The Old GPUs Held Their Price: Upgrading CoreWeave to Outperform on the One Disclosure That Mattered

Results vs. Consensus

Year-Over-Year Comparison

Quarter-Over-Quarter Comparison

Quality of Beat

Revenue: High quality and the cleanest line in the report. The $74.7M beat over consensus came despite the powered-shell delay that management flagged, meaning the base business outperformed while a piece of planned capacity failed to arrive. Sequential growth of 12.5% on a $1.2B base is a deceleration from Q2's 23.5%, and the deceleration is entirely explained by the capacity that did not energize. Management noted it also signed a number of deals for older-generation GPUs in the quarter, adding customers and recontracting existing capacity, so the revenue mix broadened rather than concentrating further into the newest silicon.

Margins: Adjusted operating margin held at 16%, and the CFO's explanation is unusually candid about why: adjusted operating income was better than expected partly because of "lower costs due to timing of data center deliveries from our third-party partners." In other words, some of the margin was a benefit of the same delay that cut the revenue guide, because costs that would have been incurred energizing that capacity were not incurred either. That is a temporary flattering effect and it reverses in Q4, which is precisely what the implied Q4 margin of 7.4–8.8% is telling you.

EPS: Both bases beat and both beats are tax-driven. See the callout above. On the operating measure that we care about, adjusted EBITDA less the two costs that unambiguously exist for a leveraged infrastructure business, Q3 was worse than Q2: $838.1M less D&A of $630.5M less net interest of $310.6M equals negative $102.9M, against negative $73.3M in Q2. The business grew revenue 12.5% sequentially and got further from covering its own depreciation and interest, because the debt was drawn for capacity that did not arrive.

Business Drivers

CoreWeave remains a single reportable segment with no revenue disaggregation, so the analytically useful decomposition is again the physical and contractual one. Three of the four dimensions improved sharply this quarter; the fourth got worse.

Capacity: 590 MW Live, 2.9 GW Contracted, and Over 1 GW Unsold

The revenue-per-megawatt figure needs a caveat, because it is our derivation rather than a company disclosure and the denominator is a point-in-time exit number while the numerator is a full-quarter flow. Annualizing Q3 revenue of $1,364.7M gives $5,458.7M; divided by ~590 MW of exit active power that is $9.25M per megawatt, against $10.32M on the same construction in Q2. The decline is not a pricing signal. It is arithmetic: 120 MW arrived during the quarter and earned for only part of it, so the exit denominator is larger than the average denominator that actually produced the revenue.

The more important number is the one management disclosed for the first time this quarter: more than 1 GW of contracted power is not yet sold to any customer, and is expected to energize largely within 12 to 24 months. That is roughly 1.7 times the entire currently-active fleet, already secured, already financed at the power level, and available to sell into a market management describes as insatiable. It is the clearest forward revenue bridge the company has published.

"This leaves us well-positioned for future growth with more than one gigawatt of contracted capacity available to be sold to customers that we expect to largely come online within the next twelve to twenty-four months." — Michael Intrator, Co-Founder, Chairman and CEO

Assessment: The capacity position strengthened materially even as the near-term delivery schedule slipped. Contracted power grew faster than active power, which mechanically holds the deployment ratio flat at around 20% and looks like stasis, but the composition improved: eight new data centers, no single provider above roughly 20% of the portfolio, and the beginnings of self-build. The provider-diversification disclosure is a direct response to the failure mode that caused this quarter's guidance cut, and it is the right response.

Backlog: $55.6B, Nearly Doubled, and Far Less Concentrated

Three contracts drove the $25.5B increase: a multi-year Meta agreement of up to approximately $14.2B with an option to meaningfully expand, an OpenAI expansion of up to approximately $6.5B that brings total OpenAI commitments to approximately $22.4B, and a sixth contract with a leading hyperscaler. Management noted this took the company past $50B of RPO faster than any cloud in history, a claim we cannot independently verify but which is directionally plausible given the comparison set.

The concentration collapse is the part that matters for the equity. Going from roughly 85% of backlog in one customer at the start of the year to roughly 35% in nine months, with more than 60% of the book now tied to investment-grade counterparties, converts CoreWeave's revenue base from a single-counterparty credit bet into something closer to a diversified contract portfolio. That is exactly the outcome we said would move our rating.

Assessment: Strongly positive, with one caveat about the basis. Management disclosed concentration on a revenue backlog basis, which is the more flattering measure: the Meta and OpenAI expansions are contracts that have not yet begun delivering, so they dilute the backlog denominator immediately while leaving revenue concentration untouched for several quarters. The revenue-basis figure was 71% at Q2 and is not disclosed anywhere in this release. Diversification is genuine and directionally large, but it is a forward fact rather than a present one.

Customer Credit and the NVIDIA Backstop

The $6.3B strategic collaboration with NVIDIA announced in the quarter is structurally unlike anything else in the backlog and warrants separate treatment. NVIDIA contracts and reserves capacity; CoreWeave retains the right to interrupt that reservation and resell the capacity to other customers. If the capacity is not resold, it remains committed to NVIDIA and is recognized as revenue.

The accounting consequence is that the amount expected to be resold must be excluded from RPO, so the NVIDIA agreement appears in revenue backlog but largely not in RPO. This is the first time the two metrics have meaningfully diverged, and it means backlog and RPO are no longer interchangeable in a model.

The economic consequence is more interesting. NVIDIA is underwriting the utilization risk on a tranche of capacity that CoreWeave can then sell to smaller, shorter-duration customers who cannot sign multi-year take-or-pay contracts. That is precisely the customer segment CoreWeave has been unable to serve because every megawatt it builds is pre-sold, and it is the mechanism by which the spot and on-demand products finally become possible.

Assessment: A well-designed instrument and a meaningful positive for the diversification thesis. It converts a structural constraint (no capacity available for small customers) into a funded product line, with the utilization downside sitting on NVIDIA's balance sheet rather than CoreWeave's. The obvious caution is that NVIDIA is simultaneously CoreWeave's principal supplier, a shareholder, and now a demand backstop, which makes a portion of the demand signal circular. We size that concern as small at $6.3B against a $55.6B backlog, but it is not zero and it should be tracked.

Capital Structure: $14.0B of Debt Against a Fourth Quarter That Will Not Cover It

Financing execution continued to be excellent on its own terms: $1.75B of 9.0% senior unsecured notes due 2031 priced 25 basis points below the inaugural May offering; DDTL 3.0 closed at SOFR plus 400, a 900 basis point improvement on the non-investment-grade portion of the prior facility; and a DDTL 2.0 amendment adding $0.4B of drawable capacity to create a new $3.0B tranche at SOFR plus 425. Year to date the company secured $14B of debt and equity. Automatic termination of the Series C preferred put right on share-price performance added $1.2B to stockholders' equity at no cash cost.

The problem is timing rather than pricing. Debt rose 27% sequentially to fund capacity that then did not energize, so the interest clock started while the revenue clock did not. That is the entire explanation for the implied fourth quarter, and it is why we treat the guidance cut as more expensive than management's framing suggests.

Assessment: The capital structure is being managed well and is being tested hard. A build-ahead model works when deliveries land on schedule and punishes you asymmetrically when they do not, because the financing cannot be delayed to match. This quarter is the first live demonstration of that asymmetry, and it cost roughly half of the fourth quarter's operating margin.

Key Topics & Management Commentary

Overall Management Tone: Confident on demand and notably more operationally specific than last quarter, with the specificity concentrated exactly where it was missing before: residual pricing, provider concentration, backlog concentration, and the mechanics of the delay. Management led with the bad news rather than burying it, named the cause, quantified the guidance impact, and explained the recovery path in the same breath. Where the posture was thinner was on the fourth-quarter margin consequence, which was described as a timing difference without ever being sized; the reader has to derive the implied 7–9% margin from the guidance arithmetic.

1. The H100 Recontracting: Residual Value, Finally Priced

Last quarter management confirmed that A100 and H100 clusters coming off their initial terms were being recontracted for one to three additional years, and disclosed no pricing whatsoever. We flagged that omission as the single highest-value missing datapoint in the story, because the entire bear case on GPU cloud businesses reduces to a claim about how fast the assets economically depreciate. This quarter the number arrived.

"For example, in Q3, we saw our first 10,000 plus H100 contract approaching expiration. Two quarters in advance, the customer proactively recontracted for the infrastructure at a price within 5% of the original agreement." — Michael Intrator, Co-Founder, Chairman and CEO

Three details do the work. First, the magnitude: within 5% of original price on a cluster of a generation now two architectures behind current production. Second, the timing: recontracted two quarters before expiration, which is a customer behaving as though it expects the asset to be scarce rather than abundant. Third, the scale: a 10,000-plus GPU contract is not a rounding-error renewal, it is a material block of the installed base.

The implication for the model is direct. If accelerators hold roughly their original contract pricing into a second term, then economic life extends well past the initial three-to-four year contract, reported D&A is a conservative rather than aggressive estimate of true economic depreciation, and the collateral behind CoreWeave's $14.0B of debt is durable rather than wasting. Every part of the capital-structure bear case is weaker after this disclosure than before it.

Assessment: This is a single anecdote and we treat it as such. It is one contract, one generation, one customer, and management chose to disclose it precisely because it was favorable. But it is a specific, checkable, falsifiable number where previously there was nothing, and it points hard in one direction. It is the primary reason for our upgrade.

2. The Guidance Cut: One Provider, One Shell, One Quarter

Management cut FY25 revenue to $5.05–5.15B from $5.15–5.35B, adjusted operating income to $690–720M from $800–830M, capex to $12–14B from $20–23B, and the exit-2025 active power target to over 850 MW from over 900 MW. The cause was a single third-party data center provider behind schedule on powered-shell delivery.

"In our case, we are affected by temporary delays related to a third-party data center developer who is behind schedule. This impacts fourth-quarter expectations, which Nitin will discuss shortly. Having said that, the customer affected by the current delays has agreed to adjust the delivery schedule and extend the expiration date. As a result, we maintain the total value of the original contract, and the customer preserves their capacity for the full duration of the initial agreement." — Michael Intrator, Co-Founder, Chairman and CEO

The customer-extension detail is the substantive part and it is easy to skip past. A customer facing a delivery delay had every commercial reason to demand a concession or to re-cut the contract. Instead it extended the expiration date so that CoreWeave retains the full contract value across a longer window. That is a customer that wants the capacity more than it wants leverage over its supplier, which is itself a demand datapoint.

Assessment: We accept management's characterization that this is timing rather than demand, on the evidence of the contract extension and the fact that revenue still beat consensus with the affected capacity absent. What we do not accept is the implicit framing that a timing problem is a cheap problem. It cost roughly $150M of 2025 revenue at the midpoint and roughly $110M of adjusted operating income, and it demonstrated that the build-ahead financing model has no shock absorber.

3. The Fourth Quarter the Guidance Implies

Neither the release nor the prepared remarks state a Q4 number, so it has to be derived. Full-year revenue of $5.05–5.15B less nine-month actual of $3,559.1M implies Q4 revenue of $1,490.9–1,590.9M, or 9.3% to 16.6% sequential growth. Full-year adjusted operating income of $690–720M less nine-month actual of $579.6M implies Q4 adjusted operating income of $110.4–140.4M. Together those give an implied Q4 adjusted operating margin of 7.4% to 8.8%, against 15.9% in Q3.

The interest line makes it starker. Full-year interest expense guidance of $1.21–1.25B less nine-month actual of $841.4M implies Q4 interest expense of $368.6–408.6M. At the midpoints that is $388.6M of interest against $125.4M of adjusted operating income, a ratio of 3.1 times. One year ago that ratio was under one.

Assessment: This is the bear pillar we established at initiation and it is worsening, not improving. The mitigating detail is that the Q4 spike is a direct mechanical consequence of the delay, since the debt drawn against the delayed capacity accrues interest while producing no revenue, and it should unwind through 2026 as that capacity energizes. The unmitigated fact is that FY2025 interest of $1.21–1.25B against FY2025 adjusted operating income of $690–720M means the company will pay its lenders roughly 1.7 times what its operations produced on management's own preferred measure.

4. Concentration Falls Fast, on the Right Metric and a Convenient Basis

The CFO's concentration disclosure was the most quantitatively generous moment of the call and the numbers are genuinely striking: no single customer above approximately 35% of revenue backlog, down from approximately 50% one quarter ago and approximately 85% entering the year, with more than 60% of backlog tied to investment-grade customers.

"Today, no single customer represents more than approximately 35% of our revenue backlog, down from approximately 50% last quarter and even more meaningfully from approximately 85% to begin the year. Additionally, as of Q3, more than 60% of our revenue backlog is tied to investment-grade customers. This is what successful execution against our stated goal of platform customer diversification looks like." — Nitin Agrawal, Chief Financial Officer

The investment-grade statistic is the more valuable of the two, and it is new. A take-or-pay contract is only as good as the counterparty behind it, and one of our objections at initiation was that CoreWeave's revenue base included counterparties funding compute purchases from private markets rather than from operating cash flow. Sixty percent investment grade materially changes the credit character of the book.

The basis is worth naming, though. This is backlog concentration, not revenue concentration. The Meta and OpenAI contracts signed this quarter dilute the backlog denominator on the day they are signed and will not touch revenue concentration for several quarters. The revenue-basis figure was 71% at Q2, and it does not appear anywhere in this release.

Assessment: Real, large, and forward-dated. We upgrade the concentration bear point from materializing to contained, on the strength of the investment-grade disclosure and the trajectory, while noting that the metric management chose to lead with is the one that improves first.

5. Core Scientific: Walking Away on Price

The all-stock acquisition of Core Scientific, announced in July, was terminated in October after Core Scientific shareholders voted it down. Management's framing was that the valuation demanded was not one CoreWeave was willing to pay.

"While the deal made sense strategically for both companies, the valuation required by their shareholders was simply not a price that was appropriate for CoreWeave, Inc. Class A Common Stock, particularly because the outcome of the transaction in no way adversely impacts our ability to achieve our growth ambitions in the coming years." — Michael Intrator, Co-Founder, Chairman and CEO

Two things follow. The first is a discipline signal, and it is a good one: a management team in the middle of an arms race for powered shells, which is exactly the asset Core Scientific owns, declined to raise its offer. Capital-allocation discipline is scarce in a capacity land-grab and is worth something in a business that will be issuing securities continuously for years.

The second is that the commercial relationship survives. CoreWeave continues to lease approximately 590 MW from Core Scientific, so the operational access it wanted is largely intact without the equity outlay or the dilution. The $500M of run-rate synergies claimed at announcement are gone, but we never underwrote them.

Assessment: Better outcome than the deal, on the evidence available. The dilution overhang that the Street cited as one driver of the August sell-off is removed at no strategic cost, and the vertical-integration objective is being pursued instead through self-build at Kenilworth and Lancaster, where CoreWeave controls the economics without paying a public-market premium for someone else's balance sheet.

6. Self-Build as Supply-Chain Insurance

The strategic response to the powered-shell constraint is now explicit and two-sided: keep leasing from third parties, and simultaneously build directly. Kenilworth, New Jersey and Lancaster, Pennsylvania are the flagship projects, with Lancaster carrying an intent to commit up to $6B for an initial 100 MW expandable to 300 MW. Internationally, the UK commitment rose by £1.5B to £2.5B including a new Scottish site developed with the UK government.

"We're not saying that we're going to go self-build and not use third-party data center providers. What we are saying is that self-build is a component of the way that you go about de-risking delivery across the broader portfolio. … It puts us closer to the physical infrastructure. It embeds us deeper into the supply chain around the world so that we have firsthand information." — Michael Intrator, Co-Founder, Chairman and CEO

The information argument is the strongest part of it. A company that builds shells itself knows what a shell actually costs and how long it actually takes, which makes it a far better judge of whether a third-party developer's schedule is credible. Given that a third-party developer's schedule is what caused this quarter's guidance cut, that capability has an immediate and quantifiable value.

Assessment: Correct strategy, and it raises capital intensity further. Self-build converts an operating lease obligation into owned capex plus construction risk, which is more capital-hungry per megawatt and slower to first revenue. It is the right trade for control and information, and it is another reason 2026 capex is guided to more than double.

7. 2026 Capex: More Than Double, With the Plan Deferred

The CFO guided 2026 capex to "well in excess of double" the 2025 figure of $12–14B, which implies something north of $24–28B, and explicitly deferred the 2026 revenue plan to the next earnings call. Investors are therefore asked to absorb the spending number a full quarter before they get the revenue number it supports.

The mitigating context is that a meaningful part of the 2026 figure is 2025 catch-up. The capex guide was cut by roughly 40% and management said the reduction largely reappears as construction in progress, with the vast majority of the pushed-out spending landing in Q1. Cumulative 2025-plus-2026 capital deployment is therefore not obviously higher than what was implied three months ago; it is redistributed later.

Assessment: Sequencing this disclosure ahead of the 2026 revenue plan is an unforced error in a market already anxious about AI capital intensity, and it contributed to the reaction. On substance we are comfortable: a company with $55.6B of backlog and more than 1 GW of unsold contracted power has an identifiable use for the money. But the market was handed a spending commitment without the offsetting revenue framework, and it reacted accordingly.

8. Operating Cash Flow Turns Positive, for Working-Capital Reasons

Operating cash flow was positive $1,689.1M in Q3 against negative $251.3M in Q2, and positive $1,499.1M for the nine months. The swing is worth decomposing before it is celebrated: accounts receivable released $252.2M rather than consuming $865.9M, accounts payable and accrued expenses contributed $622.4M, and deferred revenue contributed $659.3M. Accrued liabilities on the balance sheet rose to $3,172.3M from $1,411.2M at June 30.

So the cash generation is real cash, and it is substantially customer prepayments and supplier payables rather than operating profit. Receivables falling to $1,659.2M from $1,933.7M on a larger revenue base is a genuine improvement in collection: days sales outstanding fell from roughly 145 to roughly 112 on a 92-day quarter.

Assessment: A meaningful improvement that should not be extrapolated. Deferred revenue inflows are a function of when large contracts prepay, and the payables build partly reflects capex accruals that will be settled in cash. The receivable improvement is the durable part and it is the one we credit.

9. Software and Services Reach Escape Velocity in Miniature

The software layer produced its first hard number: the storage platform surpassed $100M of annual recurring revenue in Q3, following the launch of CoreWeave AI Object Storage with zero egress and transaction fees and a claimed cost reduction of more than 75% for customers. Three acquisitions closed or were announced (OpenPipe for reinforcement-learning fine-tuning, Marimo for open-source developer reach, Monolith for industrial and physical-world use cases including work with auto manufacturers), and CoreWeave Federal launched with NASA's Jet Propulsion Laboratory as an early customer.

Against $5.1B of guided 2025 revenue, $100M of storage ARR is under 2% and immaterial to the model. Its significance is directional: it is the first evidence that the Weights & Biases logic (own the developer surface, convert it into infrastructure spend) produces revenue rather than just pipeline.

Assessment: Encouraging and still small. The acquisition cadence, four deals in roughly six months, is fast enough to warrant watching integration quality. We continue to treat the software layer as optionality rather than as a driver, and we would revisit that at a $500M ARR disclosure.

10. Enterprise and Public Sector: The Diversification Is Broadening Beyond the Labs

The customer names disclosed this quarter reach materially further from the frontier-lab core than in Q2: CrowdStrike for cybersecurity agents, Rakuten for visual language models, Jasper in marketing, Poolside and Periodic Labs at the research frontier, NASA JPL in the public sector, plus a leading software design platform and a large US telecom operator left unnamed. The company also hired its first Chief Revenue Officer, from AWS.

The quantitative marker behind the anecdotes is that customers exceeding $100M of trailing-twelve-month revenue tripled year over year. That is the metric to hold management to, because it measures whether the long tail is becoming material rather than merely numerous.

Assessment: The land-and-expand claim now has support: nine of the ten largest customers have executed multiple agreements. The first CRO hire is an admission that the company has been selling through founder relationships and needs a repeatable motion to convert the enterprise pipeline. Sensible, and it takes several quarters to show up.

11. The Powered-Shell Constraint, Restated and Narrowed

Management sharpened last quarter's framing in a useful way. The constraint is not electricity and it is not GPUs; it is the partly-completed buildings into which GPUs can be installed.

"And it is a real challenge at the powered shell level. It's not a challenge for power. There's plenty of power right now. And we believe that there will be ample power for the next couple of years. But really where the challenge is is the powered shell." — Michael Intrator, Co-Founder, Chairman and CEO

This narrows the bottleneck from something structural and multi-decade (grid capacity, generation) to something that is fundamentally a construction-industry throughput problem. Construction throughput responds to price and to time, which means the constraint is real for the next two years and self-correcting after that. That is a shorter scarcity window than the "structural supply constraint" language of last quarter implied, and it is management, rather than a skeptic, drawing the distinction.

Assessment: Honest and slightly double-edged. It strengthens near-term confidence, because a construction bottleneck is knowable and CoreWeave has 2.9 GW already contracted through it. It weakens the very-long-term pricing case, because construction bottlenecks clear. An investor should hold the view that CoreWeave's advantage is a two-to-three year head start on secured capacity, not a permanent scarcity rent.

Guidance & Outlook

Implied Q4: revenue of $1,490.9–1,590.9M (+9.3% to +16.6% sequentially), adjusted operating income of $110.4–140.4M (a 7.4% to 8.8% margin), and interest expense of $368.6–408.6M. The margin compression is the whole story of the guide, and management named the mechanism without quantifying it: the largest-scale deployments in the company's history come online in Q4, and data center costs are incurred before the associated revenue is recognized.

"In Q4, we will be bringing online some of the largest scale deployments in our company's history. This will have a near-term impact on adjusted operating margin due to the timing difference between when data center costs are first incurred and when we start recognizing revenue." — Nitin Agrawal, Chief Financial Officer

Where the capex went: the $8–9B reduction in the 2025 capex guide is not cancelled spending. Management stated the reduction will be "mostly reflected by a corresponding increase in construction in progress," and construction in progress already rose $2.8B sequentially to $6.9B. Because capex on CoreWeave's definition only registers when infrastructure moves into service, equipment sitting in a warehouse or a half-finished shell is invisible to the capex line while being fully visible on the balance sheet and fully funded by drawn debt. That is exactly the configuration that produces a 3.1x interest-to-operating-income quarter.

Guidance style: three quarters of public history now show a consistent pattern. Management guides revenue conservatively and beats it, guides profitability at what it believes and adjusts it when facts change, and volunteers uncomfortable detail (the Q3 interest expense range at Q2; the provider delay and its full P&L consequence at Q3) rather than letting it emerge. The FY guide was cut on the same call that backlog nearly doubled, which is not the behaviour of a team managing a narrative.

Analyst Q&A Highlights

Whether the Bottleneck Is Power, Labour, or Something Else

The opening question pressed for the precise nature of the third-party provider's failure and for read-across to the rest of the portfolio. The answer separated the systemic industry problem from CoreWeave's specific exposure to it, and made the case that portfolio diversification is the structural mitigant.

Q: "Can you speak to that situation relating to the third-party provider? Specifically, is it a shortage of power or manpower? Is it something outside of that with GPUs or memory or storage? And then have you spoken to your other third-party providers to get a sense of their own trending relative to schedule and whether they think they can hold on their deliver on their commitments into early next year?"
— Mark Murphy, JPMorgan

A: "Now, when you have a diversified portfolio of paths to infrastructure, the relative impact of each delay becomes smaller. … And it is a real challenge at the powered shell level. It's not a challenge for power. There's plenty of power right now. And we believe that there will be ample power for the next couple of years. But really where the challenge is is the powered shell."
— Michael Intrator, Co-Founder, Chairman and CEO

Assessment: The most useful reframing on the call. Power abundance with shell scarcity is a construction-throughput problem, which is more tractable and more temporary than a grid problem. It also implicitly narrows the duration of CoreWeave's scarcity advantage, which management did not draw attention to and which an investor should.

Whether Capacity Built for One Customer Can Serve Another

The overcapacity question that dominated the sector debate arrived in its sharpest form: if AI labs have over-contracted, is CoreWeave's purpose-built infrastructure stranded, or can it be redeployed? The answer connected asset fungibility to the software layer rather than to the hardware.

Q: "The question I want to ask you though is how we should think about your guys' infrastructure and the infrastructure that you build and how fungible that infrastructure really is. When you're building out for a particular customer, do those data centers, is that usable for any customer? Is it usable for inference and training? Or do you really build suit a certain customer that would lock you in and give you kind of less degrees of freedom, if you will, if one customer is doing better or worse?"
— Keith Weiss, Morgan Stanley

A: "And so in short, the infrastructure is fungible. It would be able to be transferred from one client to another. The infrastructure is built to the most demanding specs, so it's able to be used for training, able to be used for inference. … And I want to highlight for everyone that a lot of that flexibility, a lot of that fungibility really does tie back to the incredible software suite we provide that allows for such effective use of the infrastructure."
— Michael Intrator, Co-Founder, Chairman and CEO

Assessment: The fungibility claim is now supported rather than merely asserted, because the H100 recontracting shows an actual asset changing hands into a new term at a known price. Building to the most demanding specification is expensive insurance against customer-specific stranding, and it is a defensible reason for the capital intensity that shows up elsewhere in the model.

How Much of the Delay Reverses, and When

A direct question on the revenue mechanics of the shift produced the clearest recovery commitment of the call, from both executives in sequence.

Q: "So double-clicking on some of the delays that you called out in the quarter, can you just help us understand the implications on 2025? … Is this a delay that you think kind of gets fully resolved into Q1? And should we see sort of a step-up in growth rate next year relative to this year?"
— Tyler Radke, Citi

A: "So the overwhelming majority of the delay that you're seeing should be taken care of within Q1 of next year. … The vast majority of the CapEx push out that we experienced in Q4 will be done in Q1. And as you can imagine, we're going to ramp the capacity through the course of Q1 for this."
— Michael Intrator, Co-Founder, Chairman and CEO, and Nitin Agrawal, Chief Financial Officer

Assessment: An unusually specific and testable commitment, and it becomes the primary scorecard item for the Q4 report. Note the asymmetry an investor should hold in mind: the capex lands in Q1 but the capacity ramps "through the course of" Q1, so the interest cost arrives ahead of the revenue a second time. Q1 2026 margins will look like Q4 2025 margins, not like Q3's.

Why the NVIDIA Agreement Sits in Backlog but Not in RPO

The most technically consequential exchange of the call concerned an accounting distinction that materially changes how two headline metrics relate to each other.

Q: "And just as a small follow-up, does the NVIDIA deal specifically show up in the backlog metric? It might be useful to hear you expand on what that deal opens up given it's a bit of a different structure there as well."
— Michael Turrin, Wells Fargo

A: "This contract allows capacity contracted and reserved for NVIDIA to be interrupted and resold to different customers. … Given the flexibility in the contract, interrupt and to resell capacity, accounting rules require us that we exclude the amount we expect to be resold to other customers from RPO. To be clear, if not resold, capacity will remain committed to NVIDIA and will be recognized as revenue. So you see this NVIDIA contract in our revenue backlog, but not in our RPO to a large extent."
— Nitin Agrawal, Chief Financial Officer

Assessment: A modelling trap disclosed clearly and voluntarily, which is to management's credit. Backlog and RPO are no longer substitutes, and any comparison of CoreWeave's backlog growth to a peer's RPO growth is now apples to oranges. The economic substance is better than the accounting: a supplier is underwriting utilization on capacity that can be resold to customers CoreWeave otherwise cannot serve.

Whether a Forty Percent Capex Cut Really Traces to One Provider

The bluntest question of the call tested whether the single-provider explanation was sufficient to account for a reduction of that size, and got an unhedged confirmation.

Q: "And then just wanted to be clear, you cut CapEx by 40% for the year. And just to be clear, this is from one customer, correct? This is you're not assuming other delays across the board. Correct?"
— Brent Thill, Jefferies

A: "That is correct. So this is associated with a single provider, data center provider partner. The delays associated with that. As we talked in our prepared remarks, most of it, a vast majority of it is going to be recognized in Q1. And in Q4, you're going to see a major impact on the buildup of construction in progress associated with the buildup related to it."
— Nitin Agrawal, Chief Financial Officer

Assessment: The answer is categorical, which is what makes it valuable: no hedging about broader supply-chain risk, no allowance built in for a second provider slipping. That is a falsifiable commitment and it is the highest-consequence one on the call. If a second provider slips in Q4 or Q1, the credibility cost will exceed the revenue cost.

How 2026's Capital Programme Gets Funded

With capex guided to more than double, a question on funding sources drew out the deal-by-deal discipline that underlies the financing model, and a stated openness to leasing structures the company has so far used less than peers.

Q: "As we think about CapEx next year, … can you speak as well about the sources of funding a little bit? Because what we've seen for a lot of the other players is that leasing is coming up a lot more. … How do you think about that path for you going forward between the different ways of kind of funding the business, which might give you even more flexibility?"
— Raimo Lenschow, Barclays

A: "We don't sign customers without knowing where the financing is going to come from. We go deal by deal. We make sure that we have the physical data center spoken for. We have the power spoken for. We have the GPU spoken for, and we have the financing spoken for in order to ensure that we are able to successfully deliver compute to them."
— Michael Intrator, Co-Founder, Chairman and CEO

Assessment: The most reassuring sentence in the transcript for a credit-minded equity holder. Four-way pre-commitment (shell, power, silicon, funding) before a contract is signed is the discipline that keeps a build-ahead model from becoming a speculative one. It is also, notably, not what failed this quarter: the financing was arranged and the shell was contracted; the shell simply arrived late.

The Pace at Which the Unsold Gigawatt Gets Contracted

A question about the remaining uncontracted capacity drew the concentration and credit statistics into the answer, positioning capacity allocation as a portfolio-construction exercise rather than a race to fill.

Q: "Mike, with 2.9 gigawatts committed power and over a gigawatt yet to be contracted out to customers, meanwhile, we continue to see a number of other large deals get announced industry-wide. How do you think about and how might you frame for us the pacing on contracting out the remaining capacity given the demand is insatiable out there?"
— Brad Zelnick, Deutsche Bank

A: "Today, you know, approximately, no customer represents greater than 35% of our revenue backlog, which is meaningfully down from where we began the year at 85%. And 60% of our revenue backlog is with investment-grade customers. So vectors that we are very thoughtful around as we care for the capacity that we have available to be sold."
— Nitin Agrawal, Chief Financial Officer

Assessment: Management is explicitly rationing scarce capacity on counterparty quality and concentration rather than on price alone. For an equity whose principal risk is counterparty credit, that is the correct optimization, and it argues that the remaining gigawatt gets contracted more slowly and at better credit quality than a pure revenue-maximizer would achieve.

What They're NOT Saying

1. Revenue-basis customer concentration. Management volunteered backlog concentration in detail and did not mention revenue concentration once. The revenue figure was 71% for Q2. Backlog concentration improves the instant a new contract is signed; revenue concentration takes years.
2. The name of the delayed provider. Asked directly whether it related to Core Scientific, management declined to identify any specific provider. Given the terminated acquisition and the roughly 590 MW leased from that counterparty, the refusal is conspicuous.
3. The Q4 margin, in numbers. The implied 7.4–8.8% adjusted operating margin has to be derived from the full-year guide less nine-month actuals. Management described a "near-term impact" without ever sizing it, on a call where every other guidance element was given as an explicit range.
4. Any 2026 revenue framework. A capex number more than double the prior year was disclosed while the revenue plan it funds was explicitly deferred to the next call.
5. Utilization. Still never disclosed, in any form, in three quarters as a public company.
6. Pricing on the more than 1 GW of unsold contracted capacity. The forward revenue bridge depends entirely on the rate at which that capacity is contracted, and no pricing framework, floor, or comparison to the existing book was offered.
7. Depreciation life assumptions. D&A is now $630.5M, or 46% of revenue. The useful-life estimate applied to accelerators remains the most consequential accounting judgment in the statements and drew no commentary and no question, even on the same call where the H100 recontracting datapoint was volunteered.
8. The economics of the NVIDIA agreement. Management said NVIDIA "fully underwrite the economics," which was not elaborated. Whether resold capacity clears above or below the NVIDIA reservation rate determines whether the structure is accretive or merely protective.

Market Reaction

• Pre-print setup: close of $105.61 on November 10, down 23.7% over the trailing 30 days from $138.43 on October 10, and 42% below the 52-week closing high of $183.58. The stock entered the print already de-rating alongside a broader repricing of AI infrastructure risk.
• After-hours move: shares fell approximately 6% in extended trading on November 10, during and after the call.
• Next-day session: opened at $96.10, a 9.0% gap down; traded a $88.00 to $98.00 range; closed at $88.39, down 16.3% and $17.22 on the day.
• Volume: 76.9M shares against a 30-day average of 25.3M, a 3.0 times multiple, and roughly double the 37.2M that traded on the August reaction session.
• Benchmark: the S&P 500 rose 0.2% on the session and was up 16.2% year to date entering the print, so the entire move was company-specific.

The reaction deepened materially between the after-hours print and the next day's close, from roughly 6% to 16.3%. That widening is informative. The initial after-hours move priced the headline: revenue beat, guidance cut. The next-day move priced the arithmetic underneath it, which took a night to work through. Three things emerged on reflection.

The implied fourth quarter is far worse than the full-year cut suggests. A $150M reduction to a $5.25B full-year revenue midpoint is a 2.9% trim. The implied Q4 adjusted operating margin of 7.4–8.8% against Q3's 15.9% is a near-halving, and it only becomes visible when the reader subtracts nine-month actuals from the full-year guide. Management did not present that subtraction.

The capex guide moved in the wrong direction twice. A 40% cut to 2025 capex reads as a demand signal to anyone who has not followed the construction-in-progress explanation, and the simultaneous 2026 guide of more than double, without a revenue plan attached, gave the bear camp both a near-term slowdown headline and a long-term cash-burn headline from the same paragraph.

The setup offered no cushion. The stock had already fallen 23.7% in thirty days into a market growing visibly anxious about AI capital intensity. A guidance cut delivered into that tape gets a harsher hearing than the same cut would have received in August, regardless of the backlog number printed alongside it.

Our read is that the market priced the delay correctly and the disclosures around it barely at all. Backlog nearly doubling, concentration falling to roughly 35% from roughly 85%, 60% of the book investment grade, and an H100 cluster recontracting at within 5% of its original price are, in aggregate, worth considerably more than one quarter of deferred revenue recognition. The session sold the quarter that was cut and ignored the questions that were answered.

Street Perspective

Debate: Is the Guidance Cut a Timing Event or the First Crack?

Bull view: The bull case on the Street is that this is the most benign possible version of a guidance cut. One provider, one shell, one quarter, with the affected customer voluntarily extending the contract to preserve full value, and the capital spending reappearing as construction in progress rather than evaporating. Revenue still beat consensus with the affected capacity absent, and backlog nearly doubled on the same call.

Bear view: The bear camp contends that build-ahead businesses do not miss once. Powered-shell scarcity is industry-wide by management's own description, CoreWeave depends on third-party developers for the large majority of its 2.9 GW, and management has now committed categorically that no other provider is slipping. That is a promise made about counterparties it does not control, in a supply chain it says is overwhelmed.

Our take: Bull, with the caveat priced. The contract extension is the decisive evidence: a customer with the leverage to renegotiate chose instead to lengthen the term. But we agree the categorical framing is a hostage to fortune, and we would treat a second provider delay disclosed at Q4 as a thesis event rather than another timing item, because it would convert a specific failure into a systemic one.

Debate: Do GPUs Hold Their Value, or Was That One Lucky Contract?

Bull view: A growing view holds that the H100 recontracting at within 5% of the original price, agreed two quarters early, settles the question. Scarcity extends to prior-generation silicon because the constraint is deployed capacity rather than chip performance, and every quarter of shell shortage extends the economic life of the installed base. On that reading, reported depreciation overstates true economic decay and understated earnings power is hiding inside the D&A line.

Bear view: The skeptical response is that one favorable contract, disclosed selectively by a management team with every incentive to disclose it, is not a pricing curve. The renewal happened during the tightest capacity conditions in the industry's history, and says nothing about clearing prices once the more than 1 GW of unsold contracted capacity, and the far larger volumes being built industry-wide, come online in 12 to 24 months.

Our take: Bull on the direction, agnostic on the magnitude. One datapoint cannot establish a residual-value curve, and we would not underwrite a specific terminal value from it. But the prior was that prior-generation GPUs would re-price down sharply, and the observed evidence contradicts that prior at the only observation available. The asymmetry of updating matters more than the sample size: this is the first test of the central bear claim, and it failed.

Debate: Does More Than Doubling Capex Into an AI Capacity Debate Create or Destroy Value?

Bull view: Some sell-side desks argue the 2026 capex guide is the strongest possible statement of confidence and is fully underwritten by a $55.6B backlog with more than 60% investment-grade counterparties. Deal-by-deal pre-commitment of shell, power, silicon and financing means the spending follows contracts rather than anticipating them, which distinguishes CoreWeave from peers building speculatively.

Bear view: The bear framing is that capital deployed at more than double the prior year's rate, into a market where the marginal buyer's own economics are unproven, is how capacity cycles end. Guiding the spending a full quarter before the revenue plan it supports, in a tape already anxious about AI capital intensity, invited the multiple compression that followed.

Our take: We side with the bulls on substance and with the bears on the disclosure sequencing. A company with more than 1 GW of contracted, unsold capacity has an evidenced use for the capital. But the decision to publish a spending number without the revenue framework beside it cost the stock more than the guidance cut did, and it was avoidable.

Model Update & Valuation Framework

The FY2026 revenue bridge. Exit-2025 active power of over 850 MW, plus the delayed capacity landing through Q1, plus the more than 1 GW of contracted-but-unsold capacity energizing largely within 12 to 24 months, points to CoreWeave exiting 2026 somewhere in the 1.6 to 2.0 GW range and averaging roughly 1.2 to 1.4 GW of active power across the year. At $8.5–9.5M of annualized revenue per average active megawatt (below the $9.25M Q3 exit-basis figure, to allow for the drag of capacity that earns for only part of a period), that produces $10.2–13.3B. We hold our estimate at $10.0–12.0B, below the midpoint of that mechanical range, because three quarters of public history show that energization schedules slip more often than they accelerate.

Valuation framework. Unchanged in method: enterprise value against revenue, with an explicit net debt assumption stated rather than buried, because earnings-based multiples are unusable while GAAP income is negative and adjusted EBITDA excludes both the asset and the money that bought it.

At the November 11 close of $88.39 and 497.9M shares, equity value is approximately $44.0B. Adding total debt of $14.03B and subtracting cash, restricted cash and marketable securities of $3.02B gives an enterprise value of roughly $55.0B, which is 10.8 times the midpoint of the revised FY2025 revenue guide and 5.0 times the midpoint of our FY2026 estimate. Three months ago the same construction gave 6.4 times FY2026. The multiple compressed 22% while our FY2026 revenue estimate went up.

Twelve-month price target: $107 (range $96–119). The target is lower than August's $115 because our FY2026 net debt assumption widened on the doubled capex guide, and because near-term estimates came down. The rating is higher because the price fell 25% while our target fell 7%, and because two of the four things we said would change our mind actually happened.

What differentiates the scenarios. The bear case is a world where the more than 1 GW of unsold capacity contracts slowly or at compressed rates, a second provider slips, and the market applies an infrastructure multiple to an equity sitting behind $32B of net debt. The bull case does not need faster growth than the backlog already implies; it needs the H100 recontracting datapoint to generalize, because a durable residual value converts the depreciation line from a real economic cost into a conservative accounting estimate and re-rates the whole structure. As at initiation, both tails turn on the same variable. Unlike at initiation, we now have one observation on it, and it favours the bull.

Thesis Scorecard Post-Earnings

Scored against the eight pillars established at initiation in August, and against the seven commitments we said we would grade this quarter.

Grading Last Quarter's Commitments

Overall: The thesis strengthened where it mattered most and weakened where we already expected it to. Every full-year commitment was cut, the quarter itself beat on all four guided line items, and both structural questions we raised at initiation were answered favourably. Those are not the same kind of information: guidance is about one quarter, and residual value, counterparty credit and concentration are about terminal value. A quarter that trades a delayed shell for a priced GPU renewal and a 50-point reduction in backlog concentration is a good trade, and the market charged 16.3% for it.

Action: Upgrade to Outperform from Hold, with a $107 twelve-month target. Position sizing should still reflect the width of the distribution: the bear case remains worth roughly a third of the current price.

What would take us back to Hold: a second data center provider disclosed as delayed, which would convert a specific failure into a systemic one; the Q1 2026 catch-up not materializing as management categorically committed; the more than 1 GW of unsold contracted capacity contracting at visibly compressed rates; or a further residual-value datapoint that contradicts the H100 renewal.

What would take us to Underperform: a counterparty credit event among the top three customers; evidence that the interest-to-operating-income ratio is structurally rising rather than cyclically spiking on the delay; a debt raise at materially wider spreads, which would signal the credit market has repriced the collateral; or a 2026 revenue plan at the February call that fails to justify a capital programme of more than double the prior year.