目录
I. Participants and Forward-Looking Statements
II. Management Remarks
CEO: Three Deal Structures and the Capacity Strategy
CFO: Financial Performance, Capital Expenditure, and Full-Year Guidance
III. Q&A
Vineland Data Center Approval and Construction Progress
Why Nebius Won Four $1 Billion-Scale Deals
5 GW Contracted-Power Target and Capacity Ramp
Long-Term Strategy in a Rapidly Changing Market
Choosing Among Debt, Equity, and Asset-Backed Financing
Allocating 2027 Capacity Between Short- and Medium-Term Contracts
Market Implications of xAI Selling Compute at Premium Prices
Open-Weight Models and Inference Demand
Early Progress and Economics of the Asset-Light Business
2027 Capacity, Pricing, and Revenue Outlook
Year-End Target of 800 MW to 1 GW
Vera Rubin Deployment Timeline
Capacity Auctions and Short-Term Training Contracts
Financing the Gigawatt-Scale Expansion in 2027
Token Usage Growth, Token Factory, and Tavily
本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读
I. Participants and Forward-Looking Statements
Gili Naftalovich, Head of Investor Relations: Hello, and welcome to Nebius’ second-quarter 2026 earnings call. Joining us today are CEO Arkady, CFO Dado, and other members of the Nebius management team.
Certain statements regarding our operations and financial performance may be forward-looking. These statements are based on current expectations and assumptions and involve risks and uncertainties that could cause actual results to differ materially. Please refer to the company’s Form 20-F for a discussion of relevant risk factors. The company undertakes no obligation to update any forward-looking statements.
Today’s call will cover both GAAP and non-GAAP financial measures. Reconciliations are included in the earnings release issued today, and all related materials are available on the company’s investor relations website. I will now turn the call over to Arkady.
II. Management Remarks
CEO: Three Deal Structures and the Capacity Strategy
CEO Arkady: Thank you, Gili, and welcome, everyone. I would like to use today’s call to provide further context on Nebius’ business model.
We delivered an exceptional quarter. Demand for the products and services we are building remains enormous, and our business model positions us extremely well to capture it.
We build capacity ahead of signed contracts and serve customers through our multi-tenant cloud platform and software stack. These customers include AI-native companies, leaders in agentic AI, NeoLabs, and some of the world’s most technically sophisticated enterprises.
Our strategy is working. We retain control over when capacity is sold, to whom, on what terms, and how each project is financed. This flexibility allows us to serve independent AI developers while supporting an open, diverse, and competitive market.
We address customer demand through three types of transactions, each with a distinct duration, pricing structure, and strategic role in our business.
The first comprises medium-term contracts in our core AI cloud business, typically lasting one to three years and serving the world’s most ambitious AI companies. This quarter alone, we signed four landmark agreements with Reflection, Cohere, a major U.S. NeoLab, and a large U.S.-based quantitative trading firm. Each agreement has an average contract value exceeding $1 billion.
These contracts generate approximately $20 million to $25 million per MW annually, with customer prepayments covering 50% to 60% of the associated capital expenditure.
More importantly, we could sell all of our 2027 capacity today on these terms if we chose to. We have not done so because retaining some capacity for customers’ short-term and immediate needs creates greater value.
This brings us to the second category: short-term capacity contracts, typically lasting no more than six months. These customers have immediate, high-value requirements within clearly defined time windows and are therefore willing to pay a substantial premium. Transactions currently under discussion in this category are priced at $40 million to $50 million per MW annually, and sometimes higher. We recently signed one such agreement.
Both transaction types will come online later this year and therefore will not materially affect our 2026 revenue guidance, but they will contribute to revenue in 2027 and beyond.
The third category consists of the long-term contracts with investment-grade customers that we have discussed previously. These agreements play an important role by allowing us to finance capacity development more quickly and efficiently.
The secured debt financing completed in July was underpinned by one such transaction. We currently have a $40 billion contract backlog and expect to execute additional financings of this kind.
We have continued to build capacity, advance innovation, develop our software stack, and expand our services and product portfolio. This quarter, we introduced innovations across several areas.
First, we conducted our inaugural capacity auction, with highly successful results. The clearing price was the highest we have seen for the Blackwell generation, exceeding our previous peak rate by 15%. This provided a strong real-time signal of the market’s valuation of this capacity and represented an innovative go-to-market model.
We are also exploring new capacity-development models. The asset-light partnership model launched this quarter gives us another avenue for expansion while addressing two of the industry’s principal constraints: capital and capacity.
Under this model, partners finance, build, and operate the facilities, while Nebius provides the complete technology platform and customer demand. We deliver value-added services on top of our partners’ infrastructure, generating high-margin revenue with minimal balance-sheet capital.
We expect this model to unlock additional capacity beginning in 2027. In parallel, we continue to develop our future capacity pipeline through owned and colocated sites and have raised our year-end contracted power target to 5 GW.
Our capacity pipeline effectively makes Nebius one of the few companies globally capable of adding more than 1 GW of capacity annually. We plan to achieve that milestone in 2027.
We delivered on every objective set for the quarter and exceeded most of them. We signed landmark agreements in our core markets on better-than-expected terms, capacity is expanding to meet demand, and our platform is addressing the needs of the broader industry.
Today’s update is about more than a strong quarter already behind us. We have visibility into demand, supply, and transaction terms for 2027 and beyond, as well as our future capacity-development plans. We are extraordinarily excited about the outlook.
CFO: Financial Performance, Capital Expenditure, and Full-Year Guidance
CFO Dado: We successfully executed our strategy in the first half and entered the second half with strong momentum.
Although most of our incremental 2026 capacity will not come online until the second half, second-quarter revenue and annualized run-rate revenue again posted triple-digit growth. At the same time, adjusted EBITDA margins expanded significantly, while substantial customer prepayments and a more diversified capital base further strengthened our financing capacity.
This strong start keeps us firmly on track to achieve our full-year strategic and financial objectives. Unless otherwise stated, all comparisons are year over year.
Second-quarter group revenue increased 454% to $582 million, up 46% sequentially. Nebius AI grew even faster, with revenue rising 514% to $575 million and accounting for 98% of group revenue.
At the end of June, annualized run-rate revenue reached $3 billion, up 598% year over year and 56% from $1.9 billion at the end of March.
Growth was driven primarily by capacity added in the first quarter, higher utilization, and high-margin revenue from our asset-light business model, Token Factory, and recently acquired businesses. Higher utilization also reflected improved efficiency across the underlying infrastructure.
Our capacity was once again fully sold, as new capacity is absorbed rapidly after coming online.
Group adjusted EBITDA was $236 million, compared with a loss of $21 million a year earlier and $129.5 million in the prior quarter. Group adjusted EBITDA margin reached 41%, up from 32% in the first quarter.
Nebius AI generated $286 million of adjusted EBITDA at a 50% margin. The margin difference between the group and Nebius AI primarily reflects our investment in Avride and TripleTen, both of which remain at an early stage of development. We expect Nebius AI to continue contributing the substantial majority of group adjusted EBITDA.
Margin expansion was driven by revenue growth, the initial contribution from our asset-light model, and contributions from Token Factory and recently acquired businesses. Even as we continue investing for growth, we see a clear path to further margin improvement this year, in 2027, and over the longer term.
We have strong visibility into pricing and expect capacity coming online at our owned data centers beginning in the second half of next year to drive further margin improvement.
Customer prepayments generated substantial cash inflows during the quarter. Approximately 70% of the agreements signed in the second quarter included prepayment provisions. We expect customer prepayments to provide more than $9 billion of upfront funding this year, directly reducing the amount of capital we need to raise through debt and equity.
Second-quarter operating cash flow was $2.3 billion, and cash and cash equivalents totaled $8 billion at quarter-end.
During the second quarter, we also used our at-the-market offering program to issue 12.7 million Class A ordinary shares at a weighted-average price of $224 per share, generating approximately $2.8 billion in gross proceeds. As of June 30, 12.3 million shares remained available for issuance under the program.
We view the ATM program as a flexible financing tool—an option available to us, not a commitment we are required to exercise.
In July, we announced our first $775 million asset-backed debt financing. Secured by contracted cash flows, the facility carries only a modest spread over the benchmark rate, equivalent to a mid-single-digit percentage at current levels.
Beyond this financing, we have more than $40 billion of customer commitments on similar terms under strategic agreements already signed with investment-grade companies. We will continue diversifying our funding sources while maintaining an appropriate balance between debt and equity.
We are pursuing additional asset-backed financings while also evaluating corporate-level debt and other financing instruments.
Second-quarter capital expenditure was approximately $5.7 billion, primarily for GPUs, GPU-related hardware, and data center expansion. Capital deployed to date provides the foundation for achieving our full-year capacity target.
Solid progress during the first half further strengthened our confidence in the 2026 outlook. We are therefore reiterating all elements of our full-year 2026 guidance:
Annualized run-rate revenue of $7 billion to $9 billion;
Group revenue of $3 billion to $3.4 billion;
Group adjusted EBITDA margin of approximately 40%;
Capital expenditure of $20 billion to $25 billion.
We remain confident that capacity deployment will accelerate in the second half and expect capacity brought online near the end of the second quarter to begin contributing revenue in the third quarter.
As Arkady noted, we are now building capacity to meet 2027 demand. Contracted customer commitments provide strong visibility into the near-term revenue associated with these investments.
We are also actively pursuing additional high-margin revenue opportunities, including new business models, capacity auctions, short-term capacity transactions, and asset-light operations.
Our strong second-quarter performance reflects disciplined execution in scaling the business and improving profitability. We delivered robust revenue growth, increased earnings, created a more capital-efficient path for capacity expansion, and broadened our financing channels.
Looking ahead, we will continue scaling rapidly to capture the market opportunity before us while maintaining balance, discipline, and a clear focus on creating long-term shareholder value.
III. Q&A;
Vineland Data Center Approval and Construction Progress
Moderator: Morgan Stanley’s Ryan Lantz asks: The hearing for the Vineland data center project in New Jersey was adjourned without a vote. Can you provide an update on the project and the capacity buildout? How does the absence of a vote affect the site’s expansion plans?
Tom: Let me begin with the public-hearing process.
Data centers are the subject of a broader debate across the United States, which we continue to monitor closely. Overall, our approach when entering new communities has proven effective: engage early, explain openly and transparently what we are doing, communicate with the local community, and thoughtfully address questions from all stakeholders.
That is the approach we have consistently taken with the Vineland project in New Jersey.
Overall project delivery remains on schedule. Andrey will provide more detail shortly from a construction and engineering perspective.
Regarding the public hearing, the project is seeking amended final approval for revisions to the previously approved site plan. The plan was revised following the decision to switch the project’s power source to Bloom’s on-site generation solution.
We believe the switch to Bloom will materially improve the project overall, including its value to the local community. The solution generates power on-site quietly and reliably, with extremely low emissions.
Hearings of this kind are part of the normal approval process, and these steps were incorporated into the project schedule.
For this site-plan hearing, we are confident that the proposal complies with all applicable local, state, and federal laws and regulations. We are optimistic that approval will follow promptly once public comments have been fully heard. We will provide updates as further developments occur.
Andrey: To date, we have delivered every capacity milestone required under the contract, and we have strong reason to believe the remaining phases will also be completed in accordance with the contract.
To add some detail, the building shell was completed earlier this summer, and mechanical, electrical, and engineering work is progressing well. Deployment of the Bloom fuel cells should proceed quickly.
Overall, we view the switch to Bloom as both feasible and appropriate, and we do not expect it to have a material impact on the project schedule.
Why Nebius Won Four $1 Billion-Scale Deals
Moderator: The investor platform has received numerous questions about the landmark deals. Mark, can you discuss the bidding processes and why customers selected Nebius?
Mark: We are extremely pleased to have secured these projects. Signing $1 billion-scale agreements in our core AI cloud business is an important milestone in our go-to-market development and validates our core thesis: Nebius can build a diversified customer base while scaling rapidly.
All four deals were competitively won through broadly similar processes. Our scale, performance, and reliability were key differentiators. Each customer also viewed Nebius as a long-term partner that could grow alongside them.
All of these customers already had suppliers, including hyperscalers in some cases. They were seeking new partners and opportunities to expand through next-generation capacity and platforms.
In one example, a major strategic partner introduced us to the customer near the end of the first quarter. The customer wanted a large, contiguous GB300 cluster, together with flexibility for future expansion, strong technical support, and a long-term strategic relationship.
When we signed the agreement in May, the customer specifically cited Nebius’s responsiveness, transparency, dedicated high-touch support, and ability both to meet its immediate U.S. deployment requirements and support future local expansion as key differentiators.
We are already discussing additional capacity with this customer, which is also evaluating our Token Factory inference solution.
We proactively pursued and won all of these transactions; they were not simply inbound opportunities that converted naturally. Before signing, we went through multiple rounds of engagement and discussion. Customers also tested our technology through hands-on proofs of concept. One customer described it as the best proof of concept they had ever experienced.
To support customers’ own revenue plans and growth objectives, we are already discussing post-training and inference services with several of them, as well as substantial additional capacity and next-generation chips—including Vera Rubin—with all four.
Our pipeline contains more deals of this kind. Opportunity development accelerated again in the second quarter, including several opportunities exceeding $1 billion across AI-native companies, NeoLabs, and large enterprises.
5 GW Contracted-Power Target and Capacity Ramp
Moderator: Goldman Sachs’s Alex Duval asks: Most of the company’s signed deals relate to capacity coming online between late 2026 and 2027. Given market concerns about building gigawatt-scale data centers, what underpins the company’s confidence in the capacity ramp? Can you provide an updated development timeline?
Andrey: We have made excellent progress on contracted power this year and have already surpassed our original year-end 2026 target ahead of schedule. We are therefore raising our year-end 2026 contracted-power target to 5 GW.
The vast majority of this contracted power will come online progressively over approximately the next two to three-and-a-half years. Nor do we intend to stop expanding there.
This progress spans multiple regions and combines grid supply with behind-the-meter generation. To date, the company has secured access to hundreds of MW of behind-the-meter generation resources.
We are also highly optimistic about our partnership with Bloom, which can help unlock the potential of multiple sites and accelerate development.
Importantly, the vast majority of our contracts are cloud-service agreements, giving us flexibility to select the specific delivery location within the agreed region and reducing dependence on any single site.
Securing more sites and capacity than we ultimately need is currently our highest priority, and I believe it will remain so. Our overall approach is to deploy as much capacity as possible and build it as early as possible.
Long-Term Strategy in a Rapidly Changing Market
Moderator: Arkady, the market is evolving rapidly, you have noted that prices are rising, the company plans to add more than 1 GW of capacity annually, and several new initiatives were announced this quarter. How do these initiatives fit into the company’s long-term strategy?
CEO: The market is indeed evolving extremely quickly—even faster than anyone expected. The strength of Nebius’s business model and platform is that both allow us to adapt rapidly as the market changes.
This quarter provided several examples.
First, market pricing is moving quickly and continuing to rise. From the outset, our model has been to build capacity ahead of demand without preselling all of it far in advance. As a result, we now have available capacity that can be allocated to shorter-duration, higher-margin contracts.
Second, the company held its first auction of capacity under medium-term contracts this quarter. Again, we could do so because we have unallocated capacity and a multi-tenant platform. We are now beginning to realize the benefits of those advantages.
We have similar flexibility on the capacity side. The Nebius platform is general purpose and can run on any third-party capacity, allowing us to grow faster and forming the foundation of our asset-light business model.
We see multiple ways to monetize the versatility of our full-stack platform, across both software and hardware.
Our thinking is straightforward: this market will continue to change and grow rapidly. We will ensure that our business model and platform remain flexible so that we can capture the opportunities created by growth across the market.
Choosing Among Debt, Equity, and Asset-Backed Financing
Moderator: Wolfe Research’s Arsenije Matovic asks: The company recently secured $775 million of asset-backed financing. Debt markets have been volatile, and overall financing costs have generally risen. Is the company still willing to use debt, or will it rely more heavily on ATM equity issuance or convertible debt?
CFO: Our approach is not to predict the direction of markets, but to match the right financing instrument with the right asset while maintaining discipline in three areas: cost of capital, minimizing shareholder dilution, and preserving a strong balance sheet.
Our primary sources of capital remain customer prepayments and operating cash flow. We expect to receive more than $9 billion of upfront customer prepayments in 2026, directly reducing the external financing required to fund business expansion.
Beyond that, asset-backed financing is an important component of our strategy.
The $775 million financing completed in July was priced at SOFR plus 250 basis points and backed by deployed GPU infrastructure and contractual cash flows from investment-grade customers. This demonstrates that, even amid heightened market volatility, investors remain highly willing to finance these contractual cash flows on attractive terms.
With more than $40 billion of committed backlog, we believe this is a highly scalable, repeatable financing model. As we deploy more capacity, we expect to continue accessing this market.
Beyond asset-backed financing, we retain substantial flexibility. We currently have almost no corporate-level debt, so as the business scales, that could become an additional source of capital to evaluate.
Equity and equity-linked financing, including ATM issuance or potential convertible debt, could provide additional funding. We will evaluate these alongside our other financing alternatives.
We are actively considering further equity-linked and asset-backed financing while continuing to evaluate corporate-level debt and other funding options.
Overall, we are highly confident in our current financing position. The company has access to multiple sources of capital and will continue optimizing for cost of capital, reduced dilution, and a strong balance sheet.
Allocating 2027 Capacity Between Short- and Medium-Term Contracts
Moderator: Citi’s Tyler Roddy asks: How is the company thinking about allocating 2027 capacity between short-term capacity contracts and multiyear agreements? Has the company established an annualized contract-value-per-MW threshold that must be met before signing a long-term contract?
Mark: First, it is important to distinguish more precisely among the different types of deals we are signing.
Medium-term contracts are primarily agreements with AI cloud customers and represent our core business. These contracts allow us to lock in attractive unit economics with the world’s most ambitious AI companies.
Short-term opportunities are a separate category of shorter-duration, premium-priced transactions, including the capacity auction discussed today and the large-scale short-term contracts we have referenced.
In practice, we optimize across customer type, price, payment structure, contract duration, and transaction size rather than focusing on a single variable or aggregate metric.
Our current priority order is existing customers first, then new customers, followed by contract terms. Within contract terms, our priorities are price, upfront prepayments, and contract duration, in that order. We therefore consider a broader set of factors than annualized contract value per MW alone.
At the same time, we have strategically reduced how far in advance we sell capacity. In other words, we sell closer to the actual deployment date. This both improves achievable pricing and preserves flexibility.
We deliberately allocate a portion of capacity to short-term and immediate demand because these opportunities currently generate the highest overall realized value.
As capacity comes online through the second half of this year and into next year, we will continue allocating it between medium- and short-term transactions using the customer and contract-term framework just described.
Market Implications of xAI Selling Compute at Premium Prices
Moderator: Water Tower’s James Kisner asks: xAI has begun selling compute resources at premium prices. What does this indicate about pricing in the AI capacity market? Is the company seeing similar strength in new contracts and renewals? What does this mean for Nebius?
CEO: This does not change Nebius’s strategy or our plans.
Fundamentally, we participate in the same market as the three major hyperscalers. The overall AI cloud market is growing from hundreds of billions of dollars annually toward potentially $1 trillion, with some forecasts even higher.
The incumbent leaders will undoubtedly continue growing with the market, but there is also clear room for new entrants and independent players such as Nebius.
We aim to build 1 GW of capacity annually, and very few companies can do that. Those numbers appear large, but the overall market is expanding by tens of GW each year—multiple orders of magnitude larger.
Hyperscalers account for much of that buildout, but even they cannot construct all the required capacity. Other participants will therefore need to serve the remainder of the market, and that is Nebius’s opportunity.
The entry of new participants does not change Nebius’s addressable market; it further validates the market’s scale and attractiveness.
Open-Weight Models and Inference Demand
Moderator: Baird’s Rob Oliver asks: Is Nebius’s strength in open-weight models driving customers to increase their inference workloads?
Roman: We believe customers, enterprises, and society as a whole benefit from competition and diversity. Since the company’s inception, we have been committed to building an open AI ecosystem and providing open infrastructure without lock-in.
This gives customers flexibility, control over their data and models, and the freedom to choose how they deploy.
Once applications reach scale, economics become critical. Companies at the frontier of AI applications are focusing on token-consumption costs and asking a key question: AI must not only solve the task, but do so with economics that support deployment at scale.
Another important question is how to extract the specialized knowledge and data held by enterprises and their employees and convert it into higher-performing AI systems. Data is widely described as a moat—potentially the only moat. These factors are driving enterprises toward more specialized models.
The value of open models lies not only in their availability, but also in their ability to be fine-tuned, trained, and optimized for specific customers, enterprises, and use cases.
What is changing now is the rapid improvement in open-model quality. At the same time, the broader industry is moving toward the flexibility and control that Nebius has provided from the outset.
Token Factory can support frontier open models from the first day of release, and the pace of recent iteration has been remarkable. In just the past four weeks, the market has seen the release of Nemotron 3 Ultra, GLM-5.2, Kimi K3, the new DeepSeek-V4-Flash, and MiniMax-M3, followed today by Nemotron 3.5 Lightning.
But simply making models available is not enough. We are building the capabilities required to serve them without compromising quality, cost, or performance.
Take GLM-5.2 as an example. Some described its release as a “second DeepSeek moment.” Nebius’s implementation achieved a 100% quality score and demonstrated leading performance in independent benchmarks and validation by Artificial Analysis.
This leaves the company very well positioned to capture the market’s growing demand for open models.
Early Progress and Economics of the Asset-Light Business
Moderator: Rob Oliver follows up: Can you discuss the asset-light business’s early progress, engagement with potential partners, and its economics relative to the core business?
CEO: The company is expanding globally, but fundamentally we remain a startup and must be highly disciplined in deciding where to deploy our own capital. We cannot enter every market immediately and simultaneously, so we are very willing to work with partners, and our platform supports that model.
Since announcing the asset-light business model, we have received inquiries from dozens of potential partners. These partners have substantial capacity and ample capital, but lack the expertise to build the technology or sell the capacity.
As GPUs increasingly become an investable asset class, we expect more companies to enter this market, but they will need Nebius’s help to deliver their capacity to customers.
With the right technology platform and go-to-market capabilities, we can help these partners. That is the value proposition of the asset-light business model.
The model remains at a very early stage, but we are actively advancing it and believe it has enormous potential.
2027 Capacity, Pricing, and Revenue Outlook
Moderator: Morgan Stanley’s Ryan Lantz asks: The company has said it could sell out all of its 2027 capacity today. How should investors think about 2027 capacity, pricing, and revenue? When will the company formally issue guidance?
CFO: This is probably one of the questions of greatest interest to investors on this call.
Transactions signed this quarter were priced above $20 million per megawatt per year, implying a payback period of less than two years. These deals will begin coming online late in the fourth quarter and can serve as a benchmark for pricing early next year.
Under current market conditions, the company could sell out all planned capacity today. Its decision not to do so reflects confidence in the future pricing trajectory.
Capacity is clearly another important part of the equation. The capacity planned for deployment in 2027 will significantly exceed the capacity already deployed or still scheduled for deployment in 2026.
The expected trends in both capacity and pricing make us extremely enthusiastic about 2027, although formal guidance will be issued later this year.
In addition, the asset-light business model and high-value services such as agentic and inference solutions are expected to contribute an increasing share of revenue while supporting higher margins.
Year-End Target of 800 MW to 1 GW
Moderator: Brett Knobel of Cantor asks whether the company still expects to reach 800 MW to 1 GW of connected power by year-end.
Andrey: Yes. The company continues to expect 800 MW to 1 GW of connected power this year.
The question also suggested that Vineland may not be completed until 2027. That is not the case: Vineland is part of the company’s 2026 capacity and connected-power plan.
It is important to clarify that connected power represents the power available to the data center. Several steps are still required before that power generates revenue, including data-center commissioning, network construction, cluster buildout, platform deployment, and customer onboarding.
This process takes several months. Depending on GPU-generation transitions, there may be a lag between power connection and revenue generation.
Accordingly, the 800 MW to 1 GW guidance refers to connected power. This capacity is expected to become operational progressively during the first half of 2027.
Vera Rubin Deployment Timeline
Moderator: Stefan Slowinski of BNP Paribas asks whether the third-quarter monetization level of $40 million per megawatt reflects initial pricing for Vera Rubin capacity.
Andrey: Vera Rubin has been running in our labs for some time, and the results to date are in line with expectations.
From a technical standpoint, the transition from Grace Blackwell to Vera Rubin should be easier than the previous transition to Grace Blackwell.
We expect to begin deploying Vera Rubin later this year or early next year and to continue scaling deployments throughout next year.
Capacity Auctions and Short-Term Training Contracts
Moderator: The investor platform also received numerous questions about the company’s new go-to-market models. Mark, after quarter-end, the company piloted capacity auctions and short-term, large-scale training deals. Please discuss the strategy behind these models, the pricing signals they provide, and whether these transactions still include take-or-pay provisions.
Mark: The market is exceptionally active and evolving very quickly. The company is continually looking for ways to identify market signals and deepen its understanding so it can evaluate the take-or-pay business more comprehensively.
These initiatives were introduced primarily to test and validate the market while building customer relationships in a disciplined manner.
The previously mentioned short-term, large-scale training-capacity deals target customers that require dedicated resources at scale. In the current case, the customer needed a GB300 cluster for three to six months and was willing to pay a premium.
These customers typically have highly specific requirements, such as completing time-sensitive, large-scale training before a model launch or running a reinforcement-learning post-training sprint.
Across the model industry, we repeatedly hear that a few weeks can make a material difference, making timely access to reliable, high-performance AI compute equally critical. Customers are therefore willing to pay a premium for speed and certainty over a defined period.
The primary purpose of the auction model is transparent price discovery. In the current environment, market pricing can change between the launch of a sales process and contract signing. The challenge is accurately assessing the fair value of the company’s offering at any given time.
Reference points from competitor pricing, analyst views, and even prediction markets vary widely. In a market with multiple prospective buyers for every GPU, we chose to let the market provide the answer directly.
The winning auction price was 15% above the highest price the company had previously seen and 20% above prices in the Blackwell-related sales pipeline.
The winning customer was highly satisfied with the overall experience, particularly the validated pricing and the certainty of securing the required compute capacity. The customer also indicated plans to participate in future auctions.
Both transaction types validate customer value and are highly profitable in their own right. The company will use only a small portion of its total capacity for price discovery and value validation.
As the company moves into 2027, these initiatives will help us assess both our own capacity and the broader market, with wide-ranging implications for pricing, portfolio strategy, and transaction negotiations. They also have inherently shorter lead times from signing to deployment.
Financing the Gigawatt-Scale Expansion in 2027
Moderator: The company plans to deploy more than 1 GW of incremental capacity annually beginning in 2027. Dado, how does the company intend to finance capital expenditures on this scale, and how does it prioritize the available funding sources?
CFO: The most important point is that the company has multiple pools of capital available to support growth. We are highly confident in our ability to finance capacity deployments in 2027 and beyond.
The first source is operating cash flow. The company is already generating positive operating cash flow and expects it to increase significantly as the business scales.
The second is customer prepayments. Contract terms secured in the second quarter can cover approximately 50% to 60% of the associated capital expenditures. One of the company’s objectives is to raise that coverage ratio further over time.
The third is asset-backed financing supported by long-term contracts with investment-grade customers. These contracts provide a strong foundation for raising capital on attractive terms.
The company currently has approximately $40 billion of committed backlog that can support financing and began using this funding source in the second quarter.
The company can also access corporate-level debt and equity-linked financing, neither of which has yet been used at scale.
As the business expands, we will balance these funding sources in a disciplined manner while maintaining a strong balance sheet.
New financing opportunities are also emerging around GPUs as a standalone asset class. We have seen growing market receptivity and interest in financing GPUs as independent assets and are highly optimistic about the market’s potential.
Over time, this could become another attractive source of financing for the company.
Overall, we are highly confident in the current strength of the balance sheet and believe our diversified financing strategy can support growth in 2027 and beyond while preserving financial discipline.
Token Usage Growth, Token Factory, and Tavily
Moderator: An investor asks how token usage is expanding across the market and whether this is affecting usage and customer adoption of Token Factory and Tavily.
Roman: Let me begin by revisiting the company’s inaugural Inflection event in June. It was likely the first time the company publicly presented its product strategy in a systematic way.
We are building the AI cloud layer by layer to address developers’ needs at every level, from highly scalable bare-metal infrastructure through to the agentic application layer.
The clearest trend is that AI systems are moving into production. Coding is the most visible example, but adoption extends far beyond software development.
We are already seeing long-horizon agentic workflows in financial services, including use cases at Revolut and Mastercard; companies such as Shopify are also using these capabilities to improve e-commerce customer experiences and business processes. Other vertical applications include Sword Health in healthcare, Higg Health in marketing automation, and additional industry-specific use cases.
Nebius has also consistently been the first customer of its own products. In its latest cloud product release, the company introduced Iho, its proprietary infrastructure agent, which runs open models served by Token Factory.
The challenge customers now face is scaling complex systems composed of multiple models, inference engines, and tools.
We address this through a suite of services, including Token Factory for reliable, high-performance inference and post-training capabilities, and Tavily for grounding and access to external knowledge. This is especially important as enterprises move away from closed ecosystems with built-in search toward more open ecosystems.
The Agent AI and Clarify teams are now fully integrated into Token Factory development and have begun delivering features under a unified product roadmap.
We support leading open models from the first day of release and continue to deliver measurable performance optimizations thereafter. Independent benchmarks also continue to rank the company among the leading inference platforms.
The second quarter was also Tavily’s first full quarter as part of Nebius. Its developer community grew from 1 million in February to more than 2.5 million.
Tavily launched search capabilities designed for invocation by autonomous agents and completed a range of certifications required for enterprise deployment.
We are also seeing more customers post-train their own models, creating additional demand for inference and external-information connectivity across the entire development cycle, rather than only in production.
Reinforcement-learning rollouts, evaluations, synthetic-data generation, and training workflows grounded in external information all require substantial inference capacity and reliable access to external information.
These trends validate the company’s vertically integrated platform strategy. The platform can deliver attractive total cost of ownership while supporting a diverse range of workloads.
We aim to support customers across the full AI lifecycle, from training and post-training to inference and external-information connectivity. This is the powerful product proposition created by combining the company’s infrastructure and software, and development is still at an early stage.
New workloads are also placing new demands on physical infrastructure. Because agent orchestration, tool calling, and data preparation are all highly CPU-intensive, the company is expanding Arm and CPU deployments alongside its GPU clusters.