CoreWeave, Inc. (CRWV) Earnings Call Transcript

Thanks, everyone. Good afternoon, and thanks for joining us on CoreWeave's conference call for the second quarter of 2026 earnings. Here with us today to go over what we reported are our CEO, Mike Intrator, and our CFO, Nitin Agrawal. Before we dive in, I'd like to use this moment to point out that remarks we make today will contain forward-looking statements. Because of factors laid out in today's earnings press release, along with our quarterly report on Form 10-Q that will be filed with the SEC, actual results could differ materially from what those forward-looking statements contemplate. Every forward-looking statement we make on this call rests on assumptions as of today, and we take on no obligation to update these statements because of new information or future events. On this call, we'll present both GAAP and certain non-GAAP financial measures. Today's earnings press release includes a reconciliation of GAAP to non-GAAP measures. You can find the earnings press release and an accompanying investor presentation on our website at investors.coreweave.com. A replay of this call will also be available on our investor relations website. I'd like to turn the call over to Mike.

Hello, everyone, and thanks for being here this afternoon. CoreWeave just wrapped an outstanding second quarter. Across the board we came in ahead of what we'd planned, and the operating leverage we've been steadily constructing is finally showing up plainly in our numbers. Tremendous work throughout the company delivered all-time-high financial results, fast expansion of capacity, growing demand from customers, and ongoing innovation on our platform. We posted record revenue of $2.6 billion, which is a 112% increase compared with the same period last year, and our revenue backlog climbed to $104 billion, even as enterprise adoption grew at a swift pace. That total leaves out over $25 billion in net new customer commitments that came in during the first few weeks of Q3. We stayed the course on our power strategy. Active power hit 1.5 GW — we brought on close to 500 MW, a quarterly record for us, and that's over three times what it was a year ago. We're still solidly positioned to hit 8 GW or more by 2030. Adjusted operating income climbed to $128 million, and margins widened considerably, since greater scale keeps converting into operating leverage. We also continued expanding what our technology stack can do, rolling out seven new AI platform capabilities and notching several industry firsts, which let customers build, deploy, and run AI more quickly and at larger scale. This momentum reflects the work of the whole organization, and the business keeps getting stronger. During Q2, the deals we closed with customers carried contribution margins that we anticipate will land 5 percentage points to 10 percentage points higher than what was brought in over the past few quarters. This isn't just a set of milestones. It's proof that the AI sector is unfolding consistent with the beliefs that CoreWeave was founded upon. Our view is that the AI age has arrived and will eventually reach every corner of the worldwide economy; that transformation is underway at every organization, opening a chance to reinvent existing markets and build brand-new ones. And that the future belongs to the trailblazers who grasp that chance — both AI-native firms operating at the frontier and the agents of change within long-standing companies. Being able to learn quickly and iterate rapidly is simply a requirement now if you want to lead in AI. In our view, the companies breaking new ground need a different type of platform—one capable of unlocking what AI can do on a large scale. Those convictions form the core assumptions behind our approach. They shape where our products are headed, how we spend our capital, who we partner with, and, in the end, the way we deliver value to our customers. Today, I'd like to walk through how that vision is showing up across four things. First, expanding both demand and the number of customers who adopt us. Second, the ongoing tools for building AI that we offer through our platform. Third, the speed and cost advantages that come from our AI-native design. Fourth, the strength and supply chain base that will fuel growth for years ahead. Every organization is being reshaped by AI. The argument over how much demand will exist down the road for AI cloud infrastructure probably isn't going away. Certainty, though, comes from watching what our customers actually do. We see demand growing stronger as adoption spreads wider — different industries, different parts of the world, different types of workloads, and different GPU architecture generations. AI has moved past being confined to cutting-edge research facilities. These days it's embedded in software, industrial operations, financial markets, business workflows, and national security missions. That breadth is visible in our backlog, in newly inked commitments, and in the utilization and pricing picture across the entire platform. With our Blackwell and Vera Rubin SKUs, both pricing and margins are hitting all-time highs, while for older-generation SKUs, pricing remains at or above where it stood several years ago. Basically, when it comes to the capacity we can bring on in the near term, it's already completely claimed. That turns into contracts that get signed, and the terms keep getting better while the pool of customers keeps broadening, which means CoreWeave is set up to gain share in the years ahead. No organization out there is immune to the pressure to rethink what it can actually pull off. What AI does isn't just make existing processes faster — it gives companies the ability to rework their core functions, roll out brand-new offerings, and enter markets that didn't exist before. A great example of this is Caterpillar. Together, we're going to bring NVIDIA's Vera Rubin platform online so it can drive Caterpillar's physical AI training and inference on an industrial scale. CoreWeave's AI cloud infrastructure will serve as its data factory, and with that, Caterpillar will train models built for specific purposes that make autonomous construction machinery smarter and more productive. CoreWeave is finding that life sciences is becoming yet another important area for growth. Organizations tackling some of the world's toughest scientific challenges are increasingly choosing our platform. Isomorphic Labs has come aboard as a new client, and we're excited to support their goal of curing all disease. Financial services remains a significant growth category for us as well. Flow Traders and IMC have just been added to our growing list of systematic trading companies; these clients depend on CoreWeave to develop and launch the next wave of AI models used in quantitative trading. They choose us because we can coordinate high-performance workloads at the speed, reliability, and efficiency that such applications require. Regarding government work, teaming up with Leidos marks a major milestone for CoreWeave Federal's expansion. Together, both sides aim to accelerate the rollout of protected AI tools for defense, national security, and intelligence operations. A fresh wave of pioneers is building what comes next in AI. Some of them are AI-native firms operating right at the cutting edge. Others are agents of change inside long-standing corporations who are eager to upend the existing order. CoreWeave supports both types. Companies such as Decart rely on our platform to develop Oasis 3, which is the very first world model accessible through an API for physical AI. IBM turns to CoreWeave for safe experimentation with reinforcement learning, agent tool use, and model evaluation. Through Monolith, the field engineers we specialize in collaborate in person with clients including Nissan and ZF, accelerating the creation of AI applications built for enterprise use. Interest is no longer limited to our infrastructure alone. With CoreWeave Omni, we're encountering considerable enthusiasm from sovereign, enterprise, and cloud clients in equal measure; during the last several weeks, we closed our inaugural agreement, and it will begin expanding in 2027. These cases differ across sectors and applications, but the underlying trend remains consistent. AI has moved past the experimentation stage and into the core of business operations, and companies acting decisively are pulling ahead. Simply launching something is no longer the finish line. With AI now operating in production settings, the methods used to create applications are evolving, and the winners will be whoever can learn and iterate most quickly. Over the past few years, plenty of companies viewed a model as something to hand off—train it, ship it, then walk away. That's not how enterprises work anymore. Today, training, inference, evaluation, and improvement make up one unbroken cycle. Models and agents running in production produce data from the real world. Evaluation draws on that data, which sparks fresh experiments, and those experiments enhance the model or application prior to another production release. The cycle keeps turning, and capability builds on itself as time goes on. This change reshapes both the shape of demand and the economics behind AI at a fundamental level. Compute is no longer something you need once, front-loaded at the start of a model's existence. Instead it turns into a continuing need that expands alongside each application running in production and each round of enhancement. Our AI-native platform was designed precisely for this. It covers state-of-the-art cloud infrastructure, a managed inference business that's expanding quickly, top-tier developer tooling and agent solutions, and an orchestration and observability layer that leads its class, all powered by CoreWeave Mission Control. Taken together, these strengths provide customers with a single unified environment.

When customers want to roll out applications, they do it via CoreWeave inference, either with our models or with versions they've tailored using our serverless features. They can watch how things perform through Weights & Biases, test applications once live, try out fresh models, sharpen results by way of serverless reinforcement learning or CoreWeave Sandboxes, and confirm any adjustment meets standards for quality, performance, and cost prior to putting it back into production. Our AI research and iteration agent, CoreWeave ARIA, speeds this up even more—it reviews thousands of evaluation runs, brings insights to the surface within minutes, and suggests what experiment to try next. This lets customers shrink the gap from idea to experiment to production improvement at light speed; during Q2, we rolled out seven new AI platform capabilities and notched multiple industry firsts. We created these innovations hand in hand with our customers and partners to address actual production hurdles; that explains the strong adoption we're witnessing. Only this week, we crossed 1 billion model training runs tracked on our platform. That figure represents millions of experiments, thousands of research breakthroughs, and an expanding community of engineers, researchers, and organizations crafting the next generation of AI. A wider set of customers than our core cloud is embracing our AI development services, which come with higher margins. This is turning out to be a natural path for customer expansion, since the developers creating AI applications now are the AI cloud infrastructure customers of tomorrow. By serving them early, we build relationships that naturally grow as AI workloads scale. Since its launch a few months ago, our managed inference platform has seen explosive growth, held back only by our near-term capacity. Through serverless offerings and dedicated deployments alike, CoreWeave monetizes tokens while letting customers stay flexible in how they use our platform. Firms like Grammarly and You.com are shifting from experimentation to genuine production traffic, operating AI coding agents, serving their own fine-tuned models, and deploying open weight models at scale. Customers shouldn't need to sacrifice speed for cost, and with CoreWeave, they don't; they pick our platform because of the mix of total cost of ownership, quality, breadth of service, and performance; our steady leadership in cost per token and speed to first token on leaderboards—such as Artificial Analysis for open source models, among them Kimi K2.6, Kimi K 2.7 Code, GLM- 5.2, and MiniMax M3—shows exactly that. That leadership translates straight into revenue. Over the few months since launch, booked ARR for our managed inference platform has climbed from $1 million to over $100 million. We anticipate exiting 2026 with no less than $250 million of managed inference ARR. Pioneers require a platform of a different sort. You can't support the continuous AI life cycle just by tacking GPUs onto a general purpose cloud. It calls for a fresh approach spanning power, cooling, and rack design all the way through networking, orchestration, observability, developer tools, and managed services. That's the reason CoreWeave is purpose-built for AI. Our platform stands alone in its depth, breadth, and technical capability. During Q2, we were the initial cloud provider to stand up and validate NVIDIA's Vera Rubin NVL72, drawing on what we've pioneered in software-defined liquid cooling and rack management, which continues our pattern of reaching the market ahead of everyone else. On top of that, we established new MLPerf records in both training and inference using open source models on the NVIDIA Grace Blackwell platform, and through our testing we reached the lowest cost per token for inference; even so, strong performance on its own doesn't get you all the way there. What customers want is enterprise-grade security and observability, reliability, and economics that make sense. Per Signal65, CoreWeave provides total cost of ownership that's estimated to be as much as 47% lower than the typical hyperscaler. Customers also need our platform, which pulls these capabilities together alongside a wider portfolio of storage, CPU, and networking services spanning a distributed global data center footprint. Gartner, in July, called CoreWeave a visionary in its 2026 Magic Quadrant for Cloud AI Infrastructure. That acknowledgment, we believe, gives further independent confirmation of how we go about building the AI cloud; these accomplishments aren't just standalone technical wins. For customers, they convert directly into quicker deployment, greater utilization, improved application performance, and reduced cost. Combining product depth with best-in-class performance, quality, and TCO that leads the market is a formula that wins for our customers and for CoreWeave. CoreWeave is the foundation for AI at scale. This market demands a foundation at a scale unlike anything we've seen before; that requires locking in power, sites, cooling, hardware, storage, networking, and supply chain inputs far in advance of when they're needed, and running all of it as a single integrated system. As I mentioned at the start of the call, Q2 closed with 1.5 GW of active power, and we added close to 500 MW during just that quarter. For some context, the power we brought on in Q2 exceeds what any single neo cloud runs in total today, based on third-party estimates. Importantly, our scale is paying off, and the math only improves going forward. Every new deployment is landing on top of an installed base far larger than it was just one quarter ago. As that base expands, each new build represents a smaller slice of the whole, while contracted revenue from deployments already in place stays put. That's how we're turning scale into operating leverage. It explains why margins expanded in Q2 and why we anticipate they'll keep expanding sequentially through Q3 and Q4. We're also locking down what's needed to support growth across a multi-year horizon. Contracted power reached 3.7 GW in Q2. Since the quarter ended, we've brought on approximately 500 MW, which puts contracted power at 4.2 GW as of today. Those numbers don't include over 1.5 GW of additional potential power from powered land we've gathered, options we hold to expand at existing sites, and LOIs we've signed. Several of our first self-builds are already progressing well, among them our first site, which we expect to come online later this year. Owning the land underneath our powered sites is the base layer that lets us integrate more deeply in the vertical direction, which hands us tighter control over operations and backs stronger margins over the long haul; on top of that, we're pushing into markets worldwide and have signed contracts for over 1 GW of power beyond the U.S., and that includes our recent entry into the APAC region through 360 MW in Indonesia, which should start coming online in roughly 18 months. As we serve customers in the places where they and their end users run their businesses, we anticipate international markets turning into a key growth engine. Taken together, we have strong line of sight to our goal of no less than 8 GW by 2030. For years ahead, we anticipate demand will outstrip supply by a meaningful amount. Within such conditions, securing power is merely one piece of the puzzle. Equally critical is every single component that's needed to bring the AI cloud to scale. Leveraging our tight collaborations with NVIDIA alongside our OEM and ODM partners, the long-term deal we struck recently with Solidigm serves as one example of the way we're reducing risk around obtaining the essential inputs that customer service demands. Technology, capacity, vertical integration, supply chain, and worldwide expansion — all of our investments stem from one shared vision. AI keeps spreading further into everything. Those pioneers moving quickest will take the lead, and they'll need a platform that can handle ongoing learning and deployment on a scale never seen before.

Prior to handing things over to Nitin, I'd like to emphasize something: we're heading into the back half of this year carrying greater momentum than we've ever had as a company. Every sector is being transformed by AI. Those trailblazing builders out there require far more than just compute — which explains why they choose CoreWeave, an AI cloud built to support the entire AI life cycle, handling any workload ranging from frontier training all the way to inference that scales rapidly. Across industries, regions, and hardware generations, demand keeps outpacing what we can supply; we're operating at remarkable scale for a varied customer base, with operating leverage that keeps improving and clear sight into the power and essential components needed to keep growing for years ahead. What lies in front of us represents a once-in-a-generation opportunity. For AI, CoreWeave is the indispensable cloud. Our belief in this strategy has never been firmer, and the way we execute keeps confirming it. With that said, over to you, Nitin.

Thanks, Mike, and good afternoon, everyone. For CoreWeave, the second quarter was an outstanding one, defined by powerful customer demand, a substantial scaling up of our live capacity, and steady progress on both our product and financing roadmaps. Maybe most significantly, Q2 was when margins turned a corner, rising quarter over quarter in the way we had described across the last several quarters. Before getting into the results, I'd like to take a few moments to discuss how demand dynamics are shifting in today's environment, along with what that means for cash flows and for the worth of our fast-growing infrastructure footprint. Demand for CoreWeave Cloud is still extraordinarily robust throughout our entire customer base, with multiple customers seeking each GPU we bring online. We are taking a disciplined approach to allocating our limited cloud capacity. We are giving priority to opportunities of strategic importance, bringing in new customers while strengthening long-standing existing relationships and producing appealing returns that, as Mike mentioned, are growing even further. Our AI products and services beyond GPUs are also continuing to scale up considerably as customers concentrate their spending with us. As of Q2, these margin-accretive businesses — storage, CPU, networking, and software — already surpass $400 million of ARR. We anticipate that they will keep growing quickly. To put it plainly, customers are spending more with CoreWeave because they recognize the added value we deliver. That improvement in operating margin came before the pricing adjustments we made in July, when we raised prices by about 25% across SKUs in response to the current demand picture and the growing returns our customers are getting from what they invest in the CoreWeave platform as they shift toward inference. We're also passing through higher component costs. On how this reaches cash flows, as we've noted before, a typical five-year contract carries robust unit economics that keep improving over the course of its term. Even so, those economics don't appear in a uniform way. You have to spend the money first, and that spending shows up largely as CapEx. To pay for building it out, you need some combination of borrowing, money customers pay ahead of time, and other funding from the corporate level. Once that cluster is handed over, the revenue you've contracted starts to rise, becoming dependable and producing very strong cash flow. Every one of these elements goes into our assessment of the margins we expect prior to putting pen to paper on a contract. What we get from deploying it are returns that are quite attractive — they fully pay off the asset-level debt that covered the CapEx, and on top of that they generate a meaningful amount of extra free cash flow. After a first contract ends, the cluster gives us no leverage whatsoever, and we can freely recontract that cloud infrastructure or offer it to the market; we will already have made an attractive return before there is any possibility of earning more from the cloud infrastructure. Every resale or renewal adds to the returns we already generated during the initial term. What we see right now is that the benefit from recontracting is real, because we are still mostly sold out of older generations of NVIDIA GPUs as well as the current SKUs. When the fleets from our earlier generations come off those initial contracts, there's potential for them to produce solid returns over the years that follow; we're observing that with both our Ampere and Hopper fleets. Take one recent illustration: we signed an A100 deal that runs through 2029 at an appealing price point. Keep in mind, that SKU first came to market in 2020. Clusters built on prior architecture generations provide installed, energized, production-grade compute that's already operating at scale. Customers get a proven ROI from them. In a marketplace where fresh capacity is constrained on supply and expenses keep climbing, AI cloud infrastructure that's in production is a scarce and valuable asset. Even though our business model doesn't hinge on renewals after a customer's first term wraps up, we keep seeing customers stay with us longer and pay more, which could mean significant extra gains. Those favorable conditions make us more confident than ever about the long-term returns on what we've put into our products and capacity, and that's thanks to our top-tier AI cloud offerings; now let's look at how we did in Q2. We brought in $2.6 billion in revenue for the quarter, representing 112% growth compared to the same period last year and 24% growth from the prior quarter, driven by ongoing solid performance and demand from customers for CoreWeave's AI cloud platform. At the close of the quarter, our revenue backlog reached $104 billion, a 246% increase year-over-year. Like Mike said, that number doesn't count the more than $25 billion in net new customer commitments that came in during the early part of Q3. When you look at the backlog we already had, more than half of it relates to a contract where delivery to the customer is already underway. We expect that by the end of this year, it will represent over two-thirds of our Q2 backlog. Our operating expenses came to $2.6 billion in the second quarter, and that includes $165 million in stock-based compensation expense. Those operating expenses grew because we were scaling our active power and converting backlog into revenue. That drove corresponding growth in our cost of revenue as well as in technology and infrastructure spend. In addition, the rise in sales and marketing came from putting more resources into our go-to-market team as we broaden our customer mix and expand into additional products and markets. G&A increased because of personnel expenses needed to back our growth, though it still grew more slowly than revenue. For Q2, adjusted EBITDA came to $1.5 billion, compared with $753 million in Q2 of 2025 — a twofold increase year-over-year. Our adjusted EBITDA margin stood at 59%. Adjusted operating income in Q2 was $128 million, compared with $200 million in Q2 of 2025 and up from $21 million in the prior quarter, landing well above the top end of our guidance as operating leverage kicks in with scale. Adjusted operating margin was 5%. As our scale increased, margins grew wider, even while we kept absorbing substantial costs tied to the ramp. For Q2, the net loss came to $626 million, compared with a $290 million net loss in Q2 of 2025. Interest expense in Q2 totaled $640 million, up from $267 million in Q2 of 2025. That rise stems from taking on more debt to fund the ongoing build-out of our infrastructure and to meet the customer commitments we've contracted. Despite posting a net loss, we recorded an income tax provision anyway, which reflects a valuation allowance applied against net deferred tax assets. As we noted last quarter, if no material discrete items arise and conditions remain unchanged, our tax rate should hold approximately flat throughout 2026. Our adjusted net loss came in at $567 million for Q2, versus a net loss of $130 million during Q2 of 2025. Now onto capital expenditures. Q2 CapEx came to $9.4 billion, a bit above the top end of the range we had guided. The elevated CapEx this quarter is a reflection of customer deliveries picking up speed. Construction in progress, or CIP, rose to $11.9 billion, up from $9.6 billion on a quarter-over-quarter basis, which signals how much PP&E we anticipate deploying early in Q3 given the substantial amount of power that came to us very late in Q2. In fact, during June we added over 300 MW of active power, which means June on its own was bigger than any full quarter in our company's history. As Mike pointed out, the global supply chain is still complicated; we keep working through these difficulties with operational discipline and by leaning on our partner relationships, among them new ones such as Solidigm, in order to strategically source the inputs we need. Now shifting to our balance sheet and the strong liquidity we have. As of June 30th, our cash equivalents, restricted cash, and marketable securities totaled more than $6.9 billion.

During Q2, we achieved meaningful strides in fortifying our balance sheet and broadening as well as deepening the ways we can tap capital, bringing in roughly $18 billion through a mix of debt, convertibles, and equity. Among these deals were multiple inaugural moments, such as our first Eurobond, plus our debut delayed draw term loan in the public markets secured by HPC infrastructure. Our latest financing—our second term loan syndicated publicly—represented yet another major milestone, being the first to feature shorter duration customer contracts. That transaction, priced amid one of 2025's most dislocated credit weeks, drew substantial interest. Even with conditions at the time, we opted to close it at full size. Our spreads have since retraced. Maybe most crucially, the deal showed the credit market's rising confidence in the long-term worth of NVIDIA infrastructure operating on CoreWeave Cloud. This financing matters because it opens up our capacity to serve a vital segment of the enterprise market at scale, while letting us speed up the ramp of our managed inference platform and increase our exposure to shorter-dated contracts, which usually carry a higher ASP and margins. Broad and deep investor participation came to these transactions, underscoring the considerable interest we keep seeing in backing CoreWeave's journey. Together, these deals bring us past $32 billion in debt and equity capital raised to date. Across the last year, we've cut our weighted average cost of debt by nearly 300 basis points, equating to about $1.1 billion in annualized interest savings relative to our end of Q2 debt load. Now to guidance. Thanks to sustained strong execution, we now anticipate finishing the year with over 1.85 GW of active power, raised from our prior guidance of more than 1.7 GW; regarding how that plays out in the second half, we anticipate Q3 revenue between $3.45 billion and $3.6 billion. For Q3, we expect adjusted operating income of $200 million- $260 million as margins keep expanding sequentially, hitting low teens by Q4. Q3 interest expense should land in the range of $860 million- $940 million, mirroring growth in our debt balance to fund accelerating deployments. We anticipate CapEx of $11.5 billion- $13.5 billion given the sizable new capacity we'll deliver to customers. Now for the full year. Our disciplined execution and the momentum across our customer base give us confidence to raise full year 2026 revenue guidance to $12.4 billion- $13.2 billion and adjusted operating income to $960 million- $1.15 billion. Because of our higher expectations for capacity delivered to customers this year, along with some of our notable recent wins, we now expect 2026 CapEx in the range of $35 billion- $39 billion. Lastly, we're also lifting our anticipated end of year annualized run rate revenue to $18.5 billion- $19.5 billion. The contracted revenue backlog rests on long-term arrangements and appealing margins, which keeps giving us unmistakable visibility, and our confidence in the goals we've set out remains strong; to wrap up, Q2 showed how robust demand is for CoreWeave's entire technology stack, along with the rigor of how we run the business; we deliberately grew our customer base to serve the coming wave of enterprise AI adoption at margins that keep getting more attractive. To take advantage of the full depth of our AI-native platform, customers are increasing what they spend with CoreWeave; sequential margin expansion stays on schedule for the rest of the year, and on our capital structure we've achieved considerable further progress, bringing down our weighted average cost of capital and locking in the financing needed to back our long-term growth plan. At our yearly developer conference, CoreWeave Fully Connected, this September, we can't wait to see plenty of you, since you'll hear from both our leadership and our customers about how our platform is speeding up AI in production. Thank you. Having said that, let's begin taking questions.

Hey, everyone. I appreciate you fielding our questions, and congratulations on a solid quarter overall. Perhaps just a few items. First, you brought up the chance to renew with shorter agreements as certain earlier contracts reach their end, so you can take advantage of current market pricing. Could you walk us through what you're hearing from customers during those conversations regarding their usual intentions for contract length, and what portion of your deployed equipment might be eligible for renewal in the coming years? If you could give us a feel for how large that opportunity might be. I have one more question. Thanks.

Thanks for asking. I'm glad to take a moment to discuss what was an exceptional quarter for our business, spanning our infrastructure, our software, our solutions, and our sales and agreements with both new and existing clients. Among the most thrilling aspects of what we're starting to grasp—and what we think the market is now shedding real light on—is that earlier generations of infrastructure still hold considerable value for use cases among many AI consumers. We've discussed this for years: the most advanced solutions emerging from NVIDIA, which we incorporate into our cloud and provide to our most demanding customers. That matters greatly for certain cutting-edge use cases; yet inside that setting and ecosystem, countless other use cases can leverage older, later-dated SKUs. The fact that we managed to sell a GPU with a 2020 architecture under a contract priced all the way through 2029 truly sheds light on what lies ahead as this infrastructure exits its contracts.

Samik, when it comes to the capacity due for renewal, that portion of our fleet is quite small, and the ASPs for the older generation are still higher than — or at the same levels as — what we saw roughly 12 months back. The other aspect of this that's really compelling for our business is that when these fleets reach the end of maturity, we get the chance to offer a fantastic product around managed inference for our customers, which, as Mike pointed out, reflects the very quickly changing character of our business — one segment of it — and we anticipate that it will keep expanding fast, and we expect roughly $250 million in ARR by year-end.

Understood. As a second question, I appreciate the information you provided. Regarding that follow-up, could you discuss the supply chain somewhat? Clearly you are handling it quite effectively as you add capacity. With respect to the arrangement you currently have with Solidigm, say, how do you view the possibility of taking a more wide-ranging approach throughout the supply chain, such as entering longer-term agreements, to give yourself confidence in obtaining additional supply, allowing you to keep fulfilling the capacity you intend to bring online? Thank you.

Yes. That is an excellent question. Look, ultimately, my responsibility is making certain this company possesses the capacity to provide the product our customers need. To accomplish that, we must handle a complex supply chain aggressively. That chain encompasses everything from land, power, and shell, through GPUs and networking, through memory, and every part of it is under pressure from the growth and expansion of artificial intelligence. To accomplish that, over the past several years we have established truly long-standing, deep relationships with our ODMs, our OEMs, NVIDIA, the businesses that provide memory to us, every one of them. What we have done is consider what is required to guarantee we can access the infrastructure and the components and the capital we need to provide our products at a price and quality our customers find acceptable. This is among the things simply embedded within CoreWeave's DNA. That is our business. Part of what we do daily is cultivate these relationships and make sure we can access everything required to provide the product, meaning NVIDIA infrastructure delivered via our cloud.

Samik, one point worth highlighting is that the rise in what the CoreWeave Cloud produces has grown faster than the rise in what we're paying for inputs in the supply chain right now. Because of that, you're seeing margins widen, just as Mike mentioned when he talked about typical contribution margins — last quarter those came in 5 percentage points to 10 percentage points above the levels we've seen in recent quarters.

Great. Thank you. Thanks for taking my questions.

Great. Thanks very much, and congratulations on how well you executed. For my first question, I'd like to ask about the managed inference product you offer, which has gotten off to a very impressive beginning. What have you learned so far, and which considerations shape how you plan to assign capacity toward it in the future relative to the take-or-pay arrangements you've traditionally used? I also have a follow-up. Thanks.

Yeah. I appreciate the kind words. It genuinely was an excellent quarter for us. We're quite enthusiastic about it. Look, whenever we consider our offering, we approach it in a comprehensive way. Over the past few years, we've achieved a great deal by concentrating on scale via those long-term take-or-pay agreements. Now that we've reached hyperscale, we recognize the need to expand what we offer — giving clients the products they require, delivering items with better margins, supplying software solutions, and providing CPU access, everything clients must have in order to succeed. Reflecting on what we've taken away from the remarkable, one-of-a-kind expansion of our managed inference product, which jumped from $1 million to $100 million within just one quarter. We see this as a huge chance to deliver our most advanced compute, and it's also a great way for us to tap into GPUs whose contracts are ending, so we can pull the most value out of them for the business as time goes on. The market here is really deep. We believe we hold a built-in edge thanks to how we control the silicon, and we expect to do very well in that space down the road.

Brad, something worth mentioning: yesterday we made an announcement regarding the closing of our DDTL 5.5, and what that demonstrates is that there's tremendous interest from the capital markets, along with support for CoreWeave's product when it comes to underwriting contracts of shorter duration—which is certainly a favorable factor as we examine these markets to back our customers' requirements.

Thanks, Nitin. Thanks, Mike. That really sets up what I wanted to ask next. Regarding the 5%-10% improved margin you're getting on deals you've recently signed, could you break down what's driving that? How much of it comes from deals with shorter duration, versus strong competitive differentiation, or other factors? And more broadly, what are you observing in the market when it comes to pricing? Thank you.

A lot of different elements are at play, and pulling it apart isn't easy. What we provide via the CoreWeave Cloud carries greater worth for our customers than any alternative they might find. The platform's quality, how reliable the infrastructure is, the security, the TCO — every one of those factors plays a role in customers returning to us repeatedly and growing their footprint inside our cloud and infrastructure. Part of this is that they recognize just how much more valuable a given piece of infrastructure becomes when we deliver it. The other part is that a lot of our clients are monetizing their products, which makes them more aggressive in coming to us and ready to pay higher margins, since they require access to the compute that lets them succeed. This phenomenon is happening throughout the infrastructure space, yet it's especially happening inside our ecosystem, and it's very exciting to watch, as the premium product we deliver gets priced in a premium way by those consuming this compute.

Thanks.

Hello, and thank you. This is Irvin Liu, filling in for Amit. I have one question, plus a follow-up. First off, it seems like pricing faces upward pressure from several directions — the greater value delivered to your customers, higher component costs being passed through, and the upcoming recontracting chance. Given all that, should we continue to view roughly $18 billion- $19 billion in ARR as the exit target for 2027?

Yes. We have actually raised the exit ARR figure we shared with you in guidance, and it now stands at $18.5 billion- $19.5 billion for 2026; that number is included in the guidance we gave you.