Nebius (NBIS.US) Q2 Conference Call: Holding an order of 40 billion US dollars and sprinting to 5GW power supply and demand, AI computing power supply and demand into the “seller's pricing power” era

The Zhitong Finance App learned that in recent years, European technology leader Nebius Group NV (NBIS.US), which has switched to providing cloud-based AI computing resources, announced strong performance and future prospects on Wednesday. Together with the financial data from CoreWeave, another Neocloud (Neocloud) leader the day before, it shows that demand for AI computing power infrastructure continues to be strong. According to Nebius' results, the AI cloud computing business, which accounts for the vast majority of its business, reached $575 million in revenue during the quarter. Total revenue for the quarter surpassed US$582 million, up 454% year over year. This figure was higher than analysts' expectations of US$557 million.

The growing demand for AI computing power is also driving Nebius to further accelerate infrastructure expansion. The company raised its contracted power consumption target from over 4 gigawatts (GW) to 5 gigawatts (GW) by the end of 2026, and plans to add more than 1 gigawatt of capacity each year starting 2027. Looking ahead to 2026, the company expects annual capital expenditure to reach 20 billion to 25 billion US dollars, an adjusted EBITDA profit margin of about 40%, and plans to announce official 2027 performance guidelines later this year; it expects annual revenue of 3 billion to 3.4 billion US dollars, and maintains an annualized operating revenue target of 7 billion to 9 billion US dollars. In addition, it is expected to receive more than $9 billion in advance payments from customers for the full year of 2026, and the total cumulative customer commitment has exceeded 40 billion US dollars.

Nebius was spun off from Russian internet service giant Yandex in 2024 and is one of a group of cloud computing technology companies known as “new cloud computing service providers” (Neocloud). These companies rent out artificial intelligence computing capabilities. The Amsterdam-based company has reached a long-term computing power infrastructure cooperation agreement with Microsoft and Meta Platforms. Nebius and CoreWeave, an AI cloud computing power leader headquartered in the US, are both part of the “AI-specific cloud computing vendor” Neocloud circuit, focusing on faster delivery, optimized training/inference stacks and flexible contracts to serve the huge AI workloads of large-scale model developers and enterprises.

The strongest signal given by Nebius management is not simply an increase in revenue, but rather that the supply and demand of AI computing power and the transaction pricing power of both parties continue to lean towards computing power sellers such as Nebius. The company's CEO, Arkady Volozh, made it clear that the company signed 4 landmark AI cloud contracts this quarter, with an average value of more than 1 billion US dollars per contract, revenue per megawatt of 20 million to 25 million US dollars, and customer advance payments are sufficient to cover 50% to 60% of related capital expenses.

More importantly, the company said that if it wanted, it could sell out all of its 2027 planned capacity today according to these terms, but Nebius took the initiative to reserve part of the capacity because management believes the price may be higher closer to delivery. The company said that short-term, 3-6-month emergency computing power contracts can now reach 40 million to 50 million US dollars per megawatt, or even higher; the transaction price of the Nvidia Blackwell AI GPU computing power cluster at the company's first capacity auction was 15% higher than the highest price in history, indicating that scarce AI computing power is gaining more and more immediate price discovery capabilities.

Nebius management believes that AI workloads are rapidly moving from training to real production environments. Among them, AI agent mode programming is only the most obvious example. Long-running AI agent workflows have entered the fields of financial services, e-commerce, medical care, and marketing automation. Companies are increasingly concerned not only about whether models can complete tasks, but whether the cost per token (Cost per Token) is low enough to be deployed on a large scale. This is driving open weighting models, post-training, reinforcement learning, inference, synthetic data generation, and information anchoring while expanding computing power requirements.

Nebius' Token Factory (AI Token Factory) has achieved first-day support for several cutting-edge open models; the Tavily developer community expanded from 1 million in February to over 2.5 million people. The company's management's judgment is very clear: inference is not an edge workload after training, but is becoming a source of huge computing power demand throughout model development, post-training, and production deployment.

The following is the full text of the Nebius performance conference call (artificial intelligence tool assisted translation):

Operator:

Welcome to Nebius Group's Q2 2026 results conference call. [Operator Tip] I will now hand over the phone to Gili Naftalovich, Head of Investor Relations, and she will begin this conference call.

Gili Naftalovich, Head of Investor Relations:

Hello everyone, and welcome to Nebius's Q2 2026 results conference call. Joining us on the conference call today are our CEO Arkady, CFO Dado, and Nebius's broader management team.

Before we officially begin, let me quickly explain the Safe Harbor Clause. Some of our statements regarding our business operations and financial performance today may be considered forward-looking statements. These statements are based on current expectations and assumptions, which are subject to several risks and uncertainties. Actual results may vary materially. Please refer to our Form 20-F for a list of risk factors. We are under no obligation to update any forward-looking statements.

In this conference call, we'll be covering both GAAP and non-GAAP financial measures. A reconciliation table between non-GAAP measures and GAAP measures is included in the results press release issued today. All documents relating to the results have been published and made available to the public through our investor relations website, which can be accessed through nebius.com.

Now I want to hand over the phone to Arkady.

Arkady Volozh, founder, CEO and non-independent executive director of Nebius:

Thank you, Gili, and welcome everyone to today's conference call. I want to give our investors a deeper understanding of our business model.

We had an amazing quarter. The market demand for the products and capabilities we're building is still huge, and we have the right business model to capture those demands. We build capacity ahead of time before signing contracts and use our multi-tenant cloud and software stack to support customers, including AI-native enterprises, leaders in intelligent AI, new AI laboratories, and some of the world's most advanced enterprise customers.

Our strategy is working. We can choose when to sell capacity, to whom, under what terms, and how to finance all of these projects. This flexibility allows us to meet the needs of independent AI developers and support an open, diverse, and competitive marketplace. We meet customer needs through three types of transactions. Each type of transaction has a different duration, pricing, and role in our business.

The first category is our core AI cloud business. We sign medium-term contracts of 1 to 3 years with the world's most ambitious AI companies. In this quarter alone, we completed 4 landmark deals for clients including Reflection, Cohere, a new US laboratory that has reached a significant scale and a major quantitative trading company headquartered in the US. The average value of these deals is over $1 billion. They correspond to revenue levels of 20 million to 25 million US dollars per megawatt, and customer advance payments can cover 50% and 60% of related capital expenses. But most importantly, if we want, we can sell the full 2027 capacity today under these terms. But we didn't do that. We believe more value can be obtained by reserving part of the capacity to meet customer needs for a shorter period of time and timeliness.

This led to a second type of transaction, which is a capacity with a shorter term, usually up to 6 months, to serve customers with immediate, clear time boundaries, and high value requirements. For this, they are willing to pay a significant premium. Following this model, the deal we are currently negotiating is priced at between $40 million and $50 million per megawatt, sometimes even higher. BTW, we just signed one of these deals recently.

The volumes of the two types of deals I've just introduced will go live later this year. As a result, they won't have a significant impact on our 2026 revenue guidance. They will definitely impact our revenue in 2027 and beyond, though.

The third type of transaction, as we've mentioned many times before, is a long-term contract we sign with an investment-grade client. These types of contracts serve an important function: they help us finance expansion in a faster and more efficient way. The secured debt financing we raised in July was completed based on one of these deals. Since we have a $40 billion contract backlog, we will do more of this kind of financing in the future.

We're always building, innovating, and developing our software stack, technology products, and services. But today, I'd like to focus on another type of innovation this quarter. We launched our first capacity auction. The auction was very successful, and the final transaction price was the highest price we've seen for Blackwell's generation chips, 15% higher than the highest price we've ever charged. This allows us to get a very strong signal in real time about the true value of these capacities in the market. In a sense, this is an innovation in our market entry strategy.

In terms of capacity, we are also innovating construction methods. For example, the asset-light partnership model we launched this quarter provides us with another way to expand. This model targets two constraints in our industry — capital and capacity. Partners are responsible for financing, construction, and operating facilities, while Nebius provides a complete full-stack platform and requirements. We provide value-added services on top of our partner infrastructure, which brings us high-margin revenue while requiring very little balance sheet capital. This model has the potential to unlock new capacity for us from 2027 and beyond.

We continue to build future capacity pipelines through our own and managed sites, and today we are raising our contracted electricity target to 5 gigawatts by the end of the year. Our future capacity pipeline actually makes Nebius one of the few companies in the world that can build more than 1 gigawatt of additional capacity each year. Also, we plan to do that in 2027.

Finally, everything we planned to do this quarter has been completed. And in the vast majority of cases, we get more done. We have signed landmark deals in our core markets, and terms are better than we expected. Capacity is growing to meet demand, and our platform is meeting industry demands. We're reporting more than just a strong quarter that just passed. It's far more than that. We can see demand; we can see supply. We look at the terms of the deal for 2027 and beyond. With our future capacity plans, including 2027, we couldn't be more excited about the future.

With this sense of optimism, I handed over the phone to Dado.

Chief Financial Officer Dado Alonso:

Thank you, Arkady. As we successfully executed our strategy in the first half of the year, we entered the second half of this year with strong momentum. In the second quarter, we achieved another three-digit increase in revenue and ARR, and this happened even before most of our 2026 capacity went live in the second half of the year. At the same time, our adjusted EBITDA margin expanded significantly, and our financing position was further strengthened, thanks to a large number of customer advance payment terms and the expansion of capital sources. This is a strong start and keeps us on track to meet our strategic and financial goals throughout the year.

I'll cover our progress in the second quarter, share specific details of our financial results, and finish with the full year guidance we've reaffirmed today. Please note that all comparisons are year-over-year unless otherwise stated.

First, let's look at revenue and ARR. In the second quarter, our group revenue increased 454% to $582 million, up 46% from the previous quarter. Nebius's AI business grew even faster, growing 514% from last year to US$575 million, accounting for 98% of the Group's revenue. By the end of June, the annualized revenue operating rate reached $3 billion, up 598% year over year, and 56% (original text) [58%] from the end of March of $1.9 billion. Revenue growth was mainly driven by additional capacity in the first quarter, higher utilization rates, high-margin revenue from our new asset-light business model, Token Factory, and recent acquisitions. Increased utilization is due to improvements in the efficiency of the underlying infrastructure. Also, we've sold out all of our capacity again because no matter how quickly we get the capacity online, we can sell it.

Next, let's talk about profitability. The Group's adjusted EBITDA was US$236 million, compared with a loss of US$21 million in the same period last year and US$129.5 million in the previous quarter. The Group's adjusted EBITDA margin was 41%, up from 32% in the first quarter. The Nebius AI business generated an adjusted EBITDA of $286 million (original text) [$236 million], with a profit margin of 50%. The difference in profit margins between the Group and Nebius AI business reflects our investments in Avride and TripleTen, both of which are still in their early stages. We anticipate that Nebius AI business will continue to contribute the majority of the Group's adjusted EBITDA.

The increase in profitability stemmed from revenue growth and the beginning of contributions from the asset-light model, Token Factory, and recent acquisitions. Even as we continue to invest in growth, we can clearly see a path to further increase our profit margins, not only this year, but also 2027 and beyond. We can see future price trends and expect that additional capacity from our own data centers will begin to improve profit margins in the second half of next year.

Now let's look at the balance sheet. Customer prepayments reached an all-time high this quarter. About 70% of the deals we signed in the second quarter included upfront payments. Overall, this year's customer advances will bring in more than $9 billion in upfront capital, directly reducing the capital we need to raise through debt and equity. Operating cash flow reached US$2.3 billion in the second quarter, and cash and cash equivalents at the end of the period was US$8 billion.

In the second quarter, we also used a plan to issue shares at market prices. We issued 12.7 million Class A shares at a weighted average price of US$224 per share, receiving a total of approximately US$2.8 billion raised. As of June 30, there are still 12.3 million shares available for issuance under the plan. We see ATMs as a flexible financing tool; it's just an option in our toolbox, not a promise.

In July, we announced the completion of our first asset-backed debt financing of $775 million. This financing is secured by contract cash flow, and its pricing is only moderately higher than the benchmark interest rate. At today's level, the interest rate is equivalent to a mid-single digit percentage. At the same time, we also have over $40 billion in additional client commitments with similar terms based on strategic deals we've signed with investment-grade companies.

We will continue to diversify our funding sources while maintaining a balanced mix of debt and equity. We are advancing more asset-backed financing while continuing to evaluate company-level debt and other financing instruments.

Next, let's talk about capital expenditure. Capital expenditure for the second quarter was approximately US$5.7 billion, mainly to purchase GPUs, GPU-related hardware, and expand data centers. The capital investments completed so far have enabled us to meet this year's capacity plans.

Let's talk about the outlook. The solid progress made in the first half of the year further strengthened our confidence in the 2026 outlook. Therefore, today we are reaffirming our guidance for all indicators for the full year 2026. We continue to expect an annualized revenue operating rate of 7 billion to 9 billion US dollars, group revenue between 3 billion and 3.4 billion US dollars, the group's adjusted EBITDA profit margin of about 40%, and capital expenditure of 20 billion to 25 billion US dollars. We remain confident that capacity deployment will be accelerated in the second half of the year, and we expect that capacity deployed at the end of the second quarter will begin contributing to revenue in the third quarter.

As Arkady said, we are now building capacity to meet 2027 needs, and these constructions are supported by existing customer commitments, so we can clearly see the revenue corresponding to these investments in the short term.

Finally, we are actively seeking opportunities to generate additional high-margin revenue through new business models, capacity auctions, short-term volume transactions, and asset-light businesses. All in all, our strong performance in the second quarter reflected our ability to achieve large-scale profitable growth in our business through rigorous execution. We have achieved strong revenue growth, increased profitability, pioneered a more capital-efficient capacity expansion path, and expanded financing options.

Looking ahead, we will continue to expand rapidly to seize immediate opportunities while remaining balanced, disciplined, and focused on creating long-term value for our shareholders.

Next, I handed over the phone to Gili and went to the question and answer session.

Q&A session (AI-assisted summary and translation)

Operator: [Operator Reminder] I will now return the phone back to the Nebius management team for a question and answer session.

Gili Naftalovich:

Thank you, operator, and thank you, Dado. Our first question came from Morgan Stanley's [Ryan Lantz]. The hearing at the Vineland NJ data center site adjourned without a vote can we get an update on what's going on at this site? How is capacity progressing? How could this impact the site's plans to expand production since no vote has yet been cast?

Tom, can you help us answer this?

Tom Blackwell, Chief Communications Officer:

Of course. Happy to answer. Let me first address this aspect of the public hearings. I'd like to make a brief overall review first. Obviously, the US currently has a wider range of issues and debates surrounding data centers. We have all been paying close attention to these issues. Overall, we have found that we have developed an effective approach to entering a new region, which is to get involved early, explain very openly and transparently what we are doing, communicate with the local community, and really answer the questions that arise. We found this method very effective, and we used the same method in Vineland, NJ. Specifically, let's talk about Vineland.

Overall, our delivery is still progressing according to plan, and I'll let Andrey go into more detail from a construction perspective later. But as far as public hearings are concerned, what is happening now? DataOne is seeking final approval after revising its initially approved site layout plan. The reason the site layout plan was revised was because the project decided to switch the power source to Bloom.

Overall, we think switching to Bloom has significantly enhanced the project, including from a community perspective. This is an on-site power generation solution that quietly provides reliable electricity with extremely low emissions.

I would also like to say that these types of hearings are part of the normal process, and we have already taken these aspects into account when setting the schedule. As far as this hearing on the layout plan is concerned, we believe the layout complies with all applicable local, state, and federal laws and regulations. We are optimistic that once public comments are fully heard, approval will proceed quickly. Of course, as soon as we have more news, we'll be sure to keep everyone updated.

But next, maybe I could ask Andrey to add a few more points about construction.

Andrey Korolenko, Chief Product and Infrastructure Officer:

Sure, Tom. Hello everyone. As of now, we have delivered all the batches our contract requires us to deliver. We have every reason to trust that we will continue to deliver the remaining batches as required by the contract.

Adding a few more details, the main construction of the building was completed earlier this summer, and the supporting construction of the project is progressing quite smoothly. The Bloom fuel cell deployment should be very fast. Overall, we think the switch to Bloom was a very valuable and very positive adjustment to the project and is not expected to have a significant impact on the project timeline.

Gili Naftalovich:

OK, thank you. The next question is — we actually get a lot of questions about this on the platform. Marc, maybe you can help us explain these landmark deal announcements. Can you explain how these transaction processes work and why these customers chose Nebius?

Marc Boroditsky, Chief Revenue Officer:

Of course. Thanks for this question. We're very excited about these wins. Closing a $1 billion deal in our core AI cloud business is an important milestone in our market entry journey. This validates one of our basic beliefs: we can build a diverse customer base on a large scale. All 4 deals were won competitively and followed a similar pattern. Our size, performance, and reliability make up our differentiating strengths. At the same time, every customer sees us as a long-term partner who can grow with them.

In all of these cases, customers already had existing vendors, some even hyperscale cloud vendors, but they wanted to find new partners, acquire next-generation capacity and platforms, and expand. In fact, I can present one of these cases. They were introduced to us by one of our key strategic partners at the end of the first quarter. The customer wanted a large-scale, continuously deployed GB300 cluster at the time, and also needed flexibility for future expansion, strong technical support, and a long-term strategic partnership. When we signed with them in May, they stated that our responsiveness, transparency, white-glove support, and ability to meet both their current US deployment needs and future sovereign expansion requirements are our key differentiators. We're currently discussing more capacity with them, and they're also exploring our inference solution, Token Factory.

We won all of these deals through competition rather than the customer taking the initiative to come to the door. Before we signed up, we went through many rounds of contact. Customers validate our technology by personally conducting a POC proof of concept. One of our customers told us it was, without exaggeration, the best POC they've ever experienced. We've discussed training and reasoning with several of these customers to support their revenue plans and growth; at the same time, we're also discussing significant new capacity and next-generation chips, including Vera Rubin, with all of them. That's right, we have more deals like this. In the second quarter, our sales pipeline expanded again, including multiple opportunities worth more than $1 billion, across native AI companies, new AI labs, and enterprise customers.

Gili Naftalovich:

Thank you, Marc. The next question comes from Alex Duval from Goldman Sachs. Andrey, many of the deals we have signed are tied to the capacity of successive launches at the end of 2026 and throughout 2027. Given the market's focus on gigawatt data center construction, what makes you confident that capacity is climbing? Can you give us an update on our development timeline?

Andrey Korolenko, Chief Product and Infrastructure Officer:

Thank you Gili. We have made very good progress with the electricity we have contracted this year, and far ahead of expectations, surpassed the forecast for the end of 2026. As a result, we will — in fact, we have now raised our contracted electricity target to 5 gigawatts by the end of 2026. Of the contracted electricity we just mentioned, almost all will go online within the next 3 or maybe 3.5 years. And in fact, we're not going to stop.

These results are due to expansion in different regions and a combination of grid power supply and back-meter power supply. We are also now promoting relevant guarantees to obtain the right to use hundreds of megawatts of power generation resources.

We're also very excited about partnering with Bloom, which has enabled us to unleash and accelerate the development of many sites. I'd also like to point out that the vast majority of our contracts are cloud contracts, so we have the flexibility to decide exactly which site to deliver within a region, thereby reducing our reliance on any single location. Therefore, overreserving the site and overreserving capacity are our top priority tasks today, and I believe this will continue to be the case tomorrow. Our general idea is to deploy as much capacity as possible; if we can build as much capacity as possible, we can build as much as early as possible.

Gili Naftalovich:

Thank you, Andrey. Arkady, we get a lot of questions about strategy. The market is changing rapidly. You mentioned that prices are rising while planning to expand more than 1 gigawatt of capacity each year. You've also announced a number of new initiatives this quarter. Can you help us understand how these initiatives fit into our broader long-term strategy?

Founder, CEO and Non-Independent Executive Director Arkady Volozh:

Yes, you're right. The market is rapidly evolving, and in fact it is evolving even faster than anyone could have anticipated. The strength of our business model and platform is that both allow us to rapidly evolve with the market. I'll use a few examples from this quarter to illustrate.

First, as you can see, prices are changing, and they are changing fast, and prices are rising. However, our model was to build capacity in advance from the beginning, but not pre-sell capacity in advance. That's why we now have usable capacity, and we can allocate that capacity to contracts with shorter terms and much higher margins.

The second example is that, as I mentioned earlier, we also launched our first auction this quarter. This auction is for medium-term contracts. Why are we able to do this? Because we still have idle capacity that hasn't been allocated yet, and we have a multi-tenant platform. Now, we're just starting to realize the benefits of this model.

We have similar flexibility at the capacity level. Again, this is a strategic advantage of our platform. Our platform is versatile and can run on top of any 3rd party capacity, giving us an opportunity to further accelerate our growth. And this is at the heart of our asset-light model.

As a result, we see that there are many ways to monetize the versatility of a full-stack platform, both at the software level and at the hardware level. The way we think about this is actually simple: this market will continue to change rapidly and continue to grow rapidly. And we want to ensure that our business model and platform remain flexible enough to allow us to capture and capitalize on this growth.

Gili Naftalovich:

Thank you, Arkady. The next question comes from Arsenije Matovic of Wolfe Research, and it is also a key question raised by many investors on the platform. This question relates to our new $775 million asset-backed financing. The debt market has always been volatile, and the overall cost of comprehensive financing is rising. Are you still willing to continue to rely more on debt? Or should we expect companies to make more use of equity financing through ATMs or convertible bonds? Dado?

Chief Financial Officer Dado Alonso:

Thank you, Arsenije. Our approach is not to judge the direction of the market. Our core philosophy is to match the right financing instruments with the right assets while maintaining discipline in three areas: our cost of capital, minimising shareholder dilution, and maintaining a strong and healthy balance sheet.

Our primary sources of capital remain customer advances and operating cash flow. We expect to receive more than $9 billion in upfront customer payments in 2026. Obviously, as we continue to scale up, this will directly reduce the amount of external financing we need.

In addition to this, asset-backed financing is an important part of our strategy. Our $775 million financing completed in July was priced at SOFR plus 250 basis points, backed by deployed GPU infrastructure and contractual cash flow from an investment-grade customer. This proves that even as market volatility increases, investors still have strong demand to finance these contract cash flows on attractive terms. And, given our backlog of committed orders of over $40 billion, we believe this is a highly scalable and replicable financing model. Obviously, we expect to continue to finance in this market as more capacity is deployed.

In addition to asset-backed financing, we also have a lot of flexibility. We currently have almost no company-level debt, so as our business expands, this forms another source of capital we can evaluate. Equity and equity-linked financing, such as ATMs, or potential convertible bonds, can also provide additional sources of funding; we evaluate these along with all other financing options. We are actively considering further use of equity-linked and asset-backed financing while continuing to evaluate company-level debt and other financing alternatives.

So overall, we are still very confident about the current funding situation. We have multiple sources of capital at our disposal and will continue to optimize between them, while focusing on the cost of capital, limiting shareholding dilution, and maintaining a disciplined balance sheet.

Gili Naftalovich:

Thank you, Dado. Our next question comes from Citi's Tyler Radke. Marc, I leave this question to you. For 2027 capacity, how do you consider the allocation mix between short-term capacity and multi-year transactions? For longer term deals, are you waiting for each MW of ACV to reach a certain threshold before signing?

Marc Boroditsky, Chief Revenue Officer:

Thank you, Tyler. Before I answer in detail, I'd like to clarify in more detail the type of transaction we are doing. The vast majority of our long-term — to be precise, our mid-term contracts — are AI cloud customer agreements, which are at the core of our business. In these deals, we were able to lock in very strong unit economics with some of the world's most ambitious AI companies.

Then, as you pointed out, we also have shorter term opportunities. These are high premium deals. Examples include the auctions we're introducing today, and the ones we mentioned that can capture short-term large-scale contracts.

In practice, we strive to optimize between customer types, prices, payment structures, terms, and transaction size, rather than strictly focusing on a single variable or an aggregated metric. Currently, our first emphasis is on taking care of existing customers, followed by acquiring new customers, then the terms of the transaction; in terms of priority of transaction terms, the first advance payment is followed by the term. Therefore, we are considering a wider range than simply focusing on ACV per megawatt.

At the same time, we've also tactically shortened the time to sell capacity in advance. In effect, that means making sales happen closer to the point in time of deployment. This ultimately raised the price we were able to get while retaining flexibility. We are consciously allocating part of our capacity to short-term and immediate needs, as this is the area where we currently see the highest comprehensive value.

In summary, as we deploy capacity for the second half of this year and next year, we will allocate between medium- and short-term deals, and continue to use the customer and transaction terms perspective I just mentioned to determine allocations.

Gili Naftalovich:

Thank you, Marc. The next question comes from Water Tower's James Kisner. xAI has begun selling computing power at a premium. What does this say about the market pricing of AI capacity? Are you seeing similar intensity in new contracts and renewals? What does this mean for Nebius?

Founder, CEO and Non-Independent Executive Director Arkady Volozh:

This question is probably up to me to answer. I'm Arkady. There's no doubt that this won't change Nebius or our plans in any way. Our participation is essentially the same market as the three largest hyperscale cloud vendors. The entire AI cloud market is growing at a tremendous rate, from hundreds of billions of dollars a year to possibly 1 trillion dollars, and some even think it will be higher. There's no doubt that the current major players will continue to grow together with the market, right? But at the same time, there is no doubt that this market has opened a window for new entrants and independent players like us.

Yes, we want to build 1 gigawatt of capacity each year. Yes, there aren't many companies that can do this. These numbers may seem huge, but the entire market is adding tens of gigawatts of capacity every year. The entire market is several orders of magnitude higher than ours. Hyperscale cloud vendors will build a significant portion of this capacity. But even so, it's clearly impossible for them to build up all of the capacity. They need someone to help them build. But that wasn't us. Somebody — because someone always has to build the remaining part of the market, and we will build that part. This is where we belong, and where we see the greatest opportunities.

So, back to this question, new players entering the market won't change our market. In fact, in my opinion, this is just validating this market.

Gili Naftalovich:

Thank you, Arkady. The next question comes from Baird's Rob Oliver. Is your strong position in the field of open weighting models driving customers to increase their use of inference?

Roman, I leave this question to you.

Roman Chernin, Chief Commercial Officer:

OK. Thank you, Gili, and thanks for this question. First, we believe competition and diversity are beneficial to customers, businesses, and society as a whole. Since the company was founded, we have been committed to building an open AI ecosystem and providing an open infrastructure without lockdown effects. As a result, customers gain flexibility, control over their data and models, and the freedom to decide how to deploy.

As the scale increases, economy becomes critical. Companies at the cutting edge of AI adoption are seeing the cost of token consumption and are beginning to ask the question: Can AI not only solve tasks, but also complete tasks in a way that is economical enough to achieve scale?

Another important issue is how to extract value from the expertise and data held by the company and its employees and transform it into an AI system with better performance. Everyone is saying that data is the moat, maybe even the only moat.

All of these factors are driving companies to adopt a more specialized model. Open source models are valuable not only because they can be obtained, but also because they can be tuned, trained, and ultimately optimized for a specific customer, a specific enterprise, and a specific application scenario.

What's changing now is that the quality of open models is rapidly improving. At the same time, the entire industry is shifting to the flexibility and control we built right from the start. As a result, Token Factory was able to support the cutting-edge open model on day 0. And the speed of iteration is insane. If we just look at the past 4 weeks, we've welcomed Nemotron Ultra, GLM 5.2, Kimi K3, the new DeepSeek Flash, and Minimax 3. Today, the new Nemotron Lightning was released again. But just being able to provide these models is far from enough. We're building the capacity to service these models without sacrificing quality, cost, or performance.

Take GLM 5.2 for example. Some people refer to the moment this model was released as “DeepSeek Moment 2.” Our implementation received a 100% quality score and was verified by independent benchmarking and Artificial Analysis, leading the way in performance. This puts us in a very good position to drive and capture the growing demand for open source models in the current market.

Gili Naftalovich:

Thank you, Roman. Rob has a follow-up question. Can you describe some of the current early aspects of the asset-light business, including initial customer interactions, and how economical this model is compared to the core business?

Arkady, would you like to answer this question?

Founder, CEO and Non-Independent Executive Director Arkady Volozh:

Yes. First, we're expanding globally, but we're still a startup, so we have to be very — very careful about where to invest our capital. We can't access all markets at once. That's why we love to work with our partners, and our platform allows us to do that.

After we announced our asset-light model, we've received inquiries from dozens of potential partners. These partners have significant capacity and sufficient capital, but they don't know how they should be built or how they should be sold. As GPUs become increasingly an investable asset — check out recent news — we expect more and more companies will want to enter this market, but they will need our help to actually deliver this capacity to customers. By having the right technology platforms and channels to market, we can help them.

So this is actually exactly what we can offer them very easily. This is why the asset-light model exists. This model is still in its very early stages, but we are very encouraged by the early signals, and we think it has great potential.

Gili Naftalovich:

Our next question comes from Morgan Stanley's Ryan Lantz. Dado, you just said you could sell the full 2027 capacity today. How should investors understand 2027 from the three dimensions of capacity, price, and revenue? When will you officially give your outlook?

Chief Financial Officer Dado Alonso:

Thank you Ryan. This is probably one of the most noteworthy issues for investors attending this conference call. The deals we signed in this quarter exceeding $20 million per megawatt with a payback period of less than 2 years are expected to be launched one after another starting at the end of the fourth quarter. These deals could serve as a pricing benchmark for early next year, because at current prices, we can actually sell out all of our planned capacity today. However, we actively chose not to do so, which reflects our confidence in future price dynamics.

Of course, capacity is also a very important part of this equation. We will be deploying significantly more capacity in 2027 than we have already deployed in 2026 and will be deployed for the rest of the year. We expect developments in both capacity and price dimensions to make us extremely excited for 2027, but official guidance will be announced later this year.

And don't forget, in addition to this, we also anticipate that asset-light models and high-value services such as intelligence and inference solutions will contribute an increasing share of revenue while driving higher profit margins.

Gili Naftalovich:

Thank you, Dado. Our next question comes from Cantor's Brett Knoblauch. Do we still expect 800 MW to 1 GW by the end of the year?

I see -- Andrey, would you like to answer for us?

Andrey Korolenko, Chief Product and Infrastructure Officer:

Yes, of course. Yes, we still — we still expect to achieve this target of 800 megawatts to 1 gigawatt of connected electricity this year. Also — it seems that part of the question just now involves whether Vineland will be completed by 2027. No, Vineland is at 2026 capacity and has electricity.

What I want to explain is that connected electricity represents a data center, and there are still a few steps between data center and revenue. You need to debug data centers, build networks, build clusters, deploy platforms, and then let customers access before revenue starts to be generated. This will take a few months. It may also depend on GPU generational switching. As a result, there is a time lag between when electricity is connected and revenue is actually generated.

So as far as our 800 megawatt to 1 gigawatt guide is concerned, yes, this is more -- to be precise, this refers to the capacity that has already been connected; and I think this capacity will gradually enter active operation during the first half of 2027.

Gili Naftalovich:

Thank you, Andrey. We also received questions from BNP Paribas Stefan Slowinski, specifically about our third-quarter monetization level of $40 million per megawatt. Does this reflect the initial pricing for selling Vera Rubin capacity?

Andrey, can you talk about Vera Rubin's capacity climbing?

Andrey Korolenko, Chief Product and Infrastructure Officer:

Of course. In fact, we've had Vera Rubin in our lab for quite some time and have now obtained the expected test results. First, I'd like to point out that transitioning from Grace Blackwell to Vera Rubin is easier on a technical level than transitioning from a previous generation product to Grace Blackwell. We expect to begin deploying Vera Rubin at the end of this year or early next year, and continue to deploy throughout next year.

Gili Naftalovich:

OK. Going back to the first part of this question, we've received a lot of questions from the platform about new market entry models. So Marc, I think this is a good question for you to answer. Since the end of this quarter, you've mentioned both capacity auctions and short-term large-scale training deals. Can you explain to us the strategy behind these models? What did they tell you about the price? Also, do these contracts still use the pay-as-you-go model?

Marc Boroditsky, Chief Revenue Officer:

Thank you Gili. Up to now, I think everyone should agree that this market is extremely dynamic and changing very fast. We're always looking for ways to discern real signals and deepen our understanding, so we can gain a more accurate and complete understanding of our business.

We launched these initiatives with the primary purpose of learning and validation while building customer relationships in a disciplined manner. The short-term large-scale training capacity deals we mentioned earlier, specifically serve customers that require dedicated large-scale computing power — in this case the GB300 cluster — for a period of 3 to 6 months, and the customer is willing to pay a premium. These customers have clear requirements, such as completing large-scale training tasks with fixed time windows before models are released, or for post-reinforcement learning training sprints.

In the field of models, we've heard this phrase over and over again: a few weeks is very important, and being able to obtain reliable, high-performance AI computing power is just as important. These customers are willing to pay for speed and certainty, as long as they can get the computing power they need within a controlled time window. The core of auctions, on the other hand, lies entirely in clear price discovery. In the current environment, market prices may even change as the sales cycle progresses. The real challenge is to accurately determine the fair value of the products we provide at any point in time. The information given by traditional reference metrics, competitor pricing, analysts' opinions, and even predicting the market is extremely scattered. In a market where there are multiple buyers behind every GPU, we simply let the market tell us the answer directly.

The auction turned out to be 15% higher than any price we've ever seen, and 20% higher than the price in the Blackwell sales pipeline. Moreover, the winning customer was very satisfied with this experience. In particular, they pointed out that the combination of proven pricing and the certainty that the required computing power can be obtained is very valuable. They also told us that they plan to continue participating in auctions in the future.

Therefore, the conclusion you should take from this is that these two models both prove customer value and are very profitable in themselves. We use only a small portion of our overall capacity to drive price discovery and value verification. These steps will help us evaluate our capacity and the overall market as we move into 2027. They all broadly influence our pricing, product packaging strategies, and deal negotiations, and naturally have a shorter cycle from signing to deployment.

Gili Naftalovich:

Thank you, Marc. We've received a lot of questions about deploying more than 1 gigawatt of additional capacity each year starting in 2027. Dado, can you explain how we plan to finance these capital expenses on a large scale? How do we prioritize the various sources of financing?

Chief Financial Officer Dado Alonso:

Of course. The most important point is to understand that we actually have multiple capital pools to finance growth. We are very confident in our ability to finance the capacity we plan to deploy in 2027 and beyond. Obviously, the first source is operating cash flow. We are now generating positive operating cash flow, and we anticipate a significant increase in operating cash flow as the business expands.

Then there's the customer's advance payment. The contract terms we received in the second quarter were able to cover approximately 50% to 60% of the associated capital expenses. But looking ahead, one of our goals is to further increase this coverage ratio.

Third, long-term contracts with investment-grade clients provide us with a very solid foundation to raise asset-backed financing on attractive terms. We currently have a backlog of approximately $40 billion of committed orders, and we can use these contracts as a basis for borrowing, and we began using this funding source in the second quarter.

In addition to this, we can also obtain additional flexibility through company-level debt and equity-linked financing, and neither of these sources of financing is currently being used on a large scale by us. As our business expands, we will continue to balance various funding sources in a disciplined manner to maintain a strong balance sheet.

At the same time, we are also seeing additional financing opportunities emerging around GPUs as an asset class. We are seeing increasing activity and interest in the market to finance GPUs as standalone assets, and we are very convinced of the potential of this market. Over time, this may become another attractive source of funding for us.

Overall, we are indeed very confident about the state of our balance sheet today, and we are also confident that we will finance growth in 2027 and beyond through this diversified financing strategy while maintaining financial discipline.

Gili Naftalovich:

Thank you, Dado. The next question came from one of the investors on the platform. Roman, what trends are we currently seeing in the market relating to Token Maxing? Is this affecting Token Factory usage or adoption in Tbilisi?

Roman Chernin, Chief Commercial Officer:

OK. Thanks for this question. First, I'd like to remind everyone that we held our first Inflection event in June, which was probably the first time we publicly detailed our product strategy: we built an AI cloud layer by layer to meet the needs of developers at different levels, extending from large-scale bare metal infrastructure to the intelligent layer. One very clear trend is that AI systems are entering production environments. Programming is the most obvious example, but it's by no means limited to programming. We are seeing the emergence of long-running intelligent workflows in the financial services sector, such as Revolut and MasterCard applications; in the e-commerce sector, Shopify is using related applications to improve customer experience and business processes; in addition, there are many other fields, such as Sword Health in the medical field and Higgsfield in the marketing automation field.

As always, we ourselves are our own “zero customer.” Echo is our own infrastructure agent, launched in a recent cloud product launch. It runs an open source model serviced by Token Factory. We see the challenge our customers face is how to run complex systems made up of multiple models, inference engines, and tools at scale. We help clients solve this problem through a full suite of services, including Token Factory to provide reliable, high-performance inference and post-training; and Avride for information anchoring, particularly as businesses transition from a closed ecosystem with built-in search capabilities to a more open one.

The Eigen AI and Clarifai team are now fully integrated into the Token Factory project and have begun delivering features according to our product roadmap. We are now able to provide day 0 support for the release of major open models, perform quantifiable performance optimizations after launch, while independent benchmarking — as I mentioned before — continues to rank us as one of the leading inference platforms.

The second quarter was also Tavily's first full quarter after joining Nebius. Its developer community has grown from 1 million developers in February to over 2.5 million developers. Additionally, it has launched a keyless paper search specifically designed for use by autonomous agents, and has obtained a range of certifications for enterprise deployment.

We're also seeing more and more customers reinforce their models, which creates additional reasoning and information anchoring requirements throughout the development cycle, not just during production. Reinforcement learning rollouts, evaluation, synthetic data generation, and information anchoring and training workflows require significant inference capacity and reliable access to external information. These trends validate our strategy to build a vertically integrated platform—one that provides an attractive total cost of ownership while supporting diverse workloads.

We want to cover the customer's complete AI lifecycle, from training and post-training to inference and information anchoring. This is where our infrastructure and software combine to form powerful product capabilities, and we're just getting started.

New workloads are also putting new demands on physical infrastructure. Also, since we're building full stack capabilities, we're able to handle these changes. In addition to deploying the next-generation GPUs mentioned by Andrey, tasks such as agent orchestration, tool calls, and data preparation all rely heavily on the CPU. As a result, we are increasing both ARM and CPU deployments outside of the GPU cluster.

Gili Naftalovich:

Thank you. This concludes today's results conference call.

Operator:

That concludes today's conference call. Thank you all for participating. You can now disconnect.