CoreWeave (CRWV.US) Q3 Earnings Call: CEO Addresses High Debt Concerns, Says "We Will Not Use the Unsecured Debt Market," 75% of Committed Capex Tied to Investment-Grade Customers

CoreWeave (CRWV.US) Q3 Earnings Call: CEO Addresses High Debt Concerns, Says "We Will Not Use the Unsecured Debt Market," 75% of Committed Capex Tied to Investment-Grade Customers AI cloud computing service provider CoreWeave (CRWV.US) held its earnings call after releasing its Q3 2026 results (quarter ended September 30) after market close on November 10, Eastern Time. With revenue surging 168% year-over-year and revenue backlog approaching US$132.5 billion, management devoted substantial time to addressing the market's core concerns over its high debt load and financing costs. Chairman and CEO Michael Intrator called the quarter "a breakout quarter," emphasizing that the company employs "success-based capital expenditure," that more than 75% of committed capex is tied to investment-grade and enterprise customers, and made clear that at current yield levels it "will not use the unsecured debt market" for financing. He also offered a key judgment: compute capacity is "essentially sold out" for 2027, and infrastructure being brought to market in 2028 has also been booked. Quarterly revenue was US$3.661 billion, up 168% year-over-year and up more than US$1 billion sequentially, above the consensus estimate of roughly US$3.543.56 billion. Adjusted operating income was US$379 million (10% margin), well above the company's prior guidance range of US$200260 million; adjusted EBITDA reached US$2.256 billion (62% margin). On a GAAP basis, the company's net loss widened to US$709 million from US$110 million a year earlier, with a loss of US$1.27 per share, but better than the market's expected loss of US$1.46 per share. The company raised full-year revenue guidance to US$13.013.3 billion and adjusted operating income guidance to US$1.2751.375 billion. Following the report, CRWV rose as much as about 8% after hours; the stock had closed regular trading down 0.51% at US$48.73 that day. Focus and Response: High-Debt Financing Concerns vs. Strong Demand The most closely watched point of contention at this call was that, even as the company delivered standout growth, its net loss continued to widen due to high financing costsquarterly net interest expense nearly doubled year-over-year to US$846 million, already exceeding adjusted operating income of US$379 million. Reuters also reported that the company's bond yields had at one point risen to about 15.2%. Addressing this concern, Intrator responded systematically in his remarks: the company's weighted average debt cost has fallen 250 basis points year-over-year to 7.9%; 70% of this year's transactions have collected customer prepayments; roughly US$9 billion in prepayments and deposits will be collected over the next three months; the company has repaid more than US$4 billion of debt through delayed-draw term loans, and "at current yield levels will not use the unsecured debt market." He stressed that when contracts expire, the debt will have been fully repaid, and the infrastructure will be "wholly owned by CoreWeave and debt-free." Q3 Performance Characterization: Margins Reach Double Digits a Quarter Early Intrator called the quarter "a breakout quarter," citing rapid revenue growth, margin expansion ahead of plan, and extremely strong customer demand. Adjusted operating margin rose from about 1% in Q1 and 5% in Q2 to more than 10% this quarter, entering the low-double-digit range a quarter early. Management said Q3 results and the raised Q4 guidance were driven mainly by three factors: contracted power rapidly converting into active power and revenue-generating capacity, a strengthening demand and pricing environment, and expanding platform adoption across industries. Demand and Backlog: US$132.5 Billion Revenue Backlog, Essentially Sold Out for 2027 As of the end of September, the company's revenue backlog reached about US$132.5 billion, up further from about US$104.2 billion at the end of June. Intrator said compute demand is growing faster than infrastructure delivery capacity; the company is "essentially sold out" for 2027, and infrastructure brought to market in 2028 has also been sold; by the end of 2027, qualified demand is 10x available capacity. Active power has exceeded 1.7 GW, doubling since the start of the year, and is expected to exceed 1.85 GW by year-end. Financing and Capital Allocation: Success-Based Capex, No Unsecured Bond Issuance Intrator said the company employs a "success-based capex" model, allowing it to adjust the pace of investment based on customer demand, regulatory impacts, and capital market conditions, with most capital committed only after contracts and infrastructure are in place. Through 2027, more than 75% of committed capex is tied to investment-grade and enterprise customers, with the remainder tied to world-leading AI labs and AI-native companies. He expects 2027 EBITDA less interest (excluding prepayment benefits) to "more than double," strengthening internal financing capacity. The company expects its 2027 adjusted net loss to narrow significantly and adjusted net income to turn positive in the first half of 2028. Pricing and Renewals: Prior-Generation GPUs Re-Contracting at 30%50% Premiums Intrator disclosed that in Q3, pricing for Ampere and Hopper (prior-generation GPUs) rose 30%50% sequentially and above original contract prices, with multi-year commitments signed extending beyond the depreciable life. More than 130,000 GPUs will expire in 2027; renewing at current prices could generate close to US$4 billion in incremental annualized revenue; another roughly 170,000 GPUs expire in 2028, bringing cumulative incremental annualized revenue to close to US$10 billion, all "without additional capex." CFO Nitin Agrawal added that this incremental amount is not in the current revenue backlog and represents the available opportunity after re-contracting. Customer Diversification and Forge: 10% of the Fortune 500 Are Customers Management emphasized diversification of customer concentration. 10% of Fortune 500 companies are already customers, and 10% of backlog comes from financial services; more than half of new customers this year came from outside AI labs and the tech industry. Intrator introduced CoreWeave Forge, launched in September, which integrates AI development capabilities into one place; 70% of new cloud customers in Q3 were already Forge customers, making it a key entry point into the enterprise customer base. Full-Year Guidance: Revenue Raised to US$13.013.3 Billion The company raised full-year guidance, expecting 2026 revenue of US$13.013.3 billion (previously a ceiling of US$13.2 billion), adjusted operating income of US$1.2751.375 billion, year-end annualized revenue run-rate of US$2021 billion, and full-year capex of US$3739 billion (lower bound raised). Full Call Transcript Call time: 5:00 p.m. Eastern Time, November 10, 2026 (after market close) Company speakers: Michael Intrator Chairman and CEO Nitin Agrawal Chief Financial Officer Analysts participating: Gabriela Borges (Goldman Sachs Group, Inc. Research Division), Brad Zelnick (Deutsche Bank Aktiengesellschaft Research Division), Samik Chatterjee (JPMorgan Research), Karl Keirstead (UBS Group AG Research), Ben Reitzes (Melius Research), Amit Daryanani (Evercore ISI). Operator Good afternoon, and welcome to the CoreWeave Q3 2026 earnings call. Following the prepared remarks, we will open the line for questions. Michael Intrator (Chief Executive Officer) Good afternoon, everyone, and thank you for joining us. CoreWeave delivered a standout Q3: rapid revenue growth, margin expansion ahead of expectations, persistently strong customer demand, and newly signed contracts that are the most profitable batch of orders in our history. The cloud infrastructure we built early is renewing at a premium, generating more cash flow than we modeled when we originally invested. The company's funding is supported by secured loans, customer prepayments, and operating cash flow. Together, these tailwinds make our business highly resilient and well positioned to navigate the current market environment. Here are the specifics: revenue of US$3.7 billion, up 168% year-over-year and above the high end of our guidance range; more than US$1 billion of sequential revenue growth. Adjusted operating income of US$379 million, well above guidance of US$200260 million, with operating margin entering the low-double-digit range a quarter early. Revenue backlog grew to about US$133 billion; active compute power rose to more than 1.7 GW, with compute scale doubling since the start of the year. Last quarter we noted that operating leverage would become evident as infrastructure reaches scale and the business balances between long- and short-cycle contracts. This quarter's results and our guidance confirm that. We have also long believed that AI will permeate every industry, and that the people building that future need a new form of cloud infrastructure. That conviction is becoming reality every quarter. Customers use our AI compute platform to advance biological research, support clinical care, develop autonomous driving systems, and enable personalized learning. Our AI cloud platform helps them bring innovative ideas to life and deliver practical applications to the public. Bringing these ideas to life requires a stable, cost-effective compute foundation for developing and deploying AI models. Our core mission is to build that foundation while generating attractive returns on capital. Given recent capital market volatility, I want to explain how this business logic holds together: the value we create for customers earns their long-term commitments; those commitments are the foundation of our financing framework; and disciplined investment lets us keep expanding the compute platform customers rely on while continuing to compound shareholder value. Every decision follows one core principle: every choice pursues the best economic returnincluding financing method, pricing strategy, and pace of expansion. Building a hyperscale AI cloud requires enormous capital investment and strong execution discipline. CoreWeave has understood this since inception, and our business model and financing strategy have been designed around long-term customer contracts from the start, with capital investment serving contracted demand. This strategy lets us evaluate every new dollar of capital and, given market conditions, choose the financing channel that maximizes returns. As of the end of Q3, the company's weighted average cost of debt was 7.9%, down 250 basis points year-over-year. Looking ahead, at current yields we will not use unsecured debt financing, nor do we need to. Consistent with our approach, most of the remaining capex needed to support the backlog will be funded through delayed-draw term loans tied to specific customer contracts. By 2027, more than 75% of committed capex corresponds to investment-grade and large enterprise customers, with the remainder for the world's leading AI labs and AI-native companies. This type of financing is priced off the underlying customer contracts and our proven delivery capability, and can cover 70%100% of the corresponding capex. Since our IPO, these financing rates have generally ranged from SOFR plus 225550 basis points; prime investment-grade customers price near the low end of the range, while AI lab customers skew toward the high end. At current rates, the effective financing cost at the high end of the range is less than 10%. Drawdowns on these loans match the pace of equipment procurement and delivery; cash flow from the contracts covers interest and principal repayment at maturity while also generating additional cash flow for CoreWeave. As planned, we have already repaid more than US$4 billion of debt through this model, including nearly half of the two most expensive initial delayed-draw term loans. For financing projects serving long-term contracts, when the customer contract expires, the corresponding debt has been repaid and the infrastructure is wholly owned by CoreWeave. This debt-free installed compute base has itself become an important source of capital. A batch of existing contracts expires every year, generating revenue without new investment. Equipment procurement and data center construction are complete, and our capital intensity will continue to decline as we scale. Prepayments are another important source of capital. We price prepayments on a market basis, weighing the pricing discount against financing costs. We have used the unsecured debt market in the past; in the current environment, prepayments are more attractive. This year, we collected prepayments on 70% of customer transactions, in some cases covering the full contract value upfront. In the next three months alone, we will collect about US$9 billion in prepayments and deposits from signed customer contracts, more than US$3 billion of which comes from renewals of existing cloud infrastructure at a significant premium to original pricing. In addition, EBITDA less interest expense (excluding prepayment benefits) is expected to more than double in 2027, enhancing internal funding coverage of capital needs. OEM vendor financing can also support additional compute deployment, particularly for capacity placed for short-term demand. Overall, we expect these funding channels to cover the vast majority of 2027 capex needs, with only a small amount of additional financing needed at the parent level. Discipline in capex control is as important as access to capital. Beyond investments tied to locked-in contracts, our "success-based capex" model allows us to flex the pace of expansion based on customer compute demand and supply chain constraints (including regulatory and capital market changes)an important risk hedge. The vast majority of capital investment is initiated only after contracts are in place and infrastructure conditions are ready. Nitin will elaborate shortly. Next, let me connect the financing model to the profitability logic reflected in our results. As more contracted compute is brought online, revenue grows on the back of capital already deployed, and new deployments account for a declining share of the overall businessthis is the operating leverage we discussed last quarter. Adjusted operating margin was about 1% in Q1, 5% in Q2, and above 10% in Q3. Profitability on newly signed business is also improving. In July we raised prices across compute SKUs by about 25%, and raised them another 10% later in Q3. Prices continued to rise in Q4, with increases fully covering cost inflation. At the same time, we are allocating more compute to attractively priced short-term orders. Short-term compute can be priced at an annualized US$40 million per megawatt; new 45 year long-term contracts price in the US$2025 million per megawatt range. Long-term contracts provide visibility for our expansion plans; short-term contracts let us capture high-pricing windows and earn premium returns. We weigh each deal individually, evaluating the full economics of an investment: revenue generated, operating costs, financing costs. Strong demand gives us the option to build sustainable value. Our existing prior-generation compute clusters are also in hot demand: in Q3, Ampere and Hopper architecture compute rose 30%50% sequentially, above original contract prices. Different workloads make different trade-offs between compute performance and cost, and older-generation compute remains economically attractive, especially when connected to a stable, scaled cloud platform. We are signing multi-year renewals with terms extending beyond GPU depreciable lives; later in the contract, pricing and operating costs remain stable while depreciation steps down to zero, pushing margins higher. More than 130,000 GPU contracts expire in 2027, and another roughly 170,000 in 2028. There is already a large queue of customers competing for this compute. At current pricing, 2027 renewals will generate close to US$4 billion in incremental annualized revenue without additional capex; combined with compute expiring in 2028, incremental annualized revenue approaches US$10 billion. In addition, if industry-wide data center construction slows, existing compute supply will become scarce, potentially extending the high-GPU-price cycle. In that scenario, first-mover advantage matters: our existing compute base will complete debt repayment and contract expiry earlier than most peers in the market. Whatever the market direction, we are well positioned: if industry expansion accelerates, we grow with it; if supply comes online later, the value of our debt-free installed base rises further. Of course, our business plan does not depend on this scenario; it is grounded in locked-in contracted demand, success-based capex, and new orders now materializing. Last quarter I noted that model training, inference, evaluation, and iteration are forming a continuous AI development loop: online applications feed back their strengths and weaknesses, developers iterate models based on that feedback, test new versions, and redeploy optimized applications. This process requires both specialized software to support each round of experimentation and highly stable cloud infrastructure across the application lifecycle. This is precisely CoreWeave's core advantage. We built a unified platform that orchestrates hundreds of thousands of GPUs into a high-performance full-stack cloud, integrating networking, cooling, orchestration, and observability. This is why customers entrust us with their core workloads. In September we launched CoreWeave Forge, integrating the full set of AI development capabilities into a unified environment. We open products and services honed with leading AI labs to enterprise customers, enabling enterprises to build their own AI on their own expertise and data. 70% of new cloud customers in Q3 were already Forge users. Wave is a typical example: they first used our development tools for model testing and now train autonomous driving world models on the CoreWeave cloud. The Chan Zuckerberg Initiative uses our compute for computational biology research; Hippocratic AI trains and deploys agents for patient communication and triage, advancing disease research and helping patients access care information and support. In engineering, we work with organizations including Nissan and the Aston Martin F1 team to shorten simulation testing cycles and reduce the data required to reach target accuracy. The long-term vision of physical AI is to help engineers develop better systems and bring practical capabilities into the real world faster. MasterClass uses Forge to evaluate instructional agent conversations and optimize the course experience running on the CoreWeave cloud. The ultimate goal is personalized learning, delivering support when users need it. These customers are innovating in science, clinical care, engineering, and education, combining human exploration with AI. CoreWeave provides the platform to support R&D at scale and speed. This is also why our AI cloud platform matters as much as the underlying compute: whether a good idea becomes a usable application depends heavily on whether developers can experiment at low cost, evaluate results, and launch services reliably. More than half of new customers this year came from outside AI labs and the tech industry; 10% of Fortune 500 companies are already CoreWeave customers. The number of customers with individual commitments of more than US$100 million grew more than 2.5x year-over-year. As customers continue to identify AI use cases, their demands on the platform become more varied, broadening our opportunity and deepening customer stickiness. As of the end of 2027, qualified demand we have received is 10x currently available compute; demand for 2028 is already several times the compute that will be available by then. Before formal external release, our active compute power has already reached 1.7 GW. We maintain our expectation of exceeding 1.85 GW of active compute by year-end. Sites under construction with full permits are progressing steadily, and we judge there to be no material risk to our 2027 delivery targets. This year, international expansion has become a strategic priority. Of the 4.6 GW of contracted compute, more than 1.3 GW is outside the United States, spanning Canada, Europe, and Asia, in locations where customers have requested capacity. In Q3 we entered new markets in Indonesia and India while expanding in Denmark and Spain. Our global AI backbone connects compute across regions, delivering consistent performance, software capabilities, and reliability wherever customer compute is deployed. Our core goal is to deliver the compute customers have contracted and earn attractive returns on every deployment. Before handing over to Nitin, let me return to my opening point: CoreWeave's goal is to build a durable, long-term company. We have achieved scaled, profitable growth. Our financing strategy will continue to tie capital to committed customer contracts and, backed by stable returns, access capital from diverse channels. This mechanism lets us flex the pace of expansion based on customer demand, the regulatory environment, and capital market conditions. Our compute delivery is ahead of plan, and the value of our compute continues to rise. Even as early installed compute renews at a premium, we continue to earn the trust of host communities and customers. Our cloud platform balances price, performance, and reliability, with the software and development tools enterprises need to execute their AI strategies. The company has reached an inflection point, and the durability of the business continues to strengthen. Fundamentals continue to improve, and we are confident in serving customers, shareholders, and partner communities over the long term. I'll now hand over to Nitin. Question-and-Answer Session Operator The Q&A session will now begin. Each questioner is limited to one main question plus one follow-up. The first question comes from Gabriela Borges of Goldman Sachs Group, Inc. Gabriela, your line is open. Please go ahead. Gabriela Borges (Goldman Sachs Group, Inc. Analyst) Good afternoon, Mike, and thank you for all the detail. Let me start with a macro industry question: your supply-demand data is very strong. But investors broadly share one concern: a large amount of new compute supply will come online over the next 12 years, while the pace of demand growth is hard to predict precisely. There has been an ongoing debate: will supply-demand balance arrive next year, or only after five years? My question is: if the industry enters a phase of supply-demand balance or even oversupply, how would the CoreWeave business model be affected? Thank you. Michael Intrator (Chief Executive Officer) Thank you for the question, and thanks to everyone for joining this quarterly call. This quarter's results were very strong, and we're glad to share them. On your question: for the past several years there has been a mainstream view in the market that compute infrastructure would come online faster than demand grows. But what we observeand what other compute providers and compute buyers can also seeis that compute demand is growing faster than infrastructure can be built and delivered. Rising margins on every new contract confirm this. Yes, a lot of compute will come online. But AI technology is penetrating the entire economy faster than compute supply, cluster buildout, and model training capacity can expand. Our 2027 compute is essentially sold out, and demand is extremely strong. 2028 is also rich with demand. Not all 2028 compute is contracted, but all the compute resources we bring to market are subscribed. A large amount of demand is waiting for compute supply to come online. Gabriela Borges (Goldman Sachs Group, Inc. Analyst) Understood. A follow-up: I'm focused on the updated US$2025 million per megawatt pricing. For contracts already signed that lock in revenue per megawatt, how do you manage the mismatch risk between locked-in revenue pricing and costs? Between signing a customer contract and compute going live, costs may fluctuate or even rise. How do you manage the tension from that time lag? Thank you for the update. Operator The next question comes from Brad Zelnick of Deutsche Bank Aktiengesellschaft. Brad, your line is open. Please go ahead. Brad Zelnick (Deutsche Bank Aktiengesellschaft Analyst) Thank you; there was a lot of disclosure this quarter, and much of it addressed our concerns. I'd like to understand: will recent concerns in the industry credit markets change your contracting strategy for new compute after 2027? Including compute delivery timelines, customer mix, contract duration, and financing matching pace. Michael Intrator (Chief Executive Officer) Thank you for the question. Let me use this to clarify: we have not yet brought 2028 compute to market at scale, because we want more clarity on the power supply and compute planning environment. We have multiple levers to ensure supply pace aligns with demand signals. You can see this in our timeline for compute to be brought online and in our demand guidance. We don't lock in forward compute too early; instead, we align our capacity roadmap with customers' development plans as much as possible. One point I keep making is maintaining a broad, diversified customer pool, which gives us very precise demand signals: who needs compute, when, and how much, across multiple industries, including AI labs, the Siasun Robot&Automation track, and S&P 500 companies. This information is an important basis for setting forward compute contracts. Brad Zelnick (Deutsche Bank Aktiengesellschaft Analyst) Thanks. Mike, one follow-up. Many emerging AI cloud providers are now experiencing growing pains. Could that create opportunities for CoreWeave? For example, customer migration, industry consolidation, or pricing power? Michael Intrator (Chief Executive Officer) That's a great question too. We have already achieved scale and are well positioned in a volatile market. Ultimately, AI cloud consists of two layers: a software layer for managing infrastructure, and the compute hardware itself. We believe the software layer we've built is best-in-class. Compute hardware is a scale game; players must reach sufficient volume to compete effectively. We understood this from the start, and our business model is long-term contracts paired with a diversified buyer base. Now, as we continue to scale, our software products are competitive and our hardware can be supplied at scale. Together, this lets us be selective about customers, choose the right partners, and accompany them as they begin or continue expanding their AI buildout. Brad Zelnick (Deutsche Bank Aktiengesellschaft Analyst) Thank you very much. Operator The next question comes from Samik Chatterjee of JPMorgan. Samik, please go ahead. Samik Chatterjee (JPMorgan Analyst) Thank you. Mike, Nitin, margins were very strong this quarter, and guidance points to further improvement in Q4. I'd like to understand what role short-term contracts played in this margin expansion. You also mentioned that compute brought online early within the quarter lifts profit. How should we break down the incremental profit in Q3 and the Q4 guidance between execution and short-term contract contribution? I have a follow-up. Michael Intrator (Chief Executive Officer) Internally, we continually weigh: long-term contracts provide business stability, short-term contracts lift margins, and value-added services like managed inference also contribute. All three businesses are growing rapidly and together support the company's expansion. Our ability to take on more short-term orders and managed inference demand is built on the compute foundation established by long-term contracts. So going forward we will continue to optimize compute allocation, deploying scarce compute resources across these three opportunity areas. Samik Chatterjee (JPMorgan Analyst) Follow-up: after the election, the market is very focused on what changes in the U.S. regulatory environment and regulatory pressure might bring. Could you share your view, especially regarding your long-term plan for 8 GW by 2030, and the significance of international expansion in the current macro environment? Thank you. Michael Intrator (Chief Executive Officer) The conclusion first: we do not expect major changes in the regulatory landscape to have a material impact. For each data center corresponding to the backlog, permitting and construction progress are relatively solid, and we are confident in the pipeline. Regulatory changes essentially set requirements for how construction is done. We have always insisted on responsible development, engaging early and collaboratively with host communities and maintaining good neighborly relations locallythat has always been our approach. So we are confident in continuing to expand under the new policy environment. As you said, CoreWeave has made global expansion a priority, with more than 1.3 GW of compute added outside the United States. We have landed in Indonesia, India, Spain, Denmark, and Canada. These regions both meet customers' needs for diversified compute placement and let us build new infrastructure in communities that welcome compute projects, diversifying placement risk. Operator The next question comes from Karl Keirstead of UBS Group AG. Karl, please go ahead. Karl Keirstead (UBS Group AG Analyst) Thank you. Mike, you mentioned the goal of customer diversification. Recently, several frontier AI labs have signaled plans to build their own compute and reduce leasing. We'd like to understand how CoreWeave will respond, and what proportion of new compute you plan to allocate to non-lab customers, i.e., enterprise customers. Could you describe this customer mix shift plan and, in light of Q3 backlog growth, the customer composition within the backlog? Michael Intrator (Chief Executive Officer) Whatever part of the business we discuss, our core approach is to reduce risks such as customer concentration and data center concentration. One data point is compelling: 10% of Fortune 500 companies are already CoreWeave customers, and financial services accounts for 10% of our backlog. Diversification is already real. More than half of new customers this year are outside the AI track. From the start we understood that a cloud business must serve the long tail of enterprise customers. They need a compute foundation and development software like Forge to build their own AI tools. We will maintain long-term partnerships with AI labs, large public cloud providers, and social media companies, while aligning the whole company around expanding other types of customers. Nitin Karl, let me add: Forge is an excellent vehicle for us to reach enterprise customers, and we're already seeing clear growth and good results. 70% of new cloud customers in Q3 were already Forge users. Forge is the entry point for enterprise customers to begin deploying compute, helping us open up the enterprise market. Karl Keirstead (UBS Group AG Analyst) Thank you both. I'll leave it there and leave time for others. Operator The next question comes from Ben Reitzes of Melius Research. Ben, please go ahead. Ben Reitzes (Melius Research) Mike and team, thanks for the update. You mentioned recent communication with the newly elected governor of Texas; could you describe that communication and our exposure in Texas? You said you don't expect delivery delayswhat supports that judgment? And one brief follow-up. Michael Intrator (Chief Executive Officer) In Texas, the vast majority of our compute is either already live or fully permitted and under construction. That is the basis for our judgment that the data centers corresponding to the backlog will not be affected. At the same time, we believe it is important to establish communication channels with the new governor and regulators, just as we have maintained communication with local decision-makers in the past. The core message of my letter was: we remain nonpartisan and want to continue advancing compute buildout; we are a compliant, reliable operator willing to work with local governments and communities, and we look forward to continued deployment. Ben Reitzes (Melius Research) Thanks, Mike. A follow-up: I'd like to understand the arrival and ramp progress of Vera Rubin GPUs, and the situation with Vera CPU racks. How important are these products to the recent guidance raise and to next year's recurring revenue? Michael Intrator (Chief Executive Officer) CoreWeave is proud to be the world's first cloud provider to bring Vera Rubin online, delivering to Cognition, which is also the world's first customer to put Vera Rubin into production. This is significant for Cognition and for the industry. It is a brand-new, extremely high-performance architecture that will become one of the core hardware platforms for AI's next phase. We also very much look forward to standalone deployment of Vera CPU. It has enormous application space in the AI infrastructure track, and related demand is growing rapidly. This is a key focus for us over the next few years and for the industry as a whole, and we are honored to be first to bring it online and into production. Nitin Ben, let me answer your second question on the drivers of this quarter's results and the raised Q4 guidance. Three core points: first, contracted compute continues to convert rapidly into active, revenue-generating compute; second, the demand and pricing environment continues to strengthen, with investment returns continuing to improve; third, platform penetration continues to rise, with customers across industries using more of our products and services. Together these drove this quarter's results and will continue to support the raised Q4 guidance. Ben Reitzes (Melius Research) Thanks, Nitin. Operator The next question comes from Amit Daryanani of Evercore ISI. Amit, please go ahead. Amit Daryanani (Evercore ISI) Thank you, and thanks for the detailed discussion. I have two brief questions. First, you mentioned that at equal hardware utilization, short-term contracts run at an annualized US$40 million per megawatt versus US$2025 million for long-term contracts. After financing costs, how do the returns compare? Also, Mike, could you say what proportion of compute you are willing to place in short-term contracts? Nitin As Mike mentioned earlier, our business has three pillars: long-term contracts providing stable cash flow, high-value-added services such as managed inference, and short-term contracts. We evaluate each opportunity internally, modeling returns, compute financing costs, and the supply-demand environment to allocate compute. In terms of payback, short-term contracts pay back faster, roughly 1618 months; at current long-term contract pricing, long-term contract hardware capex payback is in the 12 year range. Michael Intrator (Chief Executive Officer) One more point worth noting: early contracts are progressively expiring130,000 GPUs in 2027 and another 170,000 GPUs in 2028. These GPUs are very well suited to short-term orders. And by the time the original contracts end, the loans on this infrastructure will have been fully repaid. That is the design of the delayed-draw loans: fully amortized, significantly enhancing cash flow. Amit Daryanani (Evercore ISI) Understood. Following up on that, you mentioned 2027 renewals bringing US$4 billion in incremental revenue and close to US$10 billion in 2028. Is that number gross revenue or incremental premium? Are these contracts already renewed and locked in, or expectations based on current prices? These numbers look very substantial, and EBITDA margins on this compute would also be very high. Nitin This is not included in the currently locked backlog. The backlog is compute already contracted; this number is the revenue that could be generated if existing compute is re-contracted at current market prices when it expires, without additional capex. Michael Intrator (Chief Executive Officer) To clarify further: of the US$9 billion in prepayments and deposits we mentioned, more than US$3 billion comes from renewals of existing compute whose debt has been fully repaid. This cash inflow carries no debt service, and the corresponding infrastructure has completed the full cycle of its original contract and delayed-draw loan. That is why it is so valuable. Amit Daryanani (Evercore ISI) Clear, thank you very much. Operator That concludes the Q&A session. I will now hand back to Michael Intrator for closing remarks. Michael Intrator (Chief Executive Officer) Before closing, I want to thank our team, customers, and partners for trusting CoreWeave and advancing this work with us. None of this would be possible without you. I am proud of the progress we have made and of what we are setting out to do nextwe are building the essential cloud platform for the AI era. Thank you for joining today and for your support; we will update you again next quarter. Operator This concludes today's call. Thank you for joining; you may disconnect.