The Zhitong Finance App learned that AI cloud infrastructure company Nebius Group (NBIS.US) announced pre-market results for the third quarter of 2026 on November 12, EST, and held a conference call on the same day. The company's third-quarter revenue was US$980.8 million, up 571% year on year, higher than LSEG's agreed forecast of US$907.5 million; GAAP net loss was US$37.5 million, equivalent to a loss of 13 cents per share, which was significantly narrower than the net loss of US$119 million (loss of 50 cents per share) for the same period last year; adjusted earnings per share were 31 cents, better than market expectations of a loss of 69 cents per share. After the announcement of the results, the stock fell 22.12% to $120.48 before the market, and closed down 16.84% to $128.64 during the regular trading session. The main reason for the sell-off was not quarterly results itself, but rather the company narrowed the 2026 revenue guidance to 3 billion to 3.2 billion US dollars, raised the 2026 capital expenditure guide to about 27 billion US dollars, and gave the 2027 capital expenditure guidance for the first time, all of which were significantly higher than market expectations. CEO Arkady Volo responded at the conference: “The demand for AI is real and huge. Everything else, like building capacity and financing it, is just a task.” He reiterated that the company expects revenue to reach $12 billion to $13 billion in 2027 and achieve a break-even before tax in the fourth quarter of 2027.
Focus and Response
The central focus of this conference is the relationship between the narrowing of revenue guidelines in 2026 and the sharp increase in capital expenditure, that is, the market questions “who will pay for AI expansion”. Management's response to this was: the adjustment of the guidelines only involves the time of delivery of third-party aircraft capacity and is not related to demand; CFO Dado clearly stated “we are very conservative in this guidance” and said that the 2027 forecast did not include potential upside for asset-light cooperation. CEO Voloz acknowledged that “the capital market has not been very quiet recently,” but insisted that building production capacity and financing “are all solvable tasks.” In terms of market interpretation, Barchart's analysis pointed out that the stock surged 15% with CoreWeave the day before the results were announced, and favorable expectations were overdrawn; the huge gap between the 2027 capital expenditure plan of 40 billion to 45 billion US dollars and market expectations (about 28.77 billion US dollars) was the trigger for the sell-off. Furthermore, according to reports, the news of Nvidia CEO Hwang In-hoon's resignation and US referendums restricting data center construction in many places have also increased overall volatility in the emerging cloud vendor sector.
Record revenue for the third quarter increased 571% year over year
Management characterized this quarter as a “record quarter”. The Group's revenue was US$980.8 million, up 571% year on year and 68% month on month. Among them, Nebius AI Cloud's revenue was US$973.8 million, up 643% year on year, accounting for about 99% of the group's revenue. The company said the growth was mainly due to the release of new production capacity over the past 12 months. The price of agreements signed this quarter continued to rise. The transaction price per megawatt was higher than US$25 million, higher than the range of 20 million to 25 million US dollars in the second quarter, and multi-billion dollar multi-year agreements have already begun to be signed.
Adjusted EBITDA margin rose to 47.5%
In terms of profit, the Group's adjusted EBITDA was US$465.8 million, with a loss of US$5.2 million in the same period last year and US$236.2 million in the second quarter; the adjusted EBITDA profit margin was 47.5%, higher than 40.6% in the second quarter, of which the AI Cloud business profit margin reached 53.2%. GAAP net loss narrowed to $37.5 million. The company says the difference in profit margins between the Group and AI Cloud comes from investment in Avride and TripleTen. Token Factory, a managed inference platform, is moving towards hundreds of millions of dollars in annualized recurring revenue (ARR). The quarter also confirmed non-recurring revenue of approximately $325 million, from cooperation under the asset-light model (asset-light model).
Prepayments cover more than 100% of capital expenses
The company disclosed that since the beginning of the third quarter, the total value of customer contracts (TCV) has exceeded 12 billion US dollars, with the vast majority coming from AI native (AI native) customers. More than 90% of the contracts signed in the third quarter include advance payments. The advance payments under the new contract now cover more than 100% of the capital required to launch the corresponding production capacity. The ratio for the second quarter was 50%-60%. The payback cycle for new production capacity was shortened from 1 year and 10 months in the same period last year to 1 year and 7 months. Databricks, Datadog, Snowflake, and Palantir signed multi-year agreements with the company for the first time this quarter.
Capital expenditure and financing arrangements
Capital expenditure for the third quarter was approximately US$7.3 billion, and the capital expenditure guidance for the full year 2026 was raised to approximately US$27 billion. In terms of financing, the company launched the first asset-backed financing (asset-backed financing) line of US$775 million in July, using long-term contracts for investment-grade customers, with an interest rate of SOFR plus 250 basis points; completed the issuance of US$5.75 billion convertible senior notes in August; and sold 4.6 million Class A shares at a weighted average price of US$229.1 per share through the ATM additional issuance plan, raising a net capital of about US$1 billion. By the end of the quarter, the company held $7.4 billion in cash and cash equivalents. Management expects 75%-85% of capital expenditure in 2027 to be covered by prepayments and asset-backed financing, of which customer advances are expected to reach US$17 billion to $20 billion in 2027.
2026 Guidelines Narrow 2027 Guidelines Revenue of $12 billion to $13 billion
The company narrowed its 2026 revenue guidance from the previous 3 billion to 3.4 billion US dollars to 3 billion to 3.2 billion US dollars. Management explained that the reason was the delay in the delivery of some third-party colocation (colocation) capacity rather than weakening demand; the adjusted EBITDA profit margin guidance remained unchanged at about 40%. In 2027, the company first gave revenue guidance of 12 billion to 13 billion US dollars, which is about four times the median value of the 2026 guidance range, the capital expenditure guideline is 40 billion to 45 billion US dollars, and is expected to achieve a break-even before tax (based on a five-year depreciation of servers and network equipment) in the fourth quarter of 2027. The company expects to sign 5 gigawatts (GW) of electricity by the end of 2026, and says that the supply of GPUs, CPUs, networks and storage for 2027 has all been locked in.
New Jersey project delays are power issues
In response to the delay in the Vineland project in New Jersey, management explained that the problem was in the power chain rather than demand: after the original diesel generator plan could not be implemented in late summer, the company switched to Bloom power generation equipment, and DataOne, which provides temporary power supply, decided to withdraw early at the end of the third quarter, causing a power supply gap. CEO Voloz acknowledged that “delays are common in our industry, and we are not spared”, but stressed that the affected production capacity has not been cancelled, and some have been scheduled until 2027. The company continued to deliver to Microsoft in the third quarter, and will continue to deliver in November and December. The self-owned site project (Alabama, Missouri, Finland and Pennsylvania) progressed as planned or even slightly ahead of schedule.
Token Factory and enterprise customer development
Management responded to growth quality and customer structure issues several times during the Q&A session. Token Factory is priced as a token and sold as a managed service. It has a cost advantage because it runs on its own GPU infrastructure, and will be one of the first supporters of Nvidia's new products. In terms of corporate customers, the company used Palantir as an example, saying that it combines Nebius's Token Factory with its own platform Foundry to serve sovereign AI needs; management says corporate adoption is still early but improving, and it is expected that more corporate business progress will be revealed in the next few quarters. At the same time, the company announced that it will stop guiding the ARR index and use revenue and profit before tax as the core evaluation criteria starting in 2027.
Full Meeting Minutes
Gili Naftalovich, Head of Investor Relations at Nebius Group: Welcome to the Nebius Q3 2026 results conference call. Today's conference was attended by CEO Arkady (Arkady), CFO Dado (Dado), and other members of Nebius's management team. Some statements involve forward-looking statements. Based on current expectations and assumptions, there are risks and uncertainties, and actual results may vary significantly. Please refer to the list of risk factors in the company's Form 20-F. The company is not obligated to update the forward-looking statements. This conference will present both GAAP and non-GAAP financial indicators. Reconciliation tables for non-GAAP and GAAP indicators are included in today's results announcement. All performance-related documents are posted to the public through the company's investor relations website (nebius.com). Now leave the meeting to Arkady.
Arkady Volozh (Arkady Volozh), CEO of Nebius Group: Thank you Gili. Last quarter I introduced you to our business model, and this quarter I'm going to show you how it works. First, let me say that Q3 was a record quarter — and rightly so against the backdrop of such market growth. The capital market hasn't been very quiet lately. The entire industry is being asked two questions: how fast AI can expand, and who will pay for it. These are important questions. My way of thinking is simple: the demand for AI is real and huge. Everything else, such as building production capacity and financing it, is just a task. The quests may be difficult, but they are all solvable. Demand is strong, customers promise us longer terms, higher prices, and pay more in advance, and prices continue to rise. The transaction price of the Q3 agreement was higher than US$25 million per megawatt, rising from the level of 20 million to 25 million US dollars in the second quarter. The size of customer agreements is also growing: three months ago, we reported our first batch of contracts worth more than $1 billion; entering Q4, we are already signing multi-billion dollar agreements for AI clouds. The base of demand is also broadening. Native AI customers are growing with us, and corporate AI leaders such as Databricks, Datadog, and Snowflake signed multi-year agreements with us for the first time this quarter.
Everyone knows that demand is extremely strong right now, but I'd also like to talk about how we can prepare for the future. We sell over three time spans and choose the best combination between the three: the first type is spot (spot) and short-term contracts, which are sold at a premium; the second type is a one-to three-year contract, which is our core business; and there are a limited number of five-year agreements to diversify the contract term structure and help finance the expansion. The latter two categories are covered by advance payments. Here are two key metrics: the proportion of capital expenditure (CapEx) covered by advance payments, and the payback cycle for new capacity. Customers pay in advance for production capacity to be launched within the next 12 months to lock in production capacity and lock in current prices. The most important change is that the advance payments on the new contracts now cover more than 100% of the capital needed to bring these contracted production capacity online — compared to 50%-60% in the previous quarter. As a result, the payback cycle for Q3 was shortened from 1 year and 10 months in the previous quarter to 1 year and 7 months. Customers pay more, pay faster, promise longer, and we get our capital back faster.
On how to achieve this model: First, the sales method. We're creating new tools to discover the true price of the capacity we sell, including reserved capacity auctions, public spot pricing, and short-term capacity. The market tells us what the capacity is worth, and we sell it at that price. Second, what helps profit margins is our low cost. We build the full stack, including both the computing infrastructure and the software on it. From its inception, we built Nebius layer by layer: first for large-scale training, then inference, and now post-training (post-training) and agentic (agentic) workloads. Customers start where it's right for them and then scale up with us as their needs evolve. Our managed inference platform Token Factory proved that this model works: it is moving towards hundreds of millions of dollars of ARR, and the profit margin is higher because we use our own GPUs, which are not leased production capacity. Our prices reflect the premium value of the product, and our cost structure is one of the lowest on the market—we build our own infrastructure, design and manufacture our own racks and storage, and these costs will drop further as more production capacity comes from our own site. Premium prices combined with low costs contributed to our profit margins. Our profit margins also benefit from lower capital costs: on top of the down payment, we can also finance with investment-grade contracts. All of this reduces costs, expands the market we can serve, and allows us to acquire more valuable customers. This is our path to profit, and we expect to achieve profit before tax in the fourth quarter of 2027.
At some point, supply and demand in our industry will balance. We'll be ready by then. We invest in three things at the same time: production capacity, software platforms, and go-to-market (go-to-market) where our customers are located. All three are equally important, but production capacity is the foundation of this business, and it defines our size. It is very difficult to build on this scale. Delays are the norm in this industry, and we are not spared. For example, some third-party hosting sites, including New Jersey, are lagging behind schedule. Customers have been advised that no production capacity has been cancelled and part of it has been postponed to 2027. This means that part of this year's planned revenue will move to early next year. Our own site is on track and even ahead of schedule — this is important because most of our next gigawatt (GW) production capacity will come from our own site, where our own team is responsible for everything from land, electricity, plant, hardware, software to operations. This gives us confidence for next year. We're building the next gigawatt, making us a company of all sizes, and we're fully prepared for this goal. Payments will exceed $10 billion in 2026 and will increase significantly next year; we can obtain financing with favorable terms through investment-grade contracts, and we also have a strong balance sheet and potential external financing channels. We have full confidence in 2027 and are planning for 2027 and beyond. The industry has ups and downs like any industry, and we are building a business model and financial model that can both thrive in a good market and remain resilient in the face of adversity. We expect that under the premise of maintaining a conservative judgment on the pace of construction, revenue will increase by about four times next year and achieve profit before tax in the fourth quarter. Finally, the future of our business has never been better, our confidence in the business we have built has not changed, and we will continue to move forward. Let's leave it to CFO Dado.
Nebius Group CFO Dado: Thank you Arkady. We delivered a strong third quarter. Revenue increased 571% year over year, mostly from additional production capacity over the past 12 months. EBITDA profit margins increased year-on-year and month-on-month. We've also strengthened the funding situation: advance payments have become the standard structure of the agreement, and advance payments promised by customers are expected to reach $10 billion in 2026; long-term agreements with Microsoft and Meta can be used for further asset-backed financing; and $7.4 billion in cash and cash equivalents are held on the balance sheet at the end of the quarter. Furthermore, operating cash flow from each batch of production capacity is increasing. I'll start with the third quarter results, balance sheet, and capital expenditure funding requirements, then the updated 2026 guidance and preliminary outlook for 2027. Unless otherwise specified, the following comparisons are year-on-year.
Let's look at revenue first. The Group's revenue was US$980.8 million, up 571% year on year and 68% month on month. Nebius AI Cloud's revenue was US$973.8 million, up 643% year over year, accounting for about 99% of the Group's revenue. In the third quarter, we confirmed non-recurring revenue of approximately $325 million as part of our asset-light model: in the early stages of this model, we worked with partners to deploy the Nebius architecture and Nebius AI Cloud platform on their production capacity. Under this kind of cooperation, we confirm a portion of revenue when we deploy the architecture and platform to partner production capacity, and then confirm part of the revenue when production capacity is launched. As the light asset layout expands, we expect this type of revenue to become a normal part of the business.
Let's look at EBITDA again. The Group's adjusted EBITDA was US$465.8 million, negative US$5.2 million in the same period last year, and US$236.2 million in the second quarter. The Group's adjusted EBITDA margin was 47.5%, up from 40.6% in the second quarter. Nebius AI Cloud's adjusted eBITDA margin was 53.2%, up from 50% in the second quarter. The difference in profit margins between the Group and AI Cloud reflects our investment in Avride and TripleTen.
Next is the balance sheet and funding. Advance payments became the standard structure of agreements this quarter: more than 90% of the contract value signed in Q3 included advance payments, and customers paid a higher proportion of advance payments in each agreement. The weighted average share of advance payments in the total contract value increased by about 20 percentage points from the second quarter. As stated in Arkady, the coverage rate exceeds 100%, covering all capital expenses corresponding to contracts signed in Q3, which is higher than 50%-60% in the second quarter, and this ratio continues to rise in the agreements signed in Q4 so far. These prepayments cover all capital expenses, including GPUs and networking equipment. Advances now cover both agreements for deployed capacity and agreements for new capacity, so they fund overall expansions rather than individual projects. Because we independently choose our contract package—by client, term, and attached prepayment—we're actually choosing how much external capital we need and when to obtain it, which allows us to seize opportunities rather than passively rely on them. We will finance at the cost of capital and when the time is right, weighing the impact of each funding source on capital costs, equity dilution, and balance sheet health, without relying on any single channel to finance 2027. In July, we landed an amount of 775 million US dollars. This is the first transaction under the asset-backed financing framework. The interest rate is SOFR plus 250 basis points as collateral. These terms and the flexibility of when we choose to use these amounts are the reasons we sign this type of contract. We already have over $40 billion in similar promised revenue and see this as a replicable framework. In August, we completed the issuance of US$5.75 billion convertible senior notes (convertible senior notes). This quarter, we also issued 4.6 million Class A shares at a weighted average price of US$229.1 per share through the AT-the-Market (ATM) program, raising a net capital of approximately US$1 billion, with 7.6 million shares remaining to be issued. ATMs are flexible tools, not promises.
Let's talk about capital expenditure again. Capital expenditure for the quarter was approximately $7.3 billion, mainly due to the expansion of servers and network equipment and in-house data centers. We will utilize all available financing instruments.
About the outlook. We've updated our 2026 guidance and provided an initial outlook for 2027 in the same caliber — that is, based on what we see today and a more careful judgment on the pace of delivery of all sites. 2026: As mentioned by Arkady, we see some third party production capacity postponed to 2027 (originally second quarter of 2027). We now expect revenue of $3 billion to $3.2 billion, compared to the previous range of $3 billion to $3.4 billion. The annualized run-rate revenue (run-rate revenue) reflects the same changes over time. We now expect ARR to surpass $7 billion in February — this is still before New Jersey reaches full capacity. After this update, we will no longer guide ARR metrics because the business is entering the next phase: the indicator is based on a single month in the quarter and does not reflect the full picture of our business foundation; we do not operate the business around any single month; as new partnerships expand, this type of revenue will also cause fluctuations in ARR. As our business matures, we think revenue and profit before tax are the most appropriate way to evaluate us. We continue to expect the Group's adjusted EBITDA margin of around 40%, in line with previous guidance. Capital expenditure in 2026 is currently estimated at around US$27 billion, higher than the previous range of US$20 billion to US$25 billion. The contracted electricity remained unchanged, reaching 5 GW at the end of the year, higher than the current 4 GW. Demand has never been so strong. The adjustment to the 2026 guidelines is only a matter of timing, not a matter of demand.
2027 outlook: We expect revenue to increase approximately fourfold to $12 billion to $13 billion. We are also excited about the profit path: we now expect to achieve a break-even before tax in the fourth quarter of 2027, based on a five-year depreciation of servers and network equipment, as owned sites continue to climb the decline over the course of the year. We expect capital expenditure of 40 billion to 45 billion US dollars, corresponding to more than 1 GW of additional production capacity, most of which comes from our own site, and its profit margin structure is superior to that of third-party hosting capacity. To support this expansion, we expect 75%-85% of the capital expenditure range to be covered by asset-backed financing through prepayments and existing investment-grade contracts. Customer advances are expected to be $17 billion to $20 billion in 2027, and this figure is not an upper limit; the size of the advance payment depends on how we allocate production capacity between contract types, and we set the combination for the best balance of price, term, and advance payment. We also have a very healthy balance sheet, and can flexibly seize opportunities when external capital conditions make sense. We are entering 2027 with a strong attitude: a clear break-even path for the fourth quarter, a financing plan based on customers, existing contracts, and balance sheets, and an outlook we are confident we can deliver on. We will continue to expand with consistent discipline. Let's leave it to Gili to enter the Q&A session.
Q & A session
Gili Naftalovich, Head of Investor Relations at Nebius Group: The first question comes from Alex Duval of Goldman Sachs Research (Goldman Sachs): “If additional funds are needed in addition to existing cash, to what extent can advance payments meet demand, given the conditions in some bond markets?” Marc, please answer.
Nebius Management (Sales/Commerce) Marc: Thanks Alex. The simple answer is: Yes, we can use an advance payment to cover any additional cash requirements. Let's expand a little bit to let you understand the role that advance payments play in overall execution. We won a few landmark deals last year — Microsoft and Meta, both of which come with large upfront payments, and since then we've known that prepayments are a fundamental part of our business. For the AI cloud business, we began implementing advance payments in the first quarter of this year. Today, it is the standard structure of the agreement, and more than 90% of the contract value signed in Q3 includes advance payments. How do we get these upfront payments? There are three factors: first, we are seen as a market-leading platform, and AI developers are willing to pay a premium for our reliability and service; second, we said last quarter that we could sell full production capacity under current terms, but we chose to stay on hold for better terms, and now we have done it — we have increased our advance payments while increasing prices; third, due to our overall management of the cost structure, the absolute amount of advance payments corresponds to a higher capital expenditure coverage ratio. Higher upfront payments combined with lower costs enabled more than 100% coverage of total capital expenditure, including data center costs, this quarter. Another point that may not be obvious: we received the advance payment long before delivery. The prepayments also apply to the capacity we already have, so they fund the overall expansion. I want everyone to see that the advance payment arrangement provides great flexibility for the continued steady financing of our business. As Alex pointed out, it gave us leverage to supplement our cash needs when necessary. I think Ophir could add some insight on using prepayment financing.
Nebius Management (Financing/Cooperation) Ophir: Thank you Marc. You're very comprehensive, but at the end of the day, when the contract advance covers more than 100% of capital expenses, you're in a very good position. When allocating between various contract types, we can also take into account capital market conditions. For example, in a period where corporate debt interest rates are high, we can decide to change the way we allocate production capacity, provide more long-term contracts, and obtain large advance payments to cover capital expenditure requirements; if market conditions change, we can issue corporate bonds or other instruments very opportunistically, and then reserve more production capacity for shorter term contracts. We have a great deal of flexibility. The results of Q3 and Q4's advance payments to date have enabled us to finance our business in a very attractive way.
Gili Naftalovich: Alex Duval's other question: “Can you talk about how Nebius works with local governments and communities, taking into account the situation in places like Texas? Can your approach ease deployment resistance?” Tom, please answer.
Nebius Management (Government/Community Relations) Tom: Yes. Regarding Texas, we don't currently have a business there, but I can talk about the situation in about six states where we already operate or are building. Up to now, we have not seen or anticipated that changes in the political landscape will have a substantial impact on the project schedule. More broadly speaking, the issues surrounding AI, and data centers in particular, do not only come from the Democratic Party; in fact, both parties have them. We think many of these questions are reasonable and worth answering, and we are doing our best to answer them. We haven't seen a change in the overall mood trajectory since the election. Questions will continue to be raised, and all we need to do is make sure we respond. We actively communicate with state leaders and communities to be transparent and explain what we are doing and what benefits we can bring to the community. We think we did a good job and will continue to do it.
Gili Naftalovich: Next question from Baird's Rob Oliver: “What's the 2027 financing strategy? How much additional funding will be needed to support the current backlog and data center expansion?” Dado, please answer.
Dado: Thanks Rob. Marc and Ophir have shared many details of their financing ideas. We have put a lot of effort into refining the product and marketing plan, so that the advance payment covers more than 100% of capital expenses and is received before capacity is delivered. We have also prepared tens of billion dollar investment-grade contracts, based on which asset-backed financing can be obtained on very attractive terms. These two factors gave us a lot of flexibility. Remember, we can allocate production capacity between long-term and short-term contracts for core businesses according to capital requirements: if capital market terms are unattractive, we can allocate more production capacity to long-term contracts; if we can finance on attractive terms, we act very opportunistically, shifting more allocation to short-term and spot markets. At the end of the day, the down payment covered 100% of capital expenses (including the data center), which placed us in a very good position in terms of financing needs.
Gili Naftalovich: Daniel Wang of BNP Paribas (BNP Paribas) asked: “Can you further explain the underlying causes of Vineland's delivery delays? The shareholders' letter mentioned that several hosting sites have been delayed. Which other sites have been affected?” Andrey, please answer.
Nebius Management (Operations/Infrastructure) Andrey: Thanks Daniel. Regarding Vineland, the main cause of the delay was electricity problems. When we switched to the Bloom power generation plan in the summer, the plan was to temporarily use the temporary power supply provided by DataOne during Bloom's power generation climb. However, at the end of the third quarter, DataOne decided to withdraw temporary power supply sooner than expected, so we are speeding up Bloom's deployment as fast as possible, but this situation will still cause a gap. It should be noted that we continued to deliver new batches to Microsoft throughout Q3, and will continue with electricity deliveries in November and December. Second, Bloom is actually already generating electricity on site. The generator set is already running online, but it just needs to climb the slope as soon as possible.
CEO Arkady Voloz: I'd like to add that, as I said before, this is one of the sites we built using a third party contractor rather than our own. The probability of problems with this type of site is already higher. In this case, in August, we were told that the original diesel generator plan didn't work. We came up with Bloom's new plan and implemented it, and electricity has already arrived. Of course, the delay was several months, but we are back on track, the electricity is in place, and the project is still progressing. The contractor, Microsoft, and our engineering teams all worked together, and the teams treated each other with respect. In fact, we are planning to use Bloom to generate electricity at more sites. Incidentally, this new plan may also bring some extra electricity to the site, and we'll discuss how to use it. One last point: even with this delay, our revenue guidelines remained within the range.
Gili Naftalovich: Ryan Lountzis of Morgan Stanley Research (Morgan Stanley) asked: “How do you consider the level of demand and TAM for short-term contracts compared to long-term exclusive capacity contracts? How will the short and long term combination evolve?” Marc, please answer.
Marc: Thanks Ryan, great question. This ties in to a few of the things we've talked about today; I'm trying to expand my perspective a little bit. We're trying to be market-driven, that is, to satisfy customers where they are—we're not making markets, but beneficiaries of what Arkady says is demand. As shared many times in the past, we maintain an opportunistic attitude and have been developing ways to obtain stronger pricing and term signals: through auctions, spot pricing, and in-depth communication with customers and prospects about their long-term plans, we are able to adjust market terms and offers to optimize results. We now have systems in place, and we can adjust execution to optimize these results. We seek to optimize between short-term consumer demand and long-term reserved production capacity. Just as Dado concatenated Ophir's answers to the advance payment issue, the same flexibility allows us to switch between different offers to optimize short-term income and capital security for future construction. Looking at the long term, based on Arkady's vision, we want to move all of these capabilities to a standardized trading platform to algorithmically adjust our offers for all types of customers — from small self-service customers to large-scale sales-assisted enterprise customers.
Gili Naftalovich: Tyler Radke of Citi Research (Citi) asks: “What are the key assumptions behind initial 2027 revenue guidance?” Dado, please answer.
Dado: Thanks Tyler. The 2027 guidance is $12 billion to $13 billion, and we are quite confident in this guidance. Achieving it requires two parts: production capacity to be ready on time and ready for sale. In terms of production capacity, we have taken a very, very conservative approach to time, which reassures us. In fact, a significant portion of our 2027 capacity will be owned by us rather than by a third party, which gives us control over delivery. We believe production capacity will be in place on time, which may even increase our stated revenue growth rate — if so, revenue may be higher than the guideline. On the demand side, we have no doubt that this capacity will be sold: customers are already looking for capacity for 2028, and the price of production capacity is rising in 2027, so we don't see any risk in this regard. Overall, we are very confident in the guidelines.
Gili Naftalovich: Tyler's question: “Since the beginning of Q3, customers have promised over $12 billion in TCV, but the RPO has only increased by 3%. How much of this TCV is not included in the RPO? What are the main factors limiting the growth of in-schedule orders?” Marc, please answer.
Marc: Sure, Tyler. Let me first talk about how this quarter unfolded and how it relates to the potential energy we are in. We've been saying for several quarters that we can sell all of our existing capacity at the current price — and that includes selling future capacity at the current price. What we didn't know before was the proper current terms. As I've answered before, we've been learning about pricing through projects such as auctions and spot pricing, and we've also been discussing their longer-term needs with our customers and having a clear understanding of the terms that may be agreed upon. At the beginning of September, we raised prices and added higher advance payment terms and contract terms to the new batch of production capacity released. As a result, since the beginning of Q3, customers have committed more than $12 billion in contract value, including multi-billion dollar five-year agreements. Q2 We announced that we have won four landmark orders, followed by expanded cooperation with similar customers — longer terms, higher prices, and more upfront payments. Several of these agreements were signed at the end of the quarter and will be reflected in commitments reported for the next quarter. Importantly, this momentum has continued into Q4.
Gili Naftalovich: James Kisner of Water Tower Research asked: “Token Factory was Kimi K3's July Day 0 launch partner, and OpenRouter's September report stated that the open weighting model accounts for 72% of its weekly token volume. How do you view the long-term pattern of open weights and closed source models? Would the shift to open weights be a net benefit for Nebius?” Roman, please answer.
Nebius Management (Product/Software) Roman: Thanks James. As a reminder, at the beginning of Q3 and Q4, we were Day 0 launch partners for more models, including Meta, Reflection AI, NVIDIA, etc. We have indeed seen an acceleration in the adoption of open models over the past few months, and this just happens to bring AI to large-scale applications with reasonable economics and quality improvements driven by the company's unique data and business knowledge. We've built a full suite of services to support our customers in this transformation. We've always said that the more diverse the market, the better for us — this diversity can come from open models or other ways to build dedicated models for specific use cases. Why is this good for Nebius? Because it expands the range of customers and partners we can work with and the value we can create. With more new products and more companies bringing AI into production environments, we will continue to support the adoption of diverse models in the ecosystem.
Gili Naftalovich: Piper Sandler's James Fish asked: “You plan to add more than 1 GW of computing power every year, while being careful about the commissioning time. What are the major constraints between contracted electricity and the computing power available to customers? What milestones should investors track to determine if the deployment progresses as planned? Are you adjusting the way you communicate about the schedule?” Andrey, please answer.
Andrey: I'll answer. First, electricity and electricity-related scheduling are major challenges facing the entire industry and us. The major change for us in 2027 is that most of the production capacity delivered to our cloud business will come from our own site — this has been mentioned in several answers — which gives us more confidence than third-party hosting projects, because these projects were developed by ourselves from the beginning, and we have full control over the pace of the project. In 2027, we have two projects in place, Alabama, Missouri, and Finland, which will contribute revenue in 2027, and Pennsylvania will also go online and may contribute future revenue. We are very confident in the execution of these sites. Another recognized major bottleneck is the chip supply chain—GPU, CPU, networking, and storage. We're in a great position in this area: we've never had issues before, and I don't expect 2027 to be an issue. As early as 2026, we began related work for 2027, including R&D and locked in supply, and we anticipated this. When it comes to the new generation of chips launched in 2027 — Vera CPUs, Vera Rubin GPUs, and GroQ (original text) — all R&D integration, testing, and benchmarking have been completed. From a technical perspective, we are fully prepared; from a supply chain perspective, we have confirmed full supply for 2027. We are also very confident in implementation.
Gili Naftalovich: Arsenije Matovic of Wolfe Research asked: “Please take us to sort out the $325 million non-recurring revenue confirmed in Q3: under the asset-light model, what contributed to the confirmation of such a large amount of revenue during the quarter? How will the future continue?” Ophir, please answer.
Ophir: I'll answer that. We're actually very excited about this result and think it's one of the most beautiful results of the quarter. We have been exploring and adopting various mechanisms and partnerships to bring more production capacity online. One of them is the asset-light model: we export the Nebius architecture and bring production capacity online. Under this model, part of revenue is recognized when production capacity, supply chain, and structure are locked in, and the other part is recognized when production capacity is actually sold to customers. This is actually our first deal, and I expect we'll incorporate them into the underlying model as deals increase, become more regular, and become part of the business. Here I would like to highlight an important point about the 2027 revenue guide: as mentioned, we are very conservative in this guidance. For example, we made a very important assumption — there could be significant delays in capacity delivery, even if most of the capacity will be owned by us and we have more control over delivery times. So if we deliver on time, things will obviously be much better. The same is true for partnerships and asset-light models: we have taken a conservative approach and have not included any potential revenue from asset-light models or other partnerships in our 2027 guidance. If these opportunities are implemented and expanded, they will clearly have an impact on the guidelines.
Gili Naftalovich: There are a few questions about the composition of AI native customers and enterprise customers, including this one from Rob from Baird: “Can you sort out the growing customer composition of new customers — enterprise customers, AI native customers, and others?” Marc, please answer.
Marc: Sure, thanks Rob. We don't actually disclose results by segment, but I'd be happy to provide the details you'd like to know. The corporate business is gaining momentum, but it is still far below our expectations for the overall potential of the market. An important leading indicator we've seen in our performance is the adoption of leading data and infrastructure platforms built on it — as Arkady mentioned earlier, we have signed multi-year agreements with companies such as Databricks, Datadog, Snowflake, and Palantir. We're actually experiencing the wave of enterprise AI adoption they've experienced. Palantir is probably a good example: they specifically chose Nebius as their preferred AI cloud, combining our Token Factory solution and AI cloud capabilities with their enterprise data and operations platform called Foundry, combined with an open weighting model to serve the sovereign AI needs of large enterprises. I recently attended one of their training camps, and the company's interest in using AI to automate and transform businesses was exciting. I'm very confident about this kind of opportunity, and we're looking for similar opportunities with the other platform companies mentioned above. At the same time, we are also seeing direct adoption of leading enterprise logos. Companies are beginning to establish the first non-hyperscaler (hyperscaler) partnerships, often to achieve the innovation they seek within their organizations. I expect we'll talk more about the accumulation of corporate business in the next few quarters.
Gili Naftalovich: Michael Donovan of Compass Point Research asked: “What are the drivers behind the adjusted EBITDA margin rising from 41% in Q2 to 47.5%? Considering that around 40% of the guidance for the whole year implied that Q4 fell back to the mid-30% range, how should we view Q4?” Dado, please answer.
Dado: Thanks Michael. Our guideline for 2026 is an EBITDA margin of around 40%. From quarter to quarter, profit margins fluctuate with the time of confirmation: new sites must bear rent, electricity, and labor costs before billing; in the third quarter, revenue was confirmed in time and will not be reproduced in the same form. I don't recommend overanalyzing the ups and downs of any single quarter. We don't actually do quarterly guidance, but for 2026, as I said, we expect an EBITDA margin of around 40%. What I can say about Q4 is that it includes the full quarter of new production capacity added in Q3, further hosting capacity batches, and new ERC batches delivered in November (original text), while bearing the costs generated by the site in advance before revenue arrives. For 2027, the direction is upward: production capacity at our own site was launched in the second half of the year, there were no third party rents, and agreements signed at higher prices and larger advance payments since Q3 were converted into revenue. We are quite confident about the EBITDA margin guideline of around 40% for the full year of 2026.
Gili Naftalovich: Northland's Nehal Chokshi asked, “What percentage of the $12 billion TCV signed in Q3 came from AI-native customers and large enterprises?” Marc, please answer.
Marc: Thanks Nehal. The vast majority of the $12 billion came from AI-native customers, but I'd like to go into more detail. Most of these customers — I think we've mentioned this, but want to emphasize it again — are mostly starting from other parts of our platform, from auctions, or have deployed production capacity on the platform for other workloads. What they all have in common is rapid expansion. Also interesting is the revenue growth they have experienced: they are making significant commitments around large-scale revenue opportunities, and we are honored to be involved and excited about their prospects and growth.
Gili Naftalovich: Another question that came in: “What are the updates to the software stack? It is understood that it is driving profit margins and forming a moat against competitors.” Roman, please answer.
Roman: Thanks for asking. Talk about Token Factory. Token Factory is priced per token, sold as a managed service, and offers a variety of SLAs that differ from GPU hourly infrastructure. It provides day 0 support for the new model, advanced routing and caching capabilities, and various inference optimization techniques such as custom kernels (custom kernels) and quantization (quantization), as well as post-training capabilities and related inference optimizations. This is a relatively new service, growing very fast from a low base. We don't look at profit margins by product, but since we run on our own infrastructure, our cost structure is far superior to any inference service provider that rents GPUs. As Arkady said at its inception, this is not an isolated business—Nebius uses the entire platform to serve customers the way they want them to. The beginning of Q3 and Q4 was a period of significant acceleration for Nebius Token Factory. We have activated a marketing organization around large-scale training and inference workloads. Many customers, including Revolut, Higgsfield, and Applied Compute, are combining different use cases on Nebius to benefit from the flexibility of the integrated platform. Token Factory is the Day 0 launch partner for the leading open models we've discussed, and continues to deliver leading performance in independent benchmarks such as Artificial Analysis. More importantly, we've seen the same performance in real customer workloads. We help customers train and run specialized models at scale across diverse use cases — including personal AI assistants, code generation, and long-term enterprise agents. Token Factory benefits from running on Nebius infrastructure and can directly use the latest and most powerful GPUs. As an example, we will be the first service provider to support the new NVIDIA Groq 3 LPX product (sic). Behind this progress is aggressive investment in the team, including our previously completed acquisitions, Eigen AI and Clarifai, which are now fully integrated, and recently Inferize. Inferize brings leading-edge model loading capabilities and solves the cold start (cold start) problem, which is particularly important for improving GPU utilization under highly volatile on-demand workloads. This is what we call a vertically integrated platform: everything runs on the same platform, and customers can scale from one product to the next and choose the best way to consume every workload on the platform.
Gili Naftalovich: Thank you, this concludes this call.
Operator: This concludes today's session. Thank you all for participating, you can disconnect.