CoreWeave, Inc. (CRWV) Earnings Call Transcript

Thanks. Good afternoon, and thanks for joining us on CoreWeave's call covering the fourth quarter and fiscal year 2025 results. With us today to go over what we delivered are our CEO, Mike Intrator, and our CFO, Nitin Agrawal. Before we dive in, I want to remind everyone that forward-looking statements will be part of what we say today. What actually happens could end up differing materially from what those forward-looking statements anticipate. You can find the factors that might lead to materially different results laid out in the earnings press release we issued today as well as in the annual report on Form 10-K we will file with the SEC. Everything forward-looking we say on this call rests on assumptions that hold as of today, and we take on no duty to revise these statements because of new information or events that come later. On this call, we'll be sharing both GAAP and certain non-GAAP financial measures. Today's earnings press release contains 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. And now I'd like to turn the call over to Mike.

Good afternoon, everyone, and thanks for being here. For CoreWeave, 2025 was the year that defined us; revenue came in above $5.1 billion, a 168% rise compared with the prior year, and our contracted revenue backlog climbed to $66.8 billion — that's $11.2 billion more than the previous quarter and over $50 billion higher than a year earlier. By December 31, we had passed 850 megawatts of active power, and in Q4 we brought on roughly double the number of new reserved instance customers that we had in any other quarter in our history; all of this was achieved while we moved quickly to clear up the data center delays we talked about last quarter, getting the affected deployments live sooner than we had told you to expect on our Q3 earnings call. No cloud company has ever reached $5 billion in annual revenue faster than CoreWeave; we are still in the earliest phase of the biggest infrastructure build-out ever seen, and CoreWeave leads the way, constructing and running some of the largest purpose-built AI clusters anywhere, built for the most demanding workloads in the world. If you cut through the complexity, 4 fundamentals show where we stand: first, demand keeps coming without letup, fueling fast uptake from a customer base that grows more diversified across hyperscalers, AI native companies, and enterprises; second, we see widening chances to monetize CoreWeave Cloud through new margin-accretive paths, opened up as our platform moves past GPUs and as our partnership with NVIDIA recently grows; third, our data center footprint is expanding quickly, supported by execution no one else matches and a strategic plan for adding capacity. And fourth, a disciplined financial model intentionally built to put investment ahead of revenue so we can serve contracted demand, backed by $66.8 billion in revenue backlog and giving us clear sight lines to durable cash flows, appealing returns, and the capacity to lower our cost of capital; we'll go through each of those points today. On the demand side, the signals coming from hyperscalers, AI natives, and enterprise customers keep growing stronger as AI workloads grow more intricate, models get bigger faster, and adoption keeps spreading. This breadth of demand has meant deeper engineering ties with our biggest customers and real strides on diversification. CoreWeave stands behind the next wave of AI pioneers. As I noted in Q4, we brought in roughly double the new reserved instance customers of any earlier quarter in our history, among them AI native and enterprise names such as Cognition, Cursor, Mercado Libre, Midjourney, and Runway; we also deepened ties with some of our biggest partners, including both of the hyperscale cloud customers we already had. Every one of those customer categories showed faster demand, while pricing held steady all through 2025 — a pattern we've watched carry into the start of 2026. All told, across the year, the count of customers committed to spending at least $1 million on CoreWeave Cloud rose by almost 150% — and these aren't one-off infrastructure deployments. They are sophisticated, multiproduct opportunities, the opening chapters of long-lasting platform relationships, and a growth engine that compounds as AI becomes more deeply woven into how these businesses run. We're likewise observing a big uptick in interest for older generations of GPU architectures, and supply for those is tight as well. In Q4, average H100 pricing stayed within 10% of its level at the start of the year, whereas average A100 pricing rose in 2025. Based on what our customers tell us, the need for this infrastructure is mostly tied to inference use cases, and those are growing very quickly. We're committing this infrastructure into fresh reserved instance contracts before it's even available, which decisively eliminates any worries about appetite for older generation SKUs; since nearly all of our new 2026 capacity is already allocated, we keep working hard to broaden our footprint so we can satisfy the enormous requirements of current and potential customers alike, near term and long term. Such patterns strengthen our confidence in how lasting this demand is and how long this technology will endure while operating on CoreWeave Cloud. Given the unquenchable demand landscape and the ongoing signals coming from customers, we're speeding up our roadmap, aiming to bring on more than 5 gigawatts of extra data center capacity past what we've already contracted by 2030. Turning to fresh ways to generate revenue from CoreWeave Cloud. Our platform keeps advancing as we open up margin-accretive paths for growth via new products and services, plus making our proprietary cloud stack available outside our own data centers to the wider NVIDIA ecosystem. AI natives and enterprise customers aren't merely using our core GPU infrastructure. They're interacting with our unified platform at much greater rates spanning CPU, storage, software, and development tools. The chance to deliver extra value to these customers as their AI workloads grow is considerable and amounts to meaningful upside down the road. This is already evident throughout our platform. To illustrate, roughly 80% of CoreWeave Cloud customers paying at least $1 million per year have taken up one or more of our storage products. On top of that, we're experiencing robust cross-selling momentum with Weights & Biases, having brought in hundreds of millions of CoreWeave Cloud TCV from Weights & Biases customers during the second half of the year. We've also sped up work on CoreWeave's proprietary cloud stack, reference architecture, and associated software solutions, such as SUNK and Mission Control, which coordinate every layer of our purpose-built cloud and increasingly shape the CoreWeave customer experience. In January, we announced NVIDIA intends to test and validate our platform, including our software and reference architectures, working toward incorporating those offerings into NVIDIA's reference architecture for cloud, enterprise, and sovereign customers. Already, select customers are licensing SUNK as their default research cluster management platform spanning their multi-cloud footprint; we anticipate wider distribution of our proprietary cloud stack will turn into an expanding source of higher-margin revenue over time. The capacity to monetize our platform, both within and now past our own data centers via third-party licensing agreements, considerably widens our addressable market. That amounts to concrete long-term upside potential that isn't captured in the 2026 guidance as we're presenting it today. Moving to how we carried things out. We closed out the year with over 850 megawatts of power that's live, having brought on roughly 260 megawatts during Q4 by itself, spread across 43 data centers in operation, compared with 32 when the year began. During 2025, we signed contracts for nearly 2 gigawatts of extra power, finishing the year above 3.1 gigawatts of contracted capacity, and we anticipate essentially all of it being energized by the close of 2027. That capacity, contracted yet not yet active, is deferred revenue potential we intend to capture as it gets built and handed over. We'll keep sourcing land, power, and data center shell infrastructure in a strategic way, and especially with an eye past 2026, we're seeing strong opportunities in today's market for CoreWeave to expand our contracted power capacity, and we will also tap our broadened partnership with NVIDIA on a selective basis to speed our roadmap along and serve demand better. Running at this kind of scale and speed is complex by nature. Whenever issues arise, we act swiftly, with disciplined coordination spanning teams and partners. The holdups we talked about on our Q3 earnings call were resolved rapidly. All told, we've now handed over in excess of 50,000 Grace Blackwells to the customer affected, putting servers in place on a rolling schedule and getting them delivered within weeks of gaining access to the needed data center infrastructure.

We're operating at an extremely rapid clip across many locations, delivering tens of thousands of GPUs to a range of customers simultaneously. It's our belief that no other cloud platform besides CoreWeave can run at this velocity and still provide the top-tier performance and dependability that wins customer confidence and allows us to gain more wallet share. The feedback and observations coming from those customers we work with most closely are truly motivating. Grace Blackwell operating at scale on CoreWeave Cloud represents a breakthrough; in Q4, we were the first cloud platform to earn NVIDIA's Exemplar Cloud Status for GB200, and we remained the only AI cloud with a platinum ranking from SemiAnalysis. We think we'll be one of the first companies to launch NVIDIA's Rubin GPU platform commercially in the back half of 2026, and we believe that keeps us leading the pack when it comes to delivering and innovating across the whole AI cloud stack. At the same time, we're expanding what we offer by adding NVIDIA's Vera CPU along with BlueField storage. Once these newer technologies are folded into the cloud platform we built ourselves, they should unlock fresh capabilities for our customers, including agentic workflows. That same pace of execution is also the reason our Q4 capital expenditures came in higher than we had guided. Our teams got infrastructure up and running sooner than anticipated, and we view that as a high-quality way to accelerate revenue capacity for 2026. So you can get a sense of just how big we've become: third-party numbers suggest CoreWeave on its own now exceeds the combined size of the top 15 Neoclouds in North America and Europe. When you bring over 260 megawatts live within one quarter, that requires coordinating servers, network gear, storage, and custom software simultaneously, spanning upwards of 100,000 GPUs along with millions of linked system parts, and doing it in almost flawless sync. Few things in the tech world are this hard to run operationally — and it's an area where CoreWeave outperforms everybody else. Now, shifting to how we make money and our business model: for 2026, we're projecting CapEx of no less than $30 billion, which is over twice what we spent on CapEx in 2025. Let me put that figure into plain perspective — it shows just how much signed-up demand sits ahead of us. We're now at $66.8 billion in revenue backlog, and nearly all of the capital we plan to put to work goes toward backing that contracted demand, which stretches far into the future. Because of that, we can see our long-term margins directly, and they rest on steady cash flow. These investments will pay off in compounding fashion — Nitin will walk you through that shortly, sharing thoughts on our 2027 targets and beyond, along with our 2026 guidance. Funding for this backlog will keep coming mainly from the asset-level delayed draw term loans we brought to this market. Along the way, we anticipate continuing to bring down our weighted average cost of capital, while also opening up wider industry participation in the facilities. Demand to take part in CoreWeave's capital market journey is varied and expanding. For our company, we've constructed an amazing financing engine — one that enables us to expand at the same pace artificial intelligence advances, while keeping us on the path we've laid out toward achieving investment grade. Before I pass things over to Nitin, I'd like to offer a few closing thoughts. Our contracted revenue backlog sits at $66.8 billion, and every agreement covering new capacity ought to begin generating revenue by the end of 2026 at the latest. At this moment, we're providing cloud infrastructure and converting that into revenue. We're constructing and running some of the biggest purpose-built AI clusters anywhere, serving the globe's most demanding workloads, and we do it at a speed and standard nobody else matches. The demand behind this build-out keeps coming, comes from many sources, and keeps expanding, with customers signing on across a product suite that keeps widening. Thanks to that contracted backlog, both we and you can see plainly where things are headed. By broadening our work with NVIDIA, CoreWeave's platform becomes the obvious choice for cloud, enterprise, and sovereign customers hunting for the best AI infrastructure performance within our own data centers and, as of now, past their boundaries too. Nobody matches our capacity to anticipate where AI innovation goes next and construct toward what it will need. We'll keep putting money into and expanding this extraordinary market edge, doing so with discipline and contracted cash flows, deploying capital strategically to add capacity, broaden our product suite, and create the AI cloud our customers want. Our top priority hasn't changed: give customers the highest-performing, most dependable, most efficient AI platform on a worldwide scale. The market keeps speeding up, and CoreWeave is set up to gain from and drive the AI revolution alike. With that, I'll hand things over to Nitin.

Thanks, Mike, and good afternoon, everyone. All through 2025, we carried out the strategy we set out for the year with discipline, starting with our IPO. We meaningfully expanded our customer base, grew our contracted and active power capacity by more than 2x, and bolstered our balance sheet by opening up new sources of funding while bringing down our weighted average cost of capital. We also widened our product portfolio, through both organic and inorganic means, and closed 4 strategic acquisitions successfully to pull forward our roadmap. The speed the market is moving at right now, combined with how much demand exists for CoreWeave Cloud, has created a clear opening. So in 2026, we're putting money to work on purpose to grab that opening, accelerating our roadmap to push our lead even further ahead. Moving on to how we did in Q4. We brought in $1.6 billion in revenue during Q4, which is 110% higher than the same period a year ago, driven by robust customer interest and fantastic execution. For the entire year, revenue landed at roughly $5.1 billion, up 168% compared with the prior year. Interest in CoreWeave Cloud continues to grow stronger, and revenue backlog for the quarter finished at $66.8 billion, which is more than 4x higher this year by itself. During the call, Mike pointed out that we've achieved a lot in broadening who our customers are — spanning hyperscalers, AI-native firms, and enterprises — which was one of the objectives we set when we went public last year. On top of that, clients are now placing their core AI work on CoreWeave over extended timeframes, driving the average weighted duration of our contracts from about four years to about five. Our revenue backlog stands at $66.8 billion, and we expect every contract tied to our new capacity to start producing revenue before the year is out. We're fulfilling what we promised right now; these agreements cover both current and earlier generations of GPUs as part of wider customer plans, and discussions regarding upcoming SKUs are already actively taking place. During Q4, our operating expenses came to $1.7 billion. That figure includes $157 million tied to stock-based compensation. We were able to get our data center and server infrastructure up and running more quickly than we had anticipated, and we brought more capacity online during this quarter than at any other point in our company's history. That led to a matching rise in our cost of revenue along with what we spent on technology and infrastructure. On top of that, the bump in sales and marketing came from money we put into expanding our go-to-market team so we could seize the fast-paced growth of the AI opportunity. As for the higher G&A, that was the result of professional services tied to M&A and financing activities, costs associated with being a public company, and bringing on more staff to keep pace with our growth. For the fourth quarter, adjusted EBITDA reached $898 million, nearly doubling the $486 million recorded in Q4 of 2024. That works out to an adjusted EBITDA margin of 57%. Adjusted operating income in Q4 landed at $88 million, down from $121 million in Q4 of 2024, giving us an adjusted operating margin of 6% for the quarter. The shortfall in adjusted operating income versus what we anticipated stemmed from infrastructure deployment that ran ahead of our expectations. We posted a net loss of $452 million in the fourth quarter, compared with a net loss of $51 million in Q4 of 2024. Interest expense for Q4 totaled $388 million, an increase from $149 million in the same quarter of 2024, reflecting the extra debt we took on to finance our infrastructure build-out. For Q4, adjusted net loss landed at $284 million versus $36 million in Q4 of 2024. Now on to capital expenditures. Q4 CapEx reached $8.2 billion, and for the entire year it was $14.9 billion, which came in above what we had anticipated because our team managed to bring infrastructure into service sooner than we expected. As we flagged last quarter, the substantial rise in construction in progress during Q4, reaching $9.4 billion, which is $2.5 billion more than the prior quarter, mirrors the considerable scale of infrastructure we expect to deliver in the near term. Keep in mind that construction in progress refers to infrastructure that isn't yet in service and isn't yet subject to depreciation. After those assets go live, they'll bring in added revenue, and we'll record corresponding depreciation. Our funding approach is designed to match how we roll out this deployment. A large share of our term debt comes through delayed draw facilities, so we don't pull capital until a data center is up and running. Supply chains worldwide remain challenging because supply and demand are still out of balance, yet operational discipline and strategic sourcing have let us keep navigating those hurdles. Having repeatedly brought infrastructure online at scale, we're confident in our ability to adjust and continue speeding up capacity deployments in 2026 and afterward; now let's turn to our balance sheet and solid liquidity position. As of December 31, our cash, cash equivalents, restricted cash, and marketable securities totaled $4.2 billion. We keep achieving significant headway in bolstering our capital structure and reducing our weighted average cost of capital. In Q4, we brought in approximately $2.6 billion via our first convertible senior notes offering, and because investor interest far exceeded what we were offering, the deal ended up being upsized. Additionally, we expanded our revolving credit facility to $2.5 billion during the quarter, aiming to handle liquidity and back our many growth initiatives. To sum up, across 2025 we brought in more than $18 billion in debt and equity, working alongside over 200 investment partners and financial institutions — a clear sign of just how deep and varied the capital backing CoreWeave's growth really is. As Mike mentioned, back in January we revealed that our commercial ties with NVIDIA were growing, and that came with a $2 billion investment into CoreWeave, backing our platform, our team, and the vision we share for large-scale AI infrastructure. The work we did in Q4, and across the whole year, to strengthen our financial structure and bring down our weighted average cost of capital shows up in the 300 basis points of improvement in our weighted average interest rate over the year, which adds up to an almost 600 basis point cut since 2023. For some context, that 300 basis point gain works out to roughly $700 million in yearly interest savings, using our Q4 debt balance as the basis. Looking ahead, we anticipate that our weighted average cost of capital can keep coming down, since those who supply capital and the rating agencies are more and more recognizing both how well we execute—best in class—and how durable and visible the cash flows are that stand behind our take-or-pay customer contracts. Apart from self-amortizing debt backed by contracts and financing from OEM vendors, we face no maturities on our debt until 2029. Moving on to taxes: in Q4 we booked a noncash tax benefit, mostly because of what One, Big, Beautiful Bill Act did. Similar factors could cause our tax rate to swing widely going forward. Across every category of customer, demand keeps growing stronger and more varied. We are purposefully speeding up investments so we can lock in that contracted demand, and because those commitments last so long, we get obvious visibility into cash flow, letting us deliver best-in-class cloud margins as deployed capacity reaches maturity. For 2026, we anticipate CapEx between $30 billion and $35 billion—over twice what we invested in 2025. Nearly all of that connects to customer contracts already signed that we plan to bring online during this year, since we anticipate active power capacity will double to exceed 1.7 gigawatts by the time the year ends. As I've explained before, once new capacity enters service, data center lease costs begin—power and depreciation expense included—while revenue from customers builds up over the months that follow; during 2026, the size of our deployment program makes this effect even larger.

We're about to add capacity that's approximately twice what we had in 2025, and that brings with it a matching rise in depreciation that will precede the related revenue recognition. Looking at the full year 2026, our expectation is revenue between $12 billion and $13 billion, which at the midpoint works out to roughly 140% growth year-over-year; adjusted operating income should come in at $900 million to $1.1 billion. Margins, we anticipate, will climb sequentially — starting at low single digits in Q1, then widening in both Q2 and Q3, and getting back to low double-digit levels by Q4, as the capacity we've deployed matures and revenue grows against the cost base already in place. Looking ahead to 2026, the margin trajectory we've laid out stems from intentional spending choices aimed at meeting the relentless appetite customers have for what our platform offers. As our operations and expansion settle into a steadier pattern, we remain confident in our ability to eventually deliver margins somewhere in the 25% to 30% range. The contribution margins we see in the mid-20s are generated by our established revenue agreements, and pairing that with the continued growth of the higher-margin offerings and services we keep bringing to market gives us the assurance needed to achieve that goal. For 2026, our guidance does not factor in any significant revenue or margin gains that could come from monetizing CoreWeave's proprietary cloud stack further with other NVIDIA Cloud customers, enterprises, or sovereign clients. We do anticipate that beginning in 2026, and becoming more substantial in subsequent years. That represents real potential upside over the long term. Now, focusing on Q1, we expect revenue to come in between $1.9 billion and $2 billion. Adjusted operating income for Q1 should be somewhere from $0 to $40 million. Q1 represents the lowest point in our margin progression for the year, because during that quarter we anticipate deploying $6 billion to $7 billion in CapEx on infrastructure, as we continue to add considerable capacity on top of the roughly 260 megawatts that came online in Q4. For the first quarter, interest expense ought to land somewhere between $510 million and $590 million. Because so much of our revenue backlog is locked in for the long haul, we can see clearly beyond 2026. Given that we remain on a hyper-growth path, our expectation is to finish 2026 with annualized run rate revenue between $17 billion and $19 billion, and by the close of 2027 we anticipate exceeding $30 billion in annualized run rate revenue; none of this trajectory is being built on speculation. What gives us the confidence to state these figures is contracted customer demand, deep strategic partnerships, active infrastructure deployment, industry-leading capabilities, and a thoughtful approach to capital markets. We closed out a transformative 2025 with a strong fourth quarter and full year. Our contracted revenue backlog increased to $66.8 billion, and at the same time we took meaningful steps to diversify our customer base. We also brought in over $18 billion of debt and equity capital, with costs declining as we went. In addition, we made our platform stronger by adding new products and services as well as through strategic acquisitions. As we began 2026, we had 850 megawatts of active power running across 43 data centers, and we remain on course to go beyond 1.7 gigawatts by the time the year ends, with each contract covering our new capacity anticipated to start producing revenue during this year. Every element of our 2026 investment plan is supported by demand that we have already contracted. And, as we mentioned, the guidance we've provided leaves out possible upside from licensing CoreWeave's proprietary cloud stack — something we anticipate starting to contribute during 2026 and growing in the years after. Thanks. We're ready for your questions.

This is Josh Baer, filling in for Keith. Nice work on a solid quarter. You beat expectations pretty clearly when it comes to CapEx, and it's encouraging that those delivery delays got sorted out faster than we'd anticipated. I'm trying to square that with the active power figure, which looks more in line, plus revenue guidance sitting inside the range, which is far under what you'd typically see in terms of upside. So I'd love for you to break down some of those moving pieces. If things are speeding up, why isn't that reflected in active power and revenue? Perhaps it was.

Josh, thanks for asking. When we bring capacity online, much of it arrived near the close of the quarter, and the monetization from that will begin showing up in 2026. We keep adding capacity quickly. During our prepared remarks we mentioned that we'll keep deploying that capacity across 2026 all year long, Q1 included, and that's the effect you're noticing. As for the Q1 figure, essentially this is our first time giving guidance for 2026 and for Q1 right now.

All right. And it's genuinely wonderful to see that roster of enterprise clients. I was hoping you might break down how a contract and deal of that kind appears when it comes to those enterprise customers. We've got a solid grasp of what a hyperscaler mega contract involves. But could you possibly walk through the size, duration, prepayment, and pricing tied to those enterprise customers?

Well, we don't discuss specific contracts. But what you're observing is a situation where competition is extremely fierce for what we provide, which is the most advanced computing infrastructure, delivered via CoreWeave Cloud, and these contracts mostly appear quite comparable to hyperscale contracts when it comes to tenure. And we collaborate with every one of the enterprise clients individually as we assemble a suitable structure for their business model and for their own customers.

So my question, I think, centers on what it costs us to finance things, particularly when you factor in that CapEx figure of more than $30 billion that we're looking at for this year. What I'd like to know is, as capacity keeps scaling up, could you put some numbers around where you believe your blended cost of capital sits today? What's the trajectory been like across the last 12 months? And on the subject of negotiating with the operators of these data centers, in what way do they evaluate our credit profile? Does that assessment hinge on our customer contracts and the identity of those customers? Or does it come down to other factors? Then, moving to financing, do you believe the NVIDIA credit support — that guarantor framework — will produce a quantifiable reduction in what we pay to borrow during '26?

Yes. I appreciate the question, or should I say the questions. So here's how I'd frame it: this company has achieved remarkable progress as it has matured as a business, as our record of running this infrastructure, collaborating with the client, and delivering infrastructure has grown more extensive. Our cost of capital has fallen 300 basis points across the last 12 months — you've witnessed that. Over the last 2 years, you've witnessed a 600 basis point decline. Our expectation is that this continues. The trend is being propelled by our business performing increasingly well through these ETL structures. Turning to the data centers, we brought on close to 2 gigawatts worth of infrastructure during 2025. To give you a sense of scale and perspective, our portfolio stood at 1.3 gigawatts when 2024 ended. So our capacity to access, build, and drive data center contracts has risen substantially. That's something we're excited about. For us, it's a significant stepping stone as we keep driving the business forward. And again, being able to sign contracts at that scale of data center capacity reflects, once more, the business maturing along with growth in our creditworthiness and scale. When it comes to the data center operators, I can share my own view, right? My view is that these operators have a strong interest in doing business with CoreWeave. Within their data centers, they want a tenant portfolio that's diversified. They want to secure exposure to a company such as CoreWeave, which accounts for so much of the AI infrastructure that will ultimately be delivered. So in us they see a pure-play route to gaining access to artificial intelligence as it scales, and that exposure is something they desire. Regarding our relationship with NVIDIA and how it accelerates our ability to obtain access to data centers, here's the perspective I'd suggest you take: clearly, when the offtake is an investment-grade counterparty, that will influence the cost of capital, or the cost tied to the data center. Clearly, teaming up with NVIDIA — something we do on a selective basis, though definitely not only with them as we expand our data center portfolio — is going to help bring down what we spend on our data center footprint.

Congrats on the bookings — those are really, really impressive. Mike, a few AI models have made what looks like an enormous jump ahead over the past 2 months. Claude Code is the one getting all the attention. Still, my sense is we haven't yet seen models that were completely, deeply trained on these massive Blackwell, GB200, or NVL data centers — the very things CoreWeave pioneered and mastered. What are you picking up in the market about how those Blackwell-based models are progressing? If GBT6 or another one of them shows up within the next 3 to 6 months, do you expect it to feel like a massive advance in what they can do? Or does it look more like a gradual evolution on the Blackwell systems?

Look, the Blackwell systems are incredible, right? They're the next step function in compute power, letting these data scientists — these companies pushing the models forward — build and scale infrastructure in ways they simply couldn't before. My expectation, and certainly every signal from the model companies, is that the pace of performance gains from these models is only starting; it's still early in the Grace Blackwell deployment, right? There aren't that many clusters out there at the size and scale we discussed, which we've already delivered. As those clusters come online in our portfolio, in the global portfolio, I think it makes sense, and you'll see step functions in performance tied to this new technology. We're really excited about it. And I think our customers are extraordinarily excited about it, since they grasp what they'll be able to do with this technology that performs so incredibly well — both for training and for inference once it's available to them.

And Mike, following up on that, since you just mentioned they were inference. How do you weigh the benefits of concentrating on the NVIDIA reference architecture? It's clearly extremely strong for the huge training runs and certain work past that. On the other hand, there'd be any tendency to collaborate with custom ASICs that they do genuinely appear to provide superior inferencing price performance; and then I'm — then clearly, NVIDIA's purchase of Groq perhaps sort of — I don't know whether you believe that, that kind of resets the competitive landscape in a manner that seeing NVIDIA reference architecture might sort of dominate even when it comes to inferencing. I'm just curious how you sort of anticipate that going forward?

Yes. So look, each time we have these calls and each time I'm asked about this, I sort of talk about the manner in which we've approached growing our business, which is we are client-led. Our clients are approaching us and they are informing us that the infrastructure that they require in order to power their business. And I want to be clear that when they say infrastructure, it's not a training infrastructure, it's not inference infrastructure. It's AI infrastructure, right? And they're approaching us particularly because we're capable of delivering such an incredibly performing of the NVIDIA technology. They know we're excellent at it; that's the reason they approach us. Are they seeking other technologies from other providers? That makes sense; but what I believe is that or what I know is that we cannot keep up with the demand signals that are arriving for the product that we deliver; and so we are going to concentrate on continuing to advance the solution that we have that is so performant and that has overwhelmed our ability and the market's ability to deliver infrastructure for the past 3 years.