Nebius (Trans): Prepayments fully cover capex; revenue to quadruple next year

The following is Dolphin Research's edited Trans of $Nebius(NBIS.US) FY26 Q3 earnings call

I. Core results recap

1. FY26 guidance: some third-party capacity shifts to Q2 2027. The change is about timing, not demand.

a. Revenue: $3.0–3.2bn. Previously $3.0–3.4bn.

b. ARR: Feb ARR expected to exceed $7.0bn. Thereafter, the company will stop guiding on ARR and focus on revenue and pre-tax profit, as ARR is a single-month metric and can be distorted by asset-light revenue.

c. Margin and capex: Adj. EBITDA margin ~40%, unchanged. Capex raised to ~$27.0bn (prev. $20–25bn); contracted power by year-end still 5GW, now above 4GW.

2. Initial FY27 outlook: more conservative timing assumptions on site deliveries. A stricter stance on delivery schedules across sites.

a. Revenue: ~4x growth. Reaching $12.0–13.0bn.

b. Profitability: expects pre-tax breakeven in Q4. Servers and network equipment depreciated over 5 years.

c. Capex: $40–45bn. Implies >1GW of new capacity, mostly from higher-margin owned sites.

3. Funding sources: 75%–85% of FY27 capex expected to be covered. Via prepayments and asset-backed financing supported by existing investment-grade contracts.

a. Prepayments: in Q3, over 90% of signed TCV included prepay, with the prepay mix up ~20ppt QoQ. FY26 committed prepay about $10.0bn; FY27 expected at $17–20bn, and that is not a cap.

b. Debt financing: completed first $775mn ABS in Jul (SOFR + 250bps), with similar investment-grade committed revenue exceeding $40bn. Issued $5.75bn convertibles in Aug.

c. Equity and cash: in Q3, sold 4.6mn shares via ATM at an Avg. price of $229.1, net proceeds approx. $1.0bn, with 7.6mn shares remaining. Cash at quarter-end was $7.4bn.

4. Key quarterly metrics

a. Margins: Adj. EBITDA margin 47.5% (Q2: 40.6%). AI cloud at 53.2% (Q2: 50%).

b. One-off revenue and capex: recognized approx. $325mn of non-recurring asset-light revenue. Q3 capex was ~ $7.3bn.

II. Earnings call details

2.1 Management remarks - key points

1. Demand and pricing: customers are signing longer-term, higher-priced, more front-loaded contracts. The demand base continues to broaden.

a. Unit price: Q3 contracts priced above $25mn per MW. Up from $20–25mn in Q2.

b. Contract size: after first disclosing $1bn+ deals last quarter, several multi-billion-dollar AI cloud agreements are already in signing for Q4. Enterprise clients such as Databricks, Datadog, and Snowflake signed multi-year deals for the first time this quarter.

c. Three-track sales: spot and short-term contracts carry a premium; 1–3 year deals are the core; a smaller set of 5-year contracts helps diversify tenor and support financing. The latter two are covered by prepayments.

2. Prepay and payback: customers prepay to lock capacity and current pricing, shortening payback further. Payback dynamics continue to improve.

a. Payback: shortened to 1 year 7 months in Q3. Versus 1 year 10 months in Q2.

b. Prepay cadence: customers are paying now for capacity that ramps within 12 months. FY26 prepayments will exceed $10.0bn, and will increase meaningfully next year.

3. Cost and margins: premium pricing plus low all-in cost from full-stack self-build underpin margins. The model blends market-led pricing tools with structural cost advantages.

a. Pricing tools: launched reserved-capacity options, public spot prices, and short-term capacity. Pricing is aligned with market discovery.

b. Low cost: racks and storage are self-designed and self-built, placing the cost structure among the lowest in the market. As owned sites mix rises, costs should fall further, aided by lower financing costs from investment-grade contracts.

c. Token Factory: the managed inference platform is on track to reach several hundred million dollars of ARR. It runs on owned GPUs rather than leased capacity, supporting better margins.

4. Capacity delivery: some colocation sites delayed; owned sites are on schedule or ahead. The next GW will mainly come from owned sites.

a. Delay impact: certain colocation sites such as New Jersey are behind schedule, with no capacity canceled; some shifted to 2027, and some planned revenue slips to early next year. Timing is the issue, not demand.

b. Owned-site edge: the in-house team handles land, power, shells, hardware, software, and ops end-to-end. The industry will eventually rebalance supply and demand, and the company is preparing for that.

c. Outlook: under a conservative build-pace assumption, revenue can quadruple next year. Pre-tax profitability is expected in Q4 2027.

2.2 Q&A

Q: If more capital is needed beyond current cash, how much can prepayments contribute given parts of the debt market today?

A: Yes; if needed, prepayments can top up incremental cash needs. Since the two landmark Microsoft and Meta deals with large prepays, prepayment has been a core feature of the biz.; AI cloud began collecting prepays in Q1 this year, and over 90% of Q3 signed TCV included prepay.

Prepay is possible for three reasons: the platform is viewed as market-leading, so clients pay a premium for reliability and service; last quarter we could have sold out capacity under then-current terms but chose to hold back for better terms, enabling both higher pricing and larger prepays now. Third, the cost base is low, so the same prepay translates into a higher share of capex coverage, exceeding 100% this quarter (incl. data center costs).

The COO added: when allocating capacity across contract types, we can factor in capital market conditions. When corporate bond yields are high, allocate more to long-term contracts and collect more prepay to cover capex; when markets improve, opportunistically use debt and keep more capacity for short-term and spot.

Q: How does the company work with state and local governments and communities in places like Texas? Can that ease deployment concerns?

A: No business in Texas currently; across roughly six active states, we do not expect political shifts to impact timelines. Skepticism about AI and data centers exists across both parties, and many concerns are reasonable and should be addressed; we expect no material change in sentiment post-election. The company engages openly with state leaders and communities to explain what we do and the benefits for local communities.

Q: What is the financing strategy for FY27? How much incremental funding is needed to support the current backlog and data center builds?

A: No specific figure was given for incremental needs, while reiterating that prepayments cover over 100% of capex. Funding rests on two pillars: prepayments cover all capex and arrive well before capacity is online; we also hold tens of billions of dollars in investment-grade contracts that support attractive ABS terms.

Core capacity can be allocated between short and long tenors depending on funding conditions. When capital market terms are poor, do more long-term, and less spot or short-term; when we can finance on good terms, raise opportunistically and allocate more to short-term and spot.

Q: What is the root cause of delivery delays? The shareholder letter mentioned multiple colocation sites slipping; what else is affected?

A: At Vineland, NJ, the main issue was early removal of temporary power while Bloom ramp had gaps. When shifting to Bloom in summer, the plan was to bridge with temporary generation from DataOne, but temp power was reduced early at end-Q3. We are accelerating Bloom deployment, but there will still be a gap; we continued delivering new batches to Microsoft in Q3, and will keep delivering in Nov–Dec as power comes online. Bloom is already energizing the site; it just needs to ramp.

The CEO added: the site is third-party built, with more potential failure points. After learning in Aug that the diesel plan would not work, we switched to Bloom, causing roughly a two-month delay, and we are now back on track. We plan to use more Bloom generation at other sites, and the new setup may add extra power to this site; even with the delay, revenue remains within the guided range. No other affected sites were named.

Q: How does demand for short-term vs. long-term dedicated capacity compare to TAM? How do you balance the mix?

A: The mix is market-driven, dynamically adjusted via auctions, spot pricing, and long-term demand dialogues. We position to meet client demand rather than to make a market, using systems that optimize between short-term burn and long-term reservation based on opportunity. Similar to prepay, we switch between optimizing near-term revenue and locking future capex for builds.

The long-term vision is a standardized trading platform with algorithmic pricing for all segments, from self-serve SMBs to large enterprises. No TAM split was quantified for the two contract types.

Q: What are the key assumptions behind the initial FY27 revenue guide?

A: The $12.0–13.0bn guide bakes in very conservative delivery timing assumptions. Hitting it requires capacity to come online on time and to be sold. On supply, a large share of FY27 capacity will be owned rather than third-party, offering higher delivery control; if on-time, revenue could exceed the guide.

On demand, customers are already seeking FY28 capacity, and FY27 pricing is moving higher. Management sees no risk on that front.

Q: Since Q3, customers committed TCV of over $12.0bn, yet RPO rose only 3%. How much TCV is not yet in RPO, and what limits reported backlog growth?

A: Several agreements were signed post-quarter-end and will appear in next quarter's disclosed commitments. For multiple quarters we have said we could sell out current and even future capacity at current prices, with the open question being terms. Through auctions and spot pricing to discover price, plus LT demand dialogues, we raised prices on a newly released tranche in early Sep, while also lifting prepay and extending durations.

Since then, Q3-to-date customer commitments exceeded $12.0bn in TCV, including multi-billion 5-year deals. Following the four hallmark deals in Q2, similar clients expanded with longer terms, higher prices, and larger prepays, and this momentum has continued into Q4. No figure was given for TCV not yet reflected in RPO.

Q: OpenRouter said in Sep that open-weight models accounted for 72% of its weekly token volume. How do you see open vs. closed models long term, and is the shift to open-weight a positive for the company?

A: Adoption of open models has accelerated in recent months, and greater model diversity is positive for us. From Q3 into early Q4, we supported more models as a launch partner (incl. NVIDIA and others). Enterprises use open models to scale AI with viable economics, and enhance quality with proprietary data and domain knowledge.

Diversification broadens the range of addressable customers and partners. No share split between open and closed was provided.

Q: What are the main constraints between contracted power and deliverable compute? What milestones should investors track, and has communication on timing changed?

A: Power and timing are the main challenges; most FY27 deliveries will be from owned sites. Owned projects are developed end-to-end in-house with much greater certainty than colo: Alabama, Missouri, and Finland will contribute revenue in FY27, with Pennsylvania energizing and contributing later.

Another common bottleneck is chips for GPUs, CPUs, networks, and storage, which have not constrained us historically and we do not expect in FY27. Next-gen chips for FY27 ramps have completed integration, testing, and benchmarking, and FY27 supply is fully secured. No specific milestones were provided.

Q: How should we interpret the $325mn of non-recurring revenue recognized in Q3? Why one-off under the asset-light model, and how will it be executed going forward?

A: This was the first asset-light transaction; at scale it should become more recurring and roll into the base model. Under asset-light, partners fund capacity built on our architecture: some revenue is recognized upon locking capacity, supply chain, and design, and some when capacity goes live and is sold to customers.

The FY27 guide includes no asset-light or other partnership revenue, and even with most capacity being owned, it still embeds substantial delay assumptions. If deliveries are on time or partnership opportunities materialize, there could be upside to the guide.

Q: What is the mix of enterprise, AI-native, and other customers among new clients?

A: No split was provided by customer type, only that enterprise demand is accelerating but far from fully unlocked. A leading indicator is the adoption of our services by the data and infra platforms enterprises rely on, and Databricks, Datadog, Snowflake, and Palantir have signed multi-year deals.

Palantir chose us as its preferred AI cloud, integrating Token Factory and AI cloud with Foundry. Together with open models, this serves large enterprises' sovereign AI needs; we are seeking similar partnerships with other platform companies.

At the same time, leading enterprises are also adopting directly, establishing their first non-hyperscaler cloud relationships. Management expects to discuss more enterprise activity in the coming quarters.

Q: What drove Adj. EBITDA margin to rise from ~41% in Q2 to 47.5%? If full-year is ~40%, does that imply Q4 steps down to the mid-30s, and how do you view that?

A: Quarterly margins fluctuate with timing mismatches; full-year remains guided at ~40%, and we do not guide quarterly. New sites bear rent, power, and staffing costs before revenue starts, and Q3 included some time-based revenue that will not recur in the same form.

Q4 includes a full quarter for newly phased colo capacity and Nov deliveries at the NJ site, while also absorbing site costs ahead of revenue. FY27 margin direction is up: owned sites come online in H2 with no third-party rent, and the higher-price, larger-prepay agreements since Q3 convert to revenue.

Q: Within the $12.0bn TCV signed in Q3, how much was from AI-native vs. large enterprises?

A: The vast majority came from AI-native customers, mostly existing clients already on the platform. These clients typically entered via auctions or deployments for other workloads, share the trait of operating at scale, and are growing revenue rapidly. No specific split was provided.

Q: What updates are there in the software stack, and how does it drive margins and moats?

A: Token Factory is growing rapidly, with owned infra delivering lower costs vs. renting GPUs. Token Factory is priced per token and sold as a managed service, with SLAs distinct from GPU infra; it supports new models, advanced routing and caching, custom kernels, and other inference optimizations, and provides post-training capabilities. We do not manage margins by product.

From Q3 into early Q4, Token Factory accelerated meaningfully, with sales now activated for both large-scale training and inference. Clients like Revolut, Higgsfield, and Applied Compute use multiple scenarios; it is a launch partner for leading open models and ranks highly in independent benchmarks such as Artificial Analysis.

We will be among the first to support NVIDIA's new GRI LPX. The team is investing including through M&A; a recent acquisition brought leading model-loading capability that solves cold-starts, which is critical for handling bursts and improving GPU utilization.

Risk disclosure and statement:Dolphin Research Disclaimer and General Disclosures