Nebius Group N.V. (NBIS) Earnings

Management highlights

- Business Model & Deal Strategy * The company follows a strategy of building capacity ahead of customer contracts, using a multi-tenant cloud software stack to serve AI natives, agentic AI leaders, independent research labs, and large enterprises, retaining flexibility to choose customers, timing, and deal terms. * Three core deal types: 1) 1-3 year mid-term core AI cloud contracts, with 4 closed landmark deals this quarter averaging over $1 billion each at 20-25 million dollars per megawatt, with 50-60% of capex covered by upfront payments; 2) up to 6 month short-term contracts for immediate high-value customer needs, currently negotiated at 40-50+ million dollars per megawatt, commanding a significant premium; 3) long-term contracts with investment-grade customers to enable faster, lower-cost financing for capacity expansion, supporting $40 billion in existing contracted backlog. * Management intentionally chose not to sell all 2027 capacity today, retaining unallocated capacity to capture higher premiums from short-term deals. - Go-To-Market & Product Innovation * The company held its first successful capacity auction for Blackwell generation chips, clearing 15% above the highest price the company had previously charged, providing clear real-time market price discovery. * Token Factory, the company's inference service for open AI models, offers day-zero support for all new frontier open model releases, with independent benchmarks confirming top-tier performance and quality. The associated developer community for the Tavuli grounding service grew to 2.5 million developers, up from 1 million in February 2026. - Capacity & Construction Updates * Raised 2026 year-end contracted power target to 5 gigawatts, becoming one of the only companies globally capable of adding over 1 gigawatt of new capacity annually, with plans to hit this annual build target in 2027. * Introduced a new asset-light partnership model: third-party partners finance, build, and operate facilities, while Nebius provides its full-stack cloud platform, customer demand, and value-added services, generating high-margin revenue with minimal balance sheet capital exposure, unlocking additional capacity for 2027 and beyond. * Construction of the Vineland, New Jersey data center building is complete, with engineering fit-out progressing well. A delayed public hearing for a site layout amendment (required after switching to on-site Bloom fuel cell power) is part of the normal planning process, with no expected material impact to the project timeline. - Financial & Capital Structure * Triple-digit revenue and ARR growth was achieved in Q2 2026, even before most 2026 contracted capacity comes online in the second half of the year. Approximately 70% of all Q2 closed deals include customer upfront prepayments, with over $9 billion in total 2026 upfront prepayments expected, reducing the need for external debt and equity financing. * In Q2 2026, the company issued 12.7 million Class A shares via its at-the-market (ATM) equity program for $2.8 billion in gross proceeds, with 12.3 million shares still remaining available under the program. In July 2026, the company closed its first $775 million asset-backed debt facility, secured by contracted cash flows, priced at a modest spread over benchmark rates with a mid-single-digit effective rate.

Guidance

- Full-year 2026 guidance is reaffirmed across all metrics, unchanged from prior projections: * Expected annualized run rate revenue of $7 to $9 billion * Expected full-year group revenue of $3 to $3.4 billion * Expected full-year group adjusted EBITDA margin of approximately 40% * Expected full-year capital expenditures of $20 to $25 billion * Confirmed 2026 full-year connected power guidance of 800 megawatts to 1 gigawatt, which includes the Vineland New Jersey site - Vera Rubin GPUs are already testing in the company's labs, with commercial deployment starting in late 2026 or early 2027, ramping through the full year 2027 - Formal 2027 guidance will be released later in 2026; baseline 2027 pricing is expected to start at over $20 million per megawatt, with management confident that pricing will continue to improve. 2027 capacity deployment will be significantly higher than 2026 levels, with high-margin asset-light revenue and value-added services expected to comprise an increasing share of total revenue

Segment performance

Nebius Group reported total Q2 2026 revenue of $582 million, representing 454% year-over-year (YoY) growth and 46% quarter-over-quarter (QoQ) growth. Nebius AI business revenue grew 514% YoY to $575 million, accounting for 98% of total group revenue. Annualized run rate revenue reached $3 billion at the end of Q2 2026, up 598% YoY and 56% QoQ from $1.9 billion at the end of Q1 2026. Group adjusted EBITDA was $236 million, up from a $21 million loss YoY and $129.5 million in Q1 2026, with an adjusted EBITDA margin of 41% (up from 32% in Q1). Nebius AI business generated adjusted EBITDA of $286 million with a 50% margin; the gap between group and segment margin comes from early-stage investments in AVRIDE and Triple10. Capital expenditures for Q2 were approximately $5.7 billion, driven primarily by GPU purchases, related hardware, and data center expansion. Ending cash and cash equivalents stood at $8 billion, with customer prepayments reaching an all-time high in the quarter.

Risks & headwinds

- There is ongoing public and regulatory debate around data center development in the U.S., which can create delays in local permitting and approval processes for new sites, such as the delayed public vote for the Vineland New Jersey site layout amendment. - GPU supply and gigawatt-scale data center construction present execution risks, including potential gaps between connected power completion and revenue generation, which can take several months as infrastructure is commissioned, clustered, and customers are onboarded. - Debt market volatility has pushed overall borrowing costs higher, creating uncertainty around the cost and availability of external financing for large-scale capex plans. - The AI cloud pricing market is highly dynamic, with prices changing rapidly during sales cycles, creating uncertainty around future revenue and margin projections.

Analyst Q&A

• Q: What is the status of the Vineland New Jersey data center, and what impact could the delayed public vote have on the project timeline? /

  A: Building construction for the site is complete, and required pre-approval trenching work has been delivered on schedule. The delayed public hearing is for a site layout amendment needed after switching to on-site Bloom power, which management says improves the project for the local community by delivering ultra-low-emission reliable power. The team remains confident the amendment complies with all regulations, expects approval quickly, and notes regulatory hearings are always factored into project timelines, with no expected material impact on delivery. (303 characters)

• Q: Can you confirm how you will allocate 2027 capacity between short-term premium deals and mid-term core contracts, and do you have a minimum ACV per megawatt threshold for longer-dated deals? /

  A: Management does not allocate based on a fixed percentage or single ACV metric, instead optimizing across customer priority, price, prepayment terms, and deal structure. Priority is given to existing customers first, then new logos, followed by deal terms prioritizing price, upfront prepayment, and then duration. The company has shortened its sales cycle to sell closer to deployment to capture higher pricing and retain agility, and deliberately allocates a portion of capacity to high-premium short-term deals to maximize total realized value. (399 characters)

• Q: XAI has entered the AI capacity market at premium pricing; what does this mean for market dynamics and Nebius' opportunity? /

  A: Arkady noted the entry of new players simply validates the massive size of the fast-growing AI cloud market, which is projected to grow from hundreds of billions to over $1 trillion annually. While hyperscalers will continue to grow with the market, they cannot meet all demand, leaving a large opening for independent providers like Nebius. The market is growing by tens of gigawatts of new capacity annually, so Nebius' plan to build 1 gigawatt of new capacity per year fits easily into the available market opportunity, with no change to the company's strategy. (377 characters)

• Q: How do you plan to fund 1+ gigawatt of annual capacity build starting in 2027, and how do you prioritize funding sources? /

  A: The company has a diversified pipeline of funding sources and is fully confident in its ability to finance planned growth. The first priority is operating cash flow, which is already positive and expected to grow meaningfully as the business scales, followed by customer prepayments (which already cover 50-60% of capex for recent deals, with a target to increase this coverage further). The $40 billion investment-grade contracted backlog enables scalable low-cost asset-backed financing, and the company has not yet meaningfully tapped corporate debt or additional equity markets, providing extra flexibility. Emerging GPU as an asset class financing also opens additional new funding avenues. (416 characters)

• Q: What is the strategy behind new go-to-market initiatives like capacity auctions and short-term premium deals, and what did they reveal about market pricing? /

  A: These initiatives are designed to drive direct price discovery in a fast-moving market where standard reference pricing is highly inconsistent, while also building customer relationships. Short-term deals serve customers needing dedicated capacity for time-boxed high-priority projects like pre-release model training, who will pay a large premium for speed and certainty. The first auction cleared at 15% above the company's previous highest Blackwell price and 20% above average pipeline pricing. Using a small portion of capacity for these initiatives provides actionable pricing data that improves overall pricing strategy for all deals, with faster close and deployment cycles. (408 characters)