CoreWeave, Inc. (CRWV) Earnings Call Transcript

Thanks. Good afternoon, everyone, and thanks for joining us on CoreWeave's conference call covering results for the third quarter of 2025. With me today to walk through our results are Mike Intrator, our CEO, and Nitin Agrawal, our CFO. Before we begin, I want to take a moment to remind everyone that today's remarks will contain forward-looking statements. Actual results could differ materially from what those forward-looking statements contemplate. Factors that might cause results to differ materially are described in today's earnings press release as well as in our quarterly report on Form 10-Q that will be filed with the SEC. Any forward-looking statements we make on this call rest on assumptions as of today, and we assume 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. A reconciliation of GAAP to non-GAAP measures appears in today's earnings press release. Today's earnings press release and an accompanying investor presentation can be found on our website at investors.coreweave.com. A replay of this call will also be accessible on our Investor Relations website. And now I'd like to hand the call over to Mike.

Good afternoon, everybody, and thanks for being with us. CoreWeave has turned in another outstanding quarter, demonstrating the quickening momentum behind our business as AI adoption spreads worldwide across industries. We're still operating in an environment where supply is severely limited, with demand for CoreWeave's best-in-class AI cloud platform greatly outstripping the capacity we have available. This relentless appetite from customers makes it obvious that the world's top companies rely on CoreWeave to run their most critical AI workloads. During Q3, we surpassed expectations, posting revenue of $1.4 billion, which is up 134% year-over-year. We brought in more than $25 billion in revenue backlog during the third quarter by itself, which lifted us above $55 billion in revenue backlog as Q3 closed, nearly twice what we had in Q2 and close to 4x year-to-date. On top of that, CoreWeave has hit $50 billion in RPO, quicker than any cloud in history. These figures show how deeply customers believe in CoreWeave, the company they rely on as their essential cloud for artificial intelligence. We keep scaling aggressively even while the industry stays capacity constrained; we grew our active power footprint by 120 megawatts sequentially, reaching approximately 590 megawatts, and expanded our contracted power capacity by over 600 megawatts to 2.9 gigawatts. That positions us nicely for growth ahead, with more than 1 gigawatt of contracted capacity ready to be sold to customers, which we expect will mostly come online during the next 12 to 24 months. In Q3, we signed large-scale compute contracts with many of our biggest customers, among them Meta and OpenAI. Every one of these represents a significant broadening of an existing relationship and a step away from reliance on any single customer; we also deepened our ties with a leading hyperscaler, which became the sixth contract with this customer so far. In fact, 9 of our 10 largest customers have now signed multiple agreements with us. The sole exception is a new customer we brought on in Q3. CoreWeave is the force multiplier that enables pioneers to speed up breakthroughs in AI innovation. These are the most sophisticated AI organizations in the world, and after they encounter the performance, flexibility and reliability of CoreWeave Cloud, they keep expanding with us. That's the strongest validation we could hope for. Our remarkable growth shows just how fast AI adoption is moving past the frontier AI labs and hyperscalers. Wider global demand and our recent big wins are fueling diversification of our revenue base. As an example, the count of customers surpassing $100 million of revenue over the last 12 months tripled year-over-year. AI-native enterprises across sectors are turning to CoreWeave to transform operations and unlock new sources of innovation, productivity and growth. At the leading edge of foundation model development, Poolside picked CoreWeave to power its mission of building artificial general intelligence and enabling the deployment of agents across enterprises, while Periodic Labs is using CoreWeave to push boundaries of scientific discovery and computational research. At the application layer, we brought on AI-native customers such as Jasper, who chose CoreWeave as their cloud partner as they transform the digital marketing landscape; we're also seeing incredible momentum within enterprises. For its work on AI agents in cybersecurity, CrowdStrike picked CoreWeave. Meanwhile, Rakuten relies on our platform to overhaul its visual language models, which helps them gain more transparency, reproducibility, and speed across their AI workloads. Beyond that, we saw additional expansion among a broad set of enterprise clients, among them a top software design platform and a major U.S. telco operator. Our footprint has grown into the public sector as well — a market where performance and security demands are distinctive. Not long ago we introduced CoreWeave Federal, extending our cloud offerings to U.S. government agencies and the defense industrial base. NASA already uses our services to push scientific exploration forward at its Jet Propulsion Lab. It's an honor to support the strengthening of America's AI infrastructure, letting agencies speed up innovation and take on vital missions and our national interests. These recent wins make clear that we are enterprise-ready. As our customer base widens across verticals and geographies, we're thrilled to bring on Jon Jones as our first Chief Revenue Officer. Jon comes to us from AWS, where his role was Global Head of Startups and Venture Capital. He's a powerful addition to our team and will be key in scaling our global revenue organization and fueling expansion through this next phase of growth. Now, turning to our expanding data center footprint, I'd like to say a few words about our previously proposed acquisition of Core Scientific, which was terminated in October. Though the deal was strategically sound for both sides, the valuation their shareholders demanded simply wasn't a price that suited CoreWeave, especially since the transaction's outcome does not in any way hurt our capacity to hit our growth ambitions over the coming years. Rather, we'll keep working closely with Core Scientific on the roughly 590 megawatts of capacity we've already leased. Our disciplined approach to growing our capacity footprint guarantees we satisfy the surging global demand for CoreWeave's cloud services. As I noted, we increased our contracted power capacity to 2.9 gigawatts this quarter, diversifying across size, geography, and developers, which boosts resilience and flexibility throughout our portfolio. As of Q3, no single data center provider accounts for more than approximately 20% of our contracted power portfolio. Last quarter, we brought on 8 new data centers throughout the U.S., bolstering our domestic coverage, with further expansions in progress across Europe, among them a significant new presence in Scotland, being developed alongside the U.K. government. And as we shared over the summer, we've begun self-build projects to further speed up our footprint and give us more operational control. While we see relentless demand for our platform, data center developers throughout the industry are also facing unprecedented supply chain pressure. In our situation, temporary delays tied to a third-party data center developer who has fallen behind are affecting us. This weighs on fourth quarter expectations, which Nitin will cover shortly. That said, the customer impacted by the present hold-ups has consented to modify the delivery timeline and prolong the expiry date. Consequently, we keep the overall worth of the initial contract intact, and the customer retains their capacity across the entire term of the original arrangement, which shows the trust they place in our capability to deliver the highest-performing solutions available on the market. Our technical achievements make us immensely proud, and customers keep informing us that CoreWeave is unquestionably the finest environment for running AI workloads. During the third quarter, we kept rolling out numerous of the first large-scale installations of the GB200s, and at the same time became first to market yet again, on this occasion with the GB300s, which underscores even more our remarkable history of operational excellence. CoreWeave's leadership within the industry cannot be equaled. We stand alone among cloud providers in submitting an MLPerf Inference result for GB300s, establishing the standard for AI performance in real-world conditions. And only a week ago, SemiAnalysis again acknowledged our superiority, granting CoreWeave its top honor, the Platinum ClusterMAX ranking, for a second time, ahead of over 200 providers, among them the hyperscalers and emerging neoclouds. Not a single other cloud has managed to do this even one time.

CoreWeave has now achieved this on two occasions, which drives home once more that no one else occupies the position CoreWeave holds at the leading edge of the AI cloud. Across GPU generations, appetite for AI cloud technology stays strong; take Q3, when we observed our initial contract for more than 10,000 H-100s moving toward exploration. A full two quarters ahead of time, that customer voluntarily signed a new agreement for the infrastructure, at a price that came in within 5% of what the original deal specified; that says a great deal about customer satisfaction, and it also signals the durable usefulness and distinct worth of the GPUs operating on CoreWeave's platform. CoreWeave is the first AI cloud anywhere at hyperscale, bringing together compute, storage, networking and software built specifically for AI workloads. An ever-widening set of software and services supports our expanding cloud portfolio, assisting customers in building, training and deploying new products more quickly. That includes mission control, our proprietary orchestration solution, which matters greatly for running our AI cloud autonomously at the bleeding edge. Not long ago we introduced CoreWeave AI Object Storage, a fully managed storage service that removes all friction from shifting data across regions, clouds and tiers, with no egress or transaction fees whatsoever. For AI workloads, CoreWeave's AI Object Storage provides the greatest throughput available while reducing customer costs by over 75%. Interest in this offering has already been enormous, and we have brought on a number of initial customers, among them frontier AI labs such as Mistral. Throughout our whole storage platform, customer adoption has been swift, surpassing $100 million in ARR during Q3, and paired with our distinctive global network backbone, built specifically for AI, this puts CoreWeave in the position of hub for customers and their essential AI workloads, delivering consistently best-in-class performance and smooth user experiences whether they rely on CoreWeave cloud or a secondary provider. We built on these capabilities by further extending our observability and security suites, making sure CoreWeave is ideally situated to manage every critical workload our customers have, no matter the use case or geography. Over the past few years, our role has been to back the pioneers who develop and refine AI; at present we are broadening that role to assist in putting AI to work. From the tools developers need in order to build AI to the solutions the physical world needs in order to adopt AI, we have treated M&A as a key instrument for speeding up this journey, among them the recently announced acquisitions of OpenPipe, Marimo and Monolith. With OpenPipe, we moved quickly to fold their solutions into our wider fine-tuning product suite and brought out the first publicly available serverless reinforcement learning tool; with Marimo, we are widening CoreWeave's exposure to and impact within the open-source community, beginning with entry-level exploration and prototyping. OpenPipe and Marimo both align seamlessly with what Weights & Biases can do, and there we are fast expanding the developer base that depends on CoreWeave's holistic platform. Through Monolith, we are carrying these capabilities into the physical world to unlock monetization of AI right now, starting with industrial use cases that have an established enterprise customer base and mature workloads, among them leading auto OEMs such as Nissan and Stellantis. Thanks to how quickly and successfully we've rolled out new offerings, our addressable market keeps widening and we keep growing alongside our customers. We're making fundamental changes to what CoreWeave can do, and that's opening beachheads plus chances to expand into fresh markets — all of it aimed at better supporting AI's rapid growth and helping AI builders and innovators reach the market faster and more reliably while driving ROI. The work we take on is becoming more sophisticated, and our partnership with CrowdStrike proves it — that deal will unlock and speed up growth driven by partners. Our new storage product, along with the partnership with Vast Data, is one more instance of speeding up both our product portfolio and our partner go-to-market motions, and it lets us compete in new markets where before we had few or no offerings. This drives customer-led platform adoption and product-led growth, which creates tailwinds for our business. To wrap up, I want to stress the thing that genuinely distinguishes CoreWeave. We are The Essential Cloud for AI, pairing technical and operational excellence that has no equal with a customer base that's diversifying quickly; we provide the most performant infrastructure, the quickest path to market, and the most advanced capabilities out there. The world's top AI innovators pick CoreWeave because we let them move faster, scale smarter, and reach outcomes that just aren't achievable anywhere else. Our momentum has never been stronger, and the opportunities in front of us keep growing. Fueled by outstanding products, an amazing team, and execution no one can match, CoreWeave is prepared to move into its next phase of growth as a full-stack AI service provider and hyperscale. The future runs on CoreWeave, and we're only getting started; with that, here's Nitin.

Thank you, Mike, and good afternoon to all of you. The strong results we posted for the third quarter underscore how intense the demand for CoreWeave remains, along with how sharply we are executing on our plan to build the cloud that AI fundamentally requires. As Mike mentioned, we are still operating in a setting where supply is heavily constrained, and we anticipate that condition lasting well into the future. By staying committed to bringing the highest-performing offering on the market and by putting capital to work throughout the stack, we are fueling expansion and a broader mix of customers, ranging from newly acquired enterprises and AI-native firms to deeper relationships with those we already serve; let me now move to our Q3 numbers. Revenue for the quarter came in at $1.4 billion, a 134% increase compared with the same period a year ago, powered by vigorous customer demand and disciplined execution. Our revenue backlog closed the quarter at $55.6 billion, nearly twice what it was just one quarter earlier. Interest remains strong not only for the Blackwell platform but throughout our entire GPU lineup. During the third quarter, we closed multiple agreements covering earlier GPU generations, bringing in new clients while also renewing contracts on capacity we already had in place. Because demand for CoreWeave's cloud services spans such a wide range, we have been able to bring down our customer concentration substantially. At present, no individual customer accounts for more than roughly 35% of our revenue backlog, a decline from about 50% in the prior quarter and an even steeper drop from approximately 85% when the year began. On top of that, as of Q3, over 60% of our revenue backlog comes from investment-grade customers; this demonstrates exactly what successful pursuit of our stated objective of diversifying both platform and customer base delivers. Our operating expenses for the third quarter came to $1.3 billion. That figure includes $144 million tied to stock-based compensation. We continue to ramp up what we put into data center and server infrastructure, which lets us fulfill our growing revenue backlog. That push raised our cost of revenue, as well as what we spent on technology and infrastructure, during Q3. On top of that, the increase in sales and marketing came from outlays on marketing and on expanding our go-to-market team, so we can seize the rapid growth of AI opportunities among enterprises and AI natives. G&A moved higher because of greater professional services costs and more head count. Adjusted operating income for Q3 hit $217 million, compared with $125 million in Q3 of 2024. Our adjusted operating margin for Q3 was 16%. Our adjusted operating income came in above what we had projected, driven by higher revenue, lower costs tied to when our third-party partners delivered data centers, and improved fleet efficiency. For the third quarter, the net loss stood at $110 million, versus a net loss of $360 million during Q3 of 2024. Interest expense for Q3 reached $311 million, compared with $104 million in Q3 of 2024, which mirrors the additional debt we've taken on to finance our infrastructure build-out, partly offset by the benefit of better interest rates on that debt as we keep making progress toward a lower cost of capital. For the third quarter, adjusted net loss was $41 million. That compares with approximately break-even during the same period in 2024. Adjusted EBITDA in Q3 reached $838 million, up from $379 million in Q3 of 2024 — a year-over-year gain of over 2x. Our adjusted EBITDA margin stood at 61%. Turning now to capital spending: CapEx during the third quarter totaled $1.9 billion, below our expectations, due to the shipment delays Mike mentioned involving a third-party data center supplier. Construction in progress rose substantially, hitting $6.9 billion — an increase of $2.8 billion quarter over quarter — and that increase is a direct result of those delays. Keep in mind that construction in progress covers infrastructure that isn't operational yet, so it stays out of CapEx until it gets deployed. Let's move on to the balance sheet and our solid liquidity standing. On September 30, our cash, cash equivalents, restricted cash and marketable securities amounted to $3 billion. Expanding quickly while running at scale calls for a deliberate strategy when it comes to obtaining capital. CoreWeave has cemented its position as the top AI cloud and the foremost innovator when it comes to funding the infrastructure needed to run the most sophisticated workloads for enterprises and AI labs worldwide. We're still making strong headway on bolstering our capital structure and bringing down our cost of capital; during Q3, we revised the DDTL 2.0 Facility, boosting its remaining drawable capacity by more than $400 million, which formed a fresh $3 billion tranche priced at SOFR plus 4.25% — well beneath what the facility originally cost. As we've said before, we also finalized DDTL 3.0 during the third quarter, priced at SOFR plus 4.00%, marking a 900-basis point drop from the noninvestment-grade piece of our earlier facility. As we look forward, we expect to secure financing at narrower spreads, since our capital partners are increasingly recognizing both our best-in-class performance and the dependable cash flow and predictability provided by our take-or-pay customer agreements. In addition, during July we raised $1.75 billion via senior notes, extending our presence in the high-yield market at a rate 25 basis points lower than our initial offering in May. CoreWeave has secured $14 billion in debt and equity deals so far this year to support fulfilling a rapidly growing backlog and to expand cost-effectively for long-term growth. Apart from payments tied to OEM vendor financing and self-amortizing debt covered by committed contract payments, we face no debt maturities before 2028. Switching to tax. During Q3, we booked a noncash tax benefit, largely driven by the effects of One Big Beautiful Bill. Although the magnitude of the Q3 impact was onetime in nature because of a year-to-date catch-up, we anticipate this change in law will let CoreWeave capture cash tax savings down the road; now moving to guidance. As we said, the powered-shell delivery postponements tied to the data center provider are going to affect our fourth quarter results. These postponements won't last, and as Mike pointed out, the customer impacted has consented to modify the delivery schedule so their capacity is maintained across the entire duration and full value of the initial agreement. Against that background, for 2025 we now forecast revenue between $5.05 billion and $5.15 billion; beyond that, we project adjusted operating income for 2025 of $690 million to $720 million, and we anticipate finishing the year with more than 850 megawatts of active power. During Q4, we'll be commissioning a few of the biggest deployments at scale in our company's history. The mismatch in timing—data center expenses hitting first while revenue recognition lags behind—is going to put pressure on adjusted operating margin in the near term. When we look at 2025, interest expense should land somewhere between $1.21 billion and $1.25 billion. That reflects the extra debt we're taking on to support CapEx growth driven by demand, though it's partly offset by our cost of capital, which continues to decline. Now on CapEx: our updated expectation for 2025 is $12 billion to $14 billion. Compared with what we previously guided, we believe this reduction will mostly be offset by a corresponding increase in construction in progress—essentially infrastructure that's accumulating and waiting to be deployed once powered-shell capacity is delivered. As a result, the bulk of the capital expenditures we'd previously thought would land in Q4 is now going to be booked in Q1 instead. On top of that, given how much our backlog has grown and the fact that appetite for our cloud services remains enormous and unrelenting, we expect CapEx for 2026 to come in at more than twice the 2025 level. Putting money into our infrastructure platform like this will strengthen our competitive moats and support the hyper growth we keep delivering; to sum up, we turned in a record third quarter and feel more confident than we ever have about where this business is headed over the long term. Over the course of this year, we have achieved enormous strides in speeding up the expansion of our revenue backlog, which has now passed $55 billion. At the same time, we have been widening the mix of customers we serve, carrying out strategic partnerships and acquisitions that bolster and expand our platform, tapping into new sources of capital that bring down our cost of capital substantially, and growing our capacity and our organization faster than ever before. Thanks to this momentum, we can capture the opportunities available to us right now and build a solid base for the years ahead. The market we can serve keeps growing — not just because AI is being adopted more and more widely across sectors and applications, but also because of intentional business choices we have made to widen our product offerings and win a larger share of spending throughout the industry. CoreWeave has hit escape velocity, expanding faster and more efficiently while strengthening our position as the go-to cloud for AI. We appreciate the backing and involvement of our investors and analysts. We're eager to share news about how we're advancing in upcoming quarters; at this point, let's begin the Q&A.

Michael, in every conversation we have throughout the AI space, we keep hearing that bookings are surging, and clearly that holds true for CoreWeave as well. However, the constraints involving power and manpower are growing really severe. Could you address that scenario as it pertains to the third-party provider. To be precise, is the issue insufficient power or insufficient manpower? Is it something else involving GPUs, keep in mind your storage? And after that, have you talked with your other third-party providers to understand how their own trends are shaping up against schedule and whether they believe they can stay on track or meet their commitments going into early next year?

Let me sort of break that question down in a few different ways, alright? So to start, you're right. It is extremely frustrating for our clients. It's extremely frustrating for us as well because of the sort of systemic difficulties that exist inside the supply chains that are needed to deliver a — the worldwide infrastructure that artificial intelligence requires. That said, we've implemented a number of measures along the way here to truly reinforce our capacity to handle that environment, which will remain difficult going forward. We've genuinely invested a great deal of effort in diversifying our data center providers. We've built a substantial portion of the company devoted to being able to enable and help with the operational side of delivering infrastructure. We established our own self-build initiatives, among them Kenilworth and Lancaster, Pennsylvania. Lancaster, Pennsylvania. So you can see us sort of truly expanding our — and making sure we're doing all that's feasible to reduce the harm tied to or the hold-ups tied to delivering this infrastructure, which is simply swamping the supply chains. Now once you possess a diversified portfolio of routes to infrastructure, the comparative effect of any single delay grows smaller. You'd simply be able to pull from various data centers as they're being delivered; and so we genuinely view this as, this is a major block of infrastructure that has arrived late. Yet the reality that the final end customer who will be using this infrastructure has moved the contract back to let us deliver the entire contract value despite the delays truly demonstrates the value that customers receive from our infrastructure. So you're going to hear this theme echoed over and over as you speak with not only CoreWeave, but you speak throughout the sector; and it is a genuine difficulty at the powered-shell level; it isn't a difficulty for power, correct? There's abundant power at present, and we think there will be sufficient power over the next couple of years. But truly the place where the difficulty lies is the powered shell.

So Michael, is there any connection at all between this and Core Scientific? Or has this got nothing to do with what you folks went through there?

Look, I won't comment on any particular data center provider. Every one of our data center providers is getting our full effort as we try to help them get the infrastructure finished and handed over to us. Getting infrastructure delivered has gone remarkably well for us, and you keep watching us scale. You saw us reach 500 — roughly 590 megawatts. That's 120 megawatts more than we had on the prior call. So as delivery keeps scaling, the success is clearly substantial; but which individual data center provider it is, I don't believe that truly matters. Down the road, this is a systemic issue that the whole industry will need to work through. What counts here, or what counts from where I sit, is that the infrastructure facing a delay won't hurt our backlog or our capacity to capture all the value from the contracts we intend to fulfill.

Nice work on yet another really striking quarter when it comes to expanding that backlog. You're correct, we have simply — no cloud provider has ever managed to scale up at that pace, as far as we can tell. Mike, there's something I'd like to put to you that comes up with us constantly and that we keep hearing about on CNBC. It's essentially the danger of having too much capacity. But my sense is it's actually a narrower concern than that — what worries people is excess capacity tied to — of the stuff AI labs are signing contracts for. What I'd like to ask you, though, is how are we supposed to view your team's infrastructure and the infrastructure you're putting up? And just how interchangeable is that infrastructure, really? Whenever you're constructing for one specific client, are those data centers, can any client make use of them? Can they serve inference as well as training? Or are you genuinely building to order for a given client, which would box you in and leave you with fewer degrees of freedom, so to speak, should one client perform better or worse?

Yes, Keith, great question; in fact it's a topic we've devoted considerable time to here while we move forward with our relationships across all of our customers. So to put it briefly, this infrastructure is interchangeable. It could be shifted from one client over to another. This infrastructure gets constructed to the toughest requirements; therefore it can serve training. It can serve inference. We've genuinely put a lot of thought into preserving the greatest possible optionality, the greatest possible flexibility in how we build our infrastructure; and I'd like to point out to everybody that much of that flexibility, much of that interchangeability genuinely traces back to the outstanding software suite we offer, which enables such efficient utilization of the infrastructure, correct? For instance, when SemiAnalysis carried out their yearly sort of assessment of the options available, there's a rationale behind CoreWeave repeatedly emerging as unique, as the finest answer for this sort of infrastructure anywhere on earth; and that covers the hyperscalers, the neoclouds and every other party attempting to supply this infrastructure; we simply perform excellently, and we're convinced there's substantial value we safeguard by offering such a strong software suite capable of delivering infrastructure.

That backlog increase is really something. There are 2 points I'd like to raise. First, Mike, you've mentioned before that you plan to bring in more contractors and spread things out, with the most recent one being decided. Could you give us a candid, thorough status report on how close we are to a situation where disruptions completely unrelated to your operations no longer impact your revenue guidance? What's the distance to that milestone? Second, looking at what's happened, well, perhaps a few saw it coming, but a $250 billion deal between OpenAI and Microsoft caught most off guard, and a $300 billion agreement between OpenAI and Oracle did too. Suddenly, the — CoreWeave clearly holds a distinctive advantage in being able to deploy GPU clusters rapidly and efficiently, practically at the pace of thought; yet in a market where hundreds of millions of dollars go to the giant hyperscalers, what makes you distinctive 3 to 4 years down the road once the market has stabilized with supply matching demand? Looking back at CoreWeave then, what compelling value proposition will keep you competitive at that moment?

I'll split that into 2 parts, okay? Your first point concerns diversification and when it begins — when it ceases to create swings in the figures we report each quarter. What I'd ask you to keep in mind is that once individual builds shrink in relation to the total set of data centers we operate, the effect of a delay of a couple of weeks will matter less and less in the overall accounting of things, correct? So while you're bringing 590 megawatts of power online, and you face a step function of 200 or 300, that represents a significant share of what will come online during the following quarter, right? Once we grow bigger and begin constructing all 2.9 gigawatts of power that we possess, a data center of 100 megawatts running a week or 2 behind won't carry a meaningful effect; and as Nitin mentioned, we anticipate that the vast bulk of that 2.9 gigawatts of power will be commissioned within the coming 12 to 24 months. That should give you a clear sense of how the curve starts to flatten out as we scale up and each data center's relative weight declines. So that covers the first part regarding scaling. The second part is that the question you're posing has been put to us from the day we launched this company. What enables CoreWeave to supply GPUs more quickly? What enables us to provide the GPUs that NVIDIA relies on to run its MLPerf? What makes us capable of developing software that will shape this entire category? Quarter after quarter, you can observe how our advantage grows, driven by tailoring our cloud to whatever specific use case is needed. And again, as you noticed in the SemiAnalysis piece, we stand alone in this regard. We are actively constructing our product lineup, and we are either building or acquiring extra capacity so that the compute we provide becomes further decommoditized. That means an organization purpose-built to supply this kind of compute is going to succeed going forward.

Amit, is your line muted?

Operator, could we move on to the following question, and we'll return to Amit afterward. Operator, shall we take the next question now, and then we'll come back to Amit.

I hope my audio is coming through clearly. Let's dig into a few of those pushed timelines you mentioned for the quarter. Could you walk us through what this means looking ahead to 2026? Nitin, I realize you gave some big-picture remarks regarding CapEx, but considering what you can currently see, especially that 24-month piece of RPO, what's the right way for us to frame how revenue gets affected by this movement? Do you expect this pushback to be entirely worked through by Q1? And ought we to anticipate an acceleration in the growth rate in the coming year compared with the current one? Any additional detail you can offer would be appreciated.

Sure. I'll kick things off, and then Nitin will take over. It matters to recognize that the buildout before us stems from infrastructure tied to one particular provider. Meanwhile, we're running things on a parallel track with additional providers for other contracts. As a result, a near-term effect from this delivery is going to show up; after that, you'll witness our capacity to catch up on the schedule as the year progresses. Consequently, the vast bulk of this setback ought to be handled by Q1 of the following year.

Absolutely, Tyler — you've got that right. The bulk of that CapEx shift we saw in Q4 is going to wrap up during Q1. And as you'd expect, we'll be bringing capacity up gradually over Q1 to support that. Like Mike said earlier, the total revenue tied to this customer doesn't take a hit, since we were able to move the delivery dates around for that customer, which means they still get their full capacity along with the contract value attached to it. We'll have more to share on the 2026 build and the related revenue plan at the next earnings call. But as we pointed out this quarter, with the strong customer demand you're seeing — reflected in our revenue backlog growth and the ongoing demand from customers — we're expecting 2026 CapEx to come in at well over twice what we saw in 2025.

I'd like to pull a few of these remarks together, since the bookings increase really jumps out, and plenty of questions center on how that's sequenced. So it sounds as though your point is that the supply chain effects you're encountering are tied more to one particular customer. What I'm hoping to understand better is whether this affects the pace at which you can bring new customers on board at all. Or does it relate more to signing after ramp and one specific customer situation? Then, as a quick add-on, does the NVIDIA deal appear in the backlog figure specifically? It could help to hear you say more about what that arrangement unlocks, since its structure is somewhat different too.