Nebius (Trans): Year-end contracted power target raised to 5 GW

Dolphin Research's takeaways from the $ Nebius.US FY26 Q2 earnings call

I. Core takeaways

1) Full-year guidance reaffirmed.

a) ARR of $7–9 bn; group revenue of $3–3.4 bn.

b) Group Adj. EBITDA margin of ~40%; capex of $20–25 bn.

c) Newly signed mid- and short-term deals this quarter will go live later this year, with no material impact on 2026 revenue guidance. They will clearly impact 2027 and beyond, with formal 2027 guidance due later this year.

2) Key quarterly metrics.

a) Total: group revenue of $582 mn (+454% YoY, +46% QoQ). ARR reached $3.0 bn at end-Jun (+598% YoY; +56% vs. $1.9 bn at end-Mar).

b) Segment: Nebius AI revenue of $575 mn (+514% YoY), accounting for 98% of group revenue. Adj. EBITDA of $286 mn with a 50% margin.

c) Profitability: group Adj. EBITDA of $236 mn (vs. a $21 mn loss a year ago; $129.5 mn in Q1), implying a 41% margin (vs. 32% in Q1). The gap vs. Nebius AI reflects investment in early-stage Avride and TripleTen.

d) Capex: approx. $5.7 bn this quarter, mainly on GPUs and related hardware, plus data center expansion.

3) Funding mix: customer prepayments now the primary source.

a) Customer prepayments hit a record; ~70% of Q2 bookings included prepay. Prepayments in FY26 are set to exceed $9 bn upfront, directly reducing required debt and equity funding.

b) OCF of $2.3 bn in Q2, with cash and cash equivalents of $8.0 bn at quarter-end.

c) Deployed the ATM program in Q2, issuing 12.7 mn Class A shares at a $224 WAP, raising ~$2.8 bn. As of Jun 30, 12.3 mn shares remained under the program, which the company views as a flexible tool rather than a commitment.

4) First asset-backed financing completed; $40 bn backlog underpins future borrowings.

a) In Jul, closed a $775 mn asset-backed facility secured by deployed GPU infrastructure and cash flows from an investment-grade client's contract. Priced at SOFR + 250 bps, implying a mid-single-digit current rate.

b) Based on strategic contracts with investment-grade clients, there are >$40 bn of similar commitments available for like-for-like financing. The company has minimal HoldCo debt and views this as additional accessible capital.

c) More asset-backed deals are in progress, while HoldCo debt and equity-linked options remain under evaluation.

5) Margin improvement timeline

a) Q2 margin uplift was driven by revenue scale, the asset-light model, Token Factory, and early contributions from recent acquisitions.

b) With pricing visibility, self-built DC capacity coming online from H2 2027 is expected to lift margins.

c) The asset-light approach and higher-value services such as agentic and inference should account for a growing revenue mix, supporting higher margins.

II. Call details

2.1 Management highlights

1) Business model: build capacity first, then decide whom to sell to.

a) The core playbook is to build capacity ahead of signing, using a multi-tenant cloud and a proprietary software stack to serve AI-native companies, agentic AI leaders, next-gen labs, and some of the most mature enterprises. The company controls timing, buyer selection, terms, and financing.

b) The first deal type comprises mid-term core AI cloud contracts of 1–3 years. Four milestone deals were signed this quarter with Reflection, Cohere, a scaled U.S. next-gen lab, and a large U.S. quant firm, each averaging >$1 bn, with $20–25 mn revenue per MW and prepayments covering 50–60% of associated capex.

c) Under these terms, the company could sell out all 2027 capacity today but chooses not to, believing that keeping capacity for shorter-dated, immediate needs yields higher value.

d) The second type is short-term capacity, typically up to six months, for immediate, time-constrained, high-spec customers willing to pay a significant premium. Pricing under discussion is $40–50 mn per MW, and sometimes higher, with one such deal recently signed.

e) The third type involves long-term contracts with investment-grade clients to accelerate and improve construction financing; the Jul secured financing was based on one such contract. More will follow on the back of the $40 bn backlog.

2) Capacity: year-end signed power target raised to 5 GW.

a) The company is expanding its future pipeline via self-build and colocation, positioning itself as one of the few globally able to add >1 GW per year, with a 2027 goal.

b) Almost all signed power is set to come online over the next 2–3.5 years. The path is multi-region expansion, combining grid power and behind-the-meter generation, with rights for several hundred MW of behind-the-meter capacity being pursued.

c) The Bloom partnership will unlock and accelerate multiple sites.

d) Most contracts are cloud contracts, giving the company latitude to deliver within a region and reduce reliance on any single site. Overprovisioning sites and capacity is the top priority now and going forward, with a bias to deploy more and build earlier.

3) Two new commercialization innovations

a) First capacity auction: clearing price was the highest seen on the Blackwell-generation chips, 15% above the company’s prior peak. This provided a real-time market signal for that capacity.

b) Asset-light partnership model: partners fund, build, and operate, while Nebius provides the full-stack platform and demand, layering value-added services on partner infra. This yields high-GP revenue with minimal balance-sheet use and targets the industry’s capital and capacity constraints, with potential to unlock incremental capacity in 2027+.

4) Revenue drivers and near-term cadence.

a) Q2 growth came from Q1 capacity adds, higher utilization, and high-GP contributions from the asset-light model, Token Factory, and recent acquisitions. Utilization improved on underlying infrastructure efficiency gains.

b) Capacity sold out again — sell-through matches the pace of deployment.

c) Capacity deployment will accelerate in H2, with late-Q2 deployments expected to contribute from Q3. Capacity under construction is for 2027 demand and is supported by customer commitments, providing near-term visibility.

2.2 Q&A

Q: The Vineland, NJ data center hearing was adjourned without a vote. Any impact on the ramp?

A: Data centers in the U.S. are part of a broader public debate, which the company is closely monitoring. The approach for new regions is early engagement, transparent communication, and direct responses to community questions, which also applies to Vineland.

Specifically for Vineland, overall delivery remains on plan. The current public hearing concerns final approval of a site plan amendment already approved, reflecting the decision to switch the project’s power to Bloom. From the community’s perspective, Bloom is a clear upgrade — on-site power, reliable, quiet, and very low emissions.

Such hearings are a normal process and built into the schedule; the company believes the plan complies with all applicable local, state, and federal rules, and expects approval shortly after public comments conclude.

On construction, all contracted batches due to date have been delivered, with good reasons to expect the remaining batches to be delivered per contract. The main structure was completed earlier this summer, interiors are progressing well, and Bloom fuel cells should deploy quickly. The pivot to Bloom is valuable and is not expected to materially impact the timeline.

Q: How were the four milestone deals won, and why did clients choose Nebius?

A: Billion-dollar-scale core AI cloud wins mark a major go-to-market milestone, validating the ability to build a diversified customer base at scale. All four were competitive wins, with scale, performance, and reliability as key differentiators; each client sees Nebius as a long-term partner.

Each customer already had suppliers, including hyperscalers, and was seeking a next-gen capacity and platform partner for expansion. In one case, a strategic partner introduced the customer late in Q1 with a need for large, continuous GB300 clusters, plus future expansion flexibility, strong technical support, and a long-term strategic relationship.

At closing in May, differentiators cited included speed of response, transparency, white-glove support, and the ability to meet immediate U.S. deployment while supporting future sovereign expansion. The company is in talks to add capacity for this customer, which is also evaluating Token Factory for inference.

These were outbound, not walk-in orders, with multiple engagements and hands-on POCs validating the tech — one client called it their best POC ever. The company is discussing post-training and inference with several of them to support their revenue plans and growth, and is in talks with all four for significant capacity additions and next-gen chips (incl. Vera Rubin).

Similar deals are in the pipeline, which stepped up further in Q2, with multiple >$1 bn opportunities across AI-native firms, next-gen labs, and enterprises.

Q: Many contracts tie to capacity arriving in late 2026–2027. With skepticism around GW-scale DC builds, where does confidence come from, and what is the timeline?

A: Signed power progressed significantly this year, exceeding year-end targets far ahead of plan, so the 2026 year-end target is now raised to 5 GW — and momentum continues. Nearly all signed power will come online within the next 2–3.5 years.

The execution path is multi-region expansion and a mix of grid power and behind-the-meter generation, with several hundred MW of behind-the-meter rights being secured. The Bloom partnership will unlock and accelerate many sites.

Also, most contracts are cloud contracts, allowing flexible delivery within regions and reducing reliance on any one site. Overprovisioning sites and capacity is the top priority now and in the future, with a strategy to deploy as much as possible as early as possible.

Q: With a fast-moving market and rising prices, and plans to add >1 GW annually plus new initiatives, how does this fit the long-term strategy?

A: The market is moving even faster than expected, and the business model and platform allow rapid adaptation. Two examples emerged this quarter.

First, prices are rising quickly, and because the model builds ahead of selling, the company has available capacity it can allocate to shorter-dated, much higher-margin contracts. Second, the first auction targeted mid-term capacity.

This was possible because of unallocated free capacity and the multi-tenant platform, and the company is now monetizing the benefits. On the capacity side, similar flexibility is a strategic advantage: the platform is general-purpose and can run on any third-party capacity, enabling faster growth — the essence of the asset-light model.

There are multiple monetization paths for full-stack platform generality across software and hardware. The market will keep moving and growing quickly, and the plan is to preserve model and platform flexibility to capture that growth.

Q: With bond market volatility and higher costs, after the $775 mn ABS, will the company keep using debt, or should we expect more ATM or converts?

A: The approach is not to predict markets but to match the right instruments to the right assets, while staying disciplined on three things: cost of capital, minimizing dilution, and maintaining a healthy balance sheet.

Primary funding remains customer prepayments and OCF, with upfront prepayments expected to exceed $9 bn in 2026, directly reducing external funding needs. Asset-backed financing is a key strategic pillar.

The $775 mn raised in Jul at SOFR + 250 bps, secured by deployed GPUs and an investment-grade contract, shows strong demand to finance these cash flows on attractive terms even in volatile markets. With >$40 bn of committed backlog, this is a highly scalable, repeatable model that will be reused as new capacity is deployed.

Beyond that, there is meaningful flexibility: the company has minimal HoldCo debt, a potential incremental capital source. Equity and equity-linked (ATM or potential converts) remain additional tools under consideration.

The company is actively evaluating further equity-linked and asset-backed structures and continues to assess HoldCo debt and other alternatives, optimizing across sources for cost of capital, dilution, and balance sheet discipline.

Q: How will 2027 capacity be allocated between short-term and multi-year deals, and is there a per-MW ACV threshold for long-term deals?

A: Mid-term contracts are largely core AI cloud agreements with strong unit economics with the world’s most ambitious AI companies. Short-term opportunities are premium, including auctions and short-duration scale contracts.

In practice, the company optimizes across customer type, price, payment structure, term, and size — not a single metric or composite. The current prioritization is existing customers first, then new customers, then terms; within terms, price, prepay ratio, and then duration.

Thus, the lens is broader than a simple per-MW ACV. Tactically, the company has shortened lead times for forward selling, moving sales closer to deployment to raise achievable pricing while preserving flexibility.

It deliberately reserves some capacity for short-term and immediate demand where total value capture is currently highest. As capacity deploys from H2 this year into next year, allocation between mid-term and short-term will follow the above customer- and terms-first framework.

Q: xAI has started selling compute at a premium. What does that imply for market pricing, and are you seeing similar strength in new deals and renewals?

A: This does not change Nebius’s plans. The company essentially competes with the three hyperscaler clouds in the same market, which is scaling from hundreds of billions per year toward ~$1 tn, if not more.

Nebius aims to build 1 GW per year — few can do this — and while that looks large, the market is adding ~10 GW per year. Hyperscalers are building much of it, but even they cannot build it all; the remainder needs to be built by others.

That is Nebius’s lane and its biggest opportunity. New entrants validate, rather than alter, the market Nebius addresses.

Q: Is your open-weights positioning driving client-side inference demand?

A: The company believes customers, enterprises, and society benefit from competition and diversity, and has supported an open AI ecosystem and infrastructure from day one. This gives customers flexibility, control over data and models, and freedom in deployment.

At scale, economics become critical. Frontier adopters are scrutinizing token cost and whether AI can solve tasks at scale-viable economics.

Another key question is how to extract and translate proprietary domain knowledge and data into better-performing AI — data is the moat, possibly the only moat. These factors push customers toward more specialized models.

The value of open models is not just availability but the ability to fine-tune, train, and optimize for specific customers, businesses, and use cases. Quality of open models is improving rapidly, and the industry is converging toward the flexibility and control the company built for from the start.

Token Factory offers day-zero support for frontier open models with rapid cadence: in just the past four weeks, Nemotron Ultra, GLM 5.2, Kimi K3, the new DeepSeek Flash, and MiniMax 3 went live, and Nemotron Lightning is out today. Availability alone is not enough, and the focus is serving these models without sacrificing quality, cost, or performance.

For GLM 5.2, some called its release a 'second DeepSeek moment'. Nebius’s implementation achieved 100% quality scores with leading performance, validated by independent benchmarks and Artificial Analysis, positioning the company well for current open-source demand.

Q: Early traction for asset-light offerings and unit economics vs. core?

A: The company is expanding globally but remains a startup that must deploy its own capital carefully and cannot enter all markets at once. It is therefore very open to partnerships, and the platform supports this.

Since announcing the asset-light model, dozens of potential partners have reached out — they have capacity and capital but do not know how to build or sell. As GPUs become an investable asset class, more entrants are expected, but they need Nebius to deliver capacity to customers.

The company has the right tech platform and routes to market to help, though the model remains very early.

Q: If you could sell all 2027 capacity today, how should investors frame 2027 by capacity, pricing, and revenue, and when will you guide?

A: This quarter’s deals imply >$20 mn per MW and paybacks under two years, with go-lives starting late Q4, offering a baseline for early next year’s pricing. The decision not to sell out reflects confidence in future pricing.

Capacity is the other key variable: 2027 deployed capacity will be significantly higher than 2026 deployed plus what will be deployed for the rest of 2026. Both capacity and pricing trends point to 2027, with formal guidance later this year.

Moreover, the asset-light model and higher-value services like agentic and inference should contribute a growing share of revenue and support higher margins.

Q: Do you still expect 800 MW–1 GW of connected power by year-end, and is Vineland in 2027 capacity?

A: Yes, the 800 MW–1 GW connected power target for this year remains in place. Vineland is part of 2026 capacity and connected power, not 2027.

Note that connected power refers to the data center itself. From there to revenue, there are multiple steps: DC commissioning, networking, cluster buildout, platform deployment, then customer onboarding before revenue starts.

This process takes months and depends on GPU generation changes, so there is a lag between connected power and revenue. The 800 MW–1 GW is connected power, with actual production load roughly in H1 2027.

Q: Does the ~$40 mn per MW monetization level in Q3 reflect initial Vera Rubin pricing, and how will the ramp unfold?

A: Vera Rubin has been running in the company’s labs for some time and is delivering expected results. The step from Grace Blackwell to Vera Rubin is technically easier than prior generational transitions to Grace Blackwell.

Deployments are expected to start late this year or early next year and continue throughout next year.

Q: Strategy behind the capacity auction pilot and short-term scale training deals since quarter-end, implications for pricing, and are they still take-or-pay?

A: The market is extremely dynamic, and the company constantly seeks better signals to understand its take-or-pay business. These two initiatives are first about learning and validation, while building relationships with discipline.

The short-term scale training deal refers to a customer needing a dedicated large cluster (GB300 in this case) for 3–6 months at a premium. Such customers have specific needs, e.g., large-scale pre-launch training windows or post-RL training sprints.

In this domain, 'weeks matter'. Reliable, high-performance AI compute availability matters too, and customers will pay for speed and certainty over a defined period.

The auction was purely about price discovery. In the current environment, prices can move before a sales cycle completes, and the challenge is to hold a reliable view of fair value at any point in time while competitor pricing, analyst takes, and market forecasts are dispersed.

In a market where each GPU has multiple buyers, the company chose to let the market set the price. The clearing price was 15% above the company’s prior high and 20% above Blackwell pipeline pricing.

The winner was very satisfied, especially with validated pricing plus certainty of capacity, and plans to participate in future rounds. Conclusion: both initiatives validate customer value and are profitable on their own.

The company is using a small slice of total capacity to drive price discovery and value validation, which informs capacity and market views into 2027. Both will influence pricing, packaging, and negotiations and have faster deployment and closing cycles.

Q: How will you finance capex to add >1 GW per year from 2027, and what is the funding stack?

A: Multiple funding pools support growth, and the company is confident about financing capacity planned for 2027 and beyond. First, OCF is already positive and should grow materially with scale.

Second, customer prepayments covered ~50–60% of associated capex in Q2, with a goal to raise that coverage. Third, long-term IG-backed contracts underpin attractive asset-backed financing, with ~$40 bn of commitments available; Q2 marked the start of tapping this source.

HoldCo debt and equity-linked financing provide further flexibility and are largely untapped. The company will balance these sources with discipline to maintain a robust balance sheet during expansion.

There are also emerging financing opportunities as GPUs become a financeable asset class in their own right, which could be another attractive long-term capital source.

Q: What is the impact of 'token maxing' on Token Factory and Tavily usage and adoption?

A: At its first Inflection event in Jun, the company outlined its product strategy — building the AI cloud layer by layer from large-scale bare-metal infra to the agentic layer, meeting developers wherever needed.

A clear trend is that AI systems are moving into production. Coding is the most visible example, but not the only one: in financial services there are longer-cycle agentic workflows like Revolut and Mastercard use cases; in e-comm, Shopify is improving CX and processes; there are also healthcare cases like Sword Health and verticals such as marketing automation.

The company is its own first customer: its infra agent Echo, launched in the last cloud release, runs open models served by Token Factory. Customers struggle to scale complex systems that combine multiple models, inference engines, and tools, which the company addresses with a suite of services.

Token Factory delivers reliable, high-performance inference and post-training, while Tavily delivers grounding, especially as enterprises move from closed, search-bundled ecosystems to more open stacks. Eigen AI and Clarifai teams are fully integrated into Token Factory and delivering to plan.

The company provides day-zero support for major open models, with measurable performance optimizations post-release, and independent benchmarks consistently rank it among leading inference platforms. Q2 was Tavily’s first full quarter within Nebius.

Its developer community grew from 1 mn in Feb to >2.5 mn, and it launched keyless paper search built for autonomous agent calls, while securing enterprise deployment certifications. Increasingly, customers are post-training their models, adding inference and grounding needs across the lifecycle, not just in production.

RL, rollout, evals, synthetic data generation, and grounding-related training all require heavy inference and reliable access to external information. These trends validate the vertically integrated platform strategy, offering attractive TCO and supporting diverse workloads across the full AI lifecycle — training, post-training, inference, and grounding.

New workloads also change physical infra needs: beyond new GPU generations, agentic orchestration, tool use, and data prep are CPU-intensive, so the company is adding ARM and CPU deployments alongside GPU clusters.

Risk disclosure and statements:Dolphin Research Disclaimer and General Disclosures