or potentially consolidate smaller neoclouds, Arkady made it clear that Nebius’ strategy remains centered on margin discipline.
Every decision the company makes, from contract negotiations and capital raises to data center design and hardware manufacturing, revolves around maximizing efficiency and protecting margins. Nebius has explored potential acquisitions in the power and land markets, but so far, its organic approach has proven to deliver stronger margins than what it could likely achieve through M&A.
Instead of buying assets or competitors, Nebius is doubling down on building and owning its own infrastructure. The company continues to decrease reliance on colocated and rented facilities, shifting toward fully owned data centers, a move that should further enhance profitability over time.
Arkady believes margins will become a key differentiator in the AI infrastructure space. While others may focus on aggressive expansion or acquisitions, Nebius is betting that long-term success will come from owning the full stack: power, land, hardware, and software, and running it as efficiently as possible.
Challenges to Scale
According to Andrey Korolenko, the company’s Chief Infrastructure and Product Officer, the biggest near-term obstacles are securing power and navigating the supply chain.
“In the near term, the key challenges to increasing capacity are securing power and managing the supply chain. We are addressing these issues. We’ve handled similar situations in the past and have quite a bit of expertise in building and operating data centers.”
In theory, Nebius could build as much capacity as possible to meet demand. In practice, though, there are physical and logistical limits, from permitting and equipment lead times to the sheer availability of power.
Arkady Volozh added more color on how the company plans and executes its CapEx, explaining that Nebius approaches data center development in three distinct stages:
Securing land and power: This is the cheapest stage, accounting for roughly 1% of total CapEx, and ensures that Nebius locks in long-term access to the most constrained resource, electricity.
Building the data centers: This phase includes construction, electrical and cooling systems, batteries, and other infrastructure. It represents about 18–20% of total CapEx.
Deploying GPUs: The final stage, and by far the most expensive, represents roughly 80% of total CapEx. Nebius only commits this spend when demand is either contracted or clearly visible.
This disciplined, phased approach allows Nebius to scale responsibly while preserving financial flexibility. The company first secures as much power as possible, then builds capacity as capital allows, and finally installs GPUs only when utilization is guaranteed.
AI Bubble
When asked whether we might be in an AI bubble, Arkady Volozh sighed and said: “Ask this question these days…” That reply drew some laughs, but also set the stage for a thoughtful answer.
He made it clear that Nebius sees the current AI cycle not as a bubble, but as part of a once-in-a-generation technological revolution.
“What we see today is that the demand is here. There’s no doubt that much more compute will be needed, and much more will be built.”
But he also acknowledged that today’s imbalance between demand and supply is temporary and will eventually stabilize. Rather than purely chasing the short-term hype around bare metal deals, Nebius is focused on building an AI cloud designed for real businesses and real industries, the kind of enterprise customers that will drive durable, long-term value creation.
“Ultimately, we just need to ensure we’re diversified in terms of customers and workloads, and that’s exactly what our software enables.”
Even with growth accelerating, Arkady emphasized that Nebius remains laser-focused on maintaining healthy margins and building a sustainable business model.
The “Old Generations” Debate
There’s been some debate recently about whether Nebius could face issues with older GPU generations as newer chips enter the market. At least for now, that concern is largely overblown.
“We sold out our capacity in Q3 across all types, and we’re nearly sold out for Q4 as well. We continue to see extremely positive demand for Hoppers.”
In other words, demand remains strong across the entire GPU portfolio, not just the latest and greatest models. Customers are using different GPU generations for different types of workloads, and that diversification in demand actually helps Nebius maintain consistently high utilization rates.
There’s a clear hierarchy of compute needs:
Training workloads, such as large-scale model development, require the most advanced chips (like Blackwell).
Inference workloads, such as text generation, recommendation systems, and many production use cases, can run extremely efficiently on slightly older GPUs like Hopper.
High-complexity inference, such as video generation, will naturally shift toward higher-performance chips as those workloads scale.
That’s why the idea of “GPU obsolescence” tends to be exaggerated. What really matters is the performance per dollar and how efficiently each GPU generation is matched to the right workload. As a result, each generation eventually finds its optimal role within the compute hierarchy, rather than becoming irrelevant. For context, there’s still strong demand for A100 GPUs, which were released five years ago (Nebius doesn’t own any A100s, but several companies have confirmed that demand remains high).
Of course, demand for NVIDIA’s Blackwell GPUs is ramping fast, especially for customers focused on training foundation models. Management noted that the arrival of these new chips has actually encouraged more customers to secure capacity in advance, often through longer-term contracts, giving Nebius better visibility and revenue stability.
“With the new generation of NVIDIA Blackwells coming online, more customers are interested in purchasing capacity in advance and securing it for a longer period of time.”
Subsidiaries
Avride
In October, Avride announced a strategic partnership with Uber, under which Uber and Nebius will make strategic investment and other commitments of up to $375M in Avride.
This deal enables Avride to rapidly scale operations and capitalize on the enormous opportunity in autonomous driving technologies and services. It also validates the company’s technical leadership in autonomous navigation and perception systems.
The investment builds on Avride’s existing commercial relationship with Uber, established through a multi-year strategic agreement signed in 2024. Nebius intends to retain a significant ownership stake, ensuring continued alignment and meaningful long-term upside for shareholders.
Operationally, Avride reached several key milestones:
Completed pre-launch testing of its autonomous vehicles in Dallas and fully integrated with Uber’s ride-hailing platform. Dallas is expected to become the first city to host Avride and Uber’s public robotaxi service in December 2025.
Achieved strong momentum with delivery robots, surpassing 215,000 deliveries YTD. Daily deliveries exceeded 1,900 in October as operations expanded in Jersey City, Dallas, Austin, and across major university campuses.
Signed partnerships with Shake Shack, Wendy’s, and White Castle, while also launching operations at the University of Arizona with a fleet of 60+ delivery robots.
Announced the launch of robot deliveries on the Grubhub marketplace, expanding its partnership with the company beyond campus delivery for the first time. It represents Grubhub’s first use of autonomous delivery on its main platform.
In Japan, Avride’s partnership with Mitsui Fudosan to provide autonomous retail logistics deliveries at the country’s largest outlet mall has been successful and continues to expand.
TripleTen
TripleTen continued to deliver exceptional revenue growth of ~100% YoY, fueled by strong new student enrollment and rising average fees. The platform added approximately 6,000 new students during the quarter, reflecting sustained demand across its core markets in the U.S. and Latin America. Customer acquisition efficiency remained strong, supported by high referral rates and sales team productivity.
During the quarter, TripleTen:
Launched AI Analyst, a new program designed to help students reskill for the AI-driven economy.
Introduced Nebius Academy, a B2B offering that enables organizations to provide training courses for employees.
Continued to integrate AI across its operations, leveraging automation for resume building, coaching, and student support.
Began using generative AI to produce course materials and videos, significantly reducing content-creation costs compared with traditional methods.
Toloka
Toloka remains a key asset in NBIS’ portfolio, serving as a leading data provider for LLM and GenAI developers. Although Nebius no longer holds voting control, it retains a significant majority economic interest, preserving substantial value participation.
In Q3, Toloka completed the core technology for its hybrid AI agent, which combines AI with human expertise to execute complex, long-horizon tasks. This system integrates reinforcement learning “gyms” and large-scale data production, allowing Toloka to serve a wider customer base, from major AI labs and global enterprises to researchers and individual developers.
The company also plans to launch Tendem.ai, a hybrid human-AI agent platform that blends automation with expert human oversight to deliver accurate and reliable results across various applications. For many workflows, Tendem.ai handles most of the process, leaving humans to complete the “last mile.”
ClickHouse
Nebius continues to hold a minority equity stake in ClickHouse, a global leader in real-time analytics, data warehousing, observability, and AI/ML infrastructure. The company continues to scale rapidly as enterprises seek faster, more cost-efficient analytical solutions. Its technology remains central to the modern data infrastructure stack and complements Nebius’ mission to enable AI at scale.
As you know, ClickHouse is my favorite non-core asset in Nebius’ portfolio, even though it’s the one where the company holds a smaller stake. With major customers like Anthropic, OpenAI, Tesla, Meta, Cloudflare, and other AI leaders, it’s no surprise that ClickHouse has quadrupled its ARR over the past 12 months.
The CEO recently mentioned that the goal is to take ClickHouse public one day, and I see that as a major catalyst for Nebius. ClickHouse is one of the hottest AI startups right now and could reach extremely high valuations in the midterm, giving Nebius the opportunity to unlock billions in non-dilutive capital.
I’ll continue to keep you updated on everything I learn about the company, feel free to follow me on X (@mvcinvesting) for more.