Within 24 hours, two companies reported quarters that did not exist in this form three years ago. CoreWeave posted revenue of $2.6 billion on Tuesday evening, up 112 percent, alongside a revenue backlog of $104.2 billion. Nebius followed on Wednesday with $582.3 million in revenue, a gain of 454 percent, and raised its target for contracted power to five gigawatts. Both stocks jumped by double digits: CoreWeave around 19 percent, Nebius 34 percent. The Nasdaq Composite closed Wednesday 0.54 percent higher at 26,588, with the S&P 500 at 7,749, within reach of a record.
There is, however, a second number attached to CoreWeave, and it does not come from analysts. It comes from the credit market. In late July, credit default swaps on CoreWeave debt were pricing a roughly 50 percent probability of default over five years. A coin flip. At the same time, the average price target across 21 analysts stood at $145.76, well above the prevailing share price, with a consensus rating of Buy.
A $104 billion order book and a 50 percent chance of default. Both numbers describe the same company, both are documented, and neither is a mistake. The interesting part of this quarter is not which side is right. It is why both can be right at once.
What the two companies actually reported
At CoreWeave, second-quarter revenue rose to $2.6 billion, up 112 percent year over year and 24 percent sequentially. Adjusted EBITDA reached $1.5 billion, a margin of 59 percent. Below that line it gets tighter: adjusted operating income came in at $128 million, a margin of five percent. On a GAAP basis the company lost $626 million, against $290 million a year earlier. The $104.2 billion backlog as of June 30 marks a 246 percent increase from $30.1 billion twelve months prior. More than $25 billion in net new commitments arrived early in the third quarter, taking the total to roughly $129 billion as of August 11. For the full year, management guided to $12.4 billion to $13.2 billion in revenue and capital expenditure of $35 billion to $39 billion.
At Nebius the absolute figures are smaller and the growth rates larger. Revenue of $582.3 million represents a 454 percent increase; the core AI business grew 514 percent to $575 million and accounts for 98 percent of group revenue. Annualized run-rate revenue hit $3.0 billion at the end of June, up 598 percent. Adjusted EBITDA swung from a $21 million loss to a $236 million profit, with the AI segment reaching a 50 percent margin versus 24 percent in the fourth quarter of 2025. Full-year guidance was reaffirmed at $3.0 billion to $3.4 billion in revenue and $20 billion to $25 billion in capital expenditure. Four contracts averaging more than $1 billion each closed during the quarter, including deals with Cohere and Reflection.
Both companies rent computing capacity for artificial intelligence, both are growing in triple digits, and both are investing a multiple of annual revenue into physical assets. CoreWeave spent $9.4 billion in the second quarter alone, and $16.1 billion in the first half against $4.8 billion in the same period last year. Nebius spent roughly $5.7 billion in the quarter. This is where the two paths diverge — not on growth, but on the question of who pays for the buildout.
The backlog is a promise about the 2030s. The interest is due every quarter.
The most revealing line in CoreWeave’s report is not the backlog. It is interest expense, which reached $640 million in the second quarter, up from $267 million a year earlier. Annualized, that is roughly $2.6 billion. Set against it is full-year guidance for adjusted operating income of $960 million to $1.15 billion. By the company’s own optimistic measure, operating earnings amount to about 40 percent of the interest bill.
That explains an income statement that looks contradictory at first glance: a 59 percent adjusted EBITDA margin and a $626 million loss in the same quarter. Between those two lines sit depreciation and interest, and in a business that buys rented graphics processors on credit, neither is a footnote. They are the business. Total debt exceeds $21 billion, the debt-to-equity ratio stands at 7.39, and roughly $7.5 billion of that debt is secured against the GPUs themselves.
One shift inside the backlog went largely unremarked: contracts with terms of 48 months or longer rose from 10 percent to 21 percent of the mix. The backlog is not only growing, it is getting longer. The average dollar in the order book moved further into the future, while the average dollar of interest remains payable this quarter. Of the $104 billion, management expects $12.4 billion to $13.2 billion to convert into revenue in 2026. The rest is a promise about the years after — contractually binding, but future.
Why the equity market and the credit market price the same company differently
This resolves the contradiction from the outset, and the resolution is not a matter of opinion. It is a matter of seniority and duration. An equity holder owns, in substance, an option on whatever remains after every obligation is served. For that investor, a $104 billion backlog against a market capitalization of roughly $48 billion is an extraordinarily attractive ratio — provided the company survives long enough to collect. A bondholder holds the mirror image: they gain nothing if the backlog doubles to $200 billion, and lose everything if a single quarterly coupon goes unpaid. For that investor, the backlog is close to irrelevant unless it converts into cash flow within the next twelve to 24 months.
The credit market is not pricing the business model. It is pricing the time to conversion. A 50 percent five-year default probability does not mean half the analysts are wrong. It means a capital structure with 7.39 times leverage, an interest bill larger than operating profit, and capital expenditure at roughly three times annual revenue leaves very little tolerance for delay. The two markets are not contradicting each other; they are measuring different things. For investors, that is the genuinely usable lesson of this quarter: buying a stock like this is not a bet on growth. It is a bet that there is enough time.
Contracted is not energized — the gigawatt gap
There is a second metric both companies now report more prominently than dollars: power, measured in gigawatts. And the real bottleneck hides in which word each company uses.
As of June 30, CoreWeave operated 51 data centers with 1.5 gigawatts of active power. Contracted power on the same date was 3.7 gigawatts, rising to 4.2 gigawatts by August 11. The full-year target for active power is more than 1.85 gigawatts. At Nebius the spread is wider still: the contracted-power target was raised within six months from above three gigawatts to above four and now to five — while connected power at year-end is guided at 800 megawatts to one gigawatt.
Between contracted and energized sits a factor of roughly two to five at both companies. That is not an accounting error or a stretch. It is the plain fact that a contract can be signed in a day while a gigawatt of delivered power takes years: interconnection queues, substations, transformers, turbines, cooling, permits. Revenue does not begin with a signature. It begins with electrons. The backlog is currently growing considerably faster than the physical ability to serve it — which is the most robust explanation for why these companies need so much capital up front, and why any interconnection delay translates directly into a weaker ability to service debt.
Two companies, two creditors: the bank or the customer
The most instructive difference between the two reports is not margins. It is where the money comes from. CoreWeave funds itself in the capital markets: more than $10 billion in unsecured notes and convertibles, a $3.1 billion term loan, and a $1 billion strategic investment from Jane Street. Nebius funds itself, to a substantial degree, from its own customers. The company expects more than $9 billion in prepayments during 2026, and 70 percent of deals closed in the second quarter included an upfront payment. That is what produced $2.3 billion of operating cash flow on $582 million of quarterly revenue — a figure that is not profit but prepaid rent. Against that sit just $775 million drawn on an asset-backed facility at SOFR plus 250 basis points, plus $2.8 billion raised through new share issuance.
Both approaches finance the same physical asset, with opposite risk distribution. A bank underwrites the borrower and expects to be paid regardless of utilization. A customer who wires 50 to 60 percent of the capital cost in advance is underwriting the asset and carrying part of the utilization risk directly. As a demand signal, a prepayment is therefore considerably harder evidence than a signature on a framework agreement: anyone paying before the first chip runs has already put up proof of conviction. For investors, the share of prepaid contracts is one of the few metrics in this sector that cannot be dressed up.
Compute now has a term structure
Alongside the results, Nebius disclosed pricing with unusual clarity. Mid-term contracts are being signed at $20 million to $25 million per megawatt, while short-term available capacity clears at $40 million to $50 million per megawatt. A capacity auction run by the company settled 15 percent above the highest price previously paid for Blackwell capacity.
Short-dated supply, in other words, costs roughly twice the long-dated equivalent. In commodity markets this pattern is called backwardation, and there it is the classic signature of physical scarcity rather than an expectations bubble. Whoever needs the goods immediately pays a premium because they do not exist; if the shortage were merely anticipated, the premium would sit at the long end of the curve. Computing capacity now behaves like a traded commodity with a spot price and a forward price. That is the strongest argument against reading this sector as pure mania — and simultaneously a precise thing to watch. The moment the short-dated price falls below the long-dated one, the shortage is over, and with it the foundation of these valuations.
Where the money lands beyond the operators
For US investors, the two operators are not the only, or necessarily the most durable, exposure to this buildout. The bottleneck, as described, is power and grid — and that demand does not depend on which cloud provider wins. Nvidia rallied alongside the sector this week, while Super Micro Computer climbed 19 percent and optical component maker Lumentum gained 14 percent on AI-driven results of their own. One step further back sit the electrical equipment suppliers: transformers, switchgear, three-phase uninterruptible power supplies and liquid cooling are the things that physically stand between a signed contract and an energized gigawatt.
The scale is set out by the International Energy Agency: global data center electricity consumption is projected to rise from 485 terawatt-hours in 2025 to 950 terawatt-hours by 2030. It is also worth noting how much exposure already exists passively. CoreWeave, Nvidia and the electrical equipment names sit inside the S&P 500 and the major technology indices, which means many investors hold this trade through a workplace retirement plan without ever having placed it. Anyone considering adding single-stock exposure on top should be clear that they are concentrating a position they may already own.
The case against
The weightiest counterargument concerns an assumption that appears in no press release: the economic useful life of the graphics processors. These balance sheets assume five to six years. If that life is really two to three, because each chip generation devalues its predecessor faster than expected, industry depreciation would be understated by something on the order of $176 billion. At CoreWeave that would strike directly at the $7.5 billion of debt secured against those same chips; analysts consider a covenant breach possible as early as 2027 if collateral values fall faster than modeled.
The second risk is customer concentration. Microsoft accounted for roughly 62 percent of CoreWeave revenue in 2024, and the three largest customers — Microsoft, OpenAI and Meta — are likely to represent more than 80 percent of 2026 revenue. A $104 billion backlog sourced four-fifths from three counterparties is a different asset than a diversified order book, particularly when those same customers are building their own data centers and can move from tenant to competitor. Finally, backwardation in capacity pricing cuts both ways: it documents today’s scarcity but says nothing about how long it lasts once new power arrives at scale.
Four measures that will settle the question
Four things to watch emerge from this quarter, all of them more concrete than any valuation debate. First, the gap between contracted and energized power: if it closes, the order book turns into cash flow; if it widens, what is really growing is the financing problem. Second, the price spread between short-dated and mid-term capacity — the most honest running thermometer of scarcity, because it is built from prices actually paid rather than forecasts. Third, the share of prepaid contracts, which reveals whether customers are co-financing the assets or whether the risk has migrated back to banks and bondholders in full. And fourth, the distance between the share price and the credit spread. Two markets valuing the same company this differently eventually converge; the interesting question is which one moves.
Until then, the finding from this quarter is remarkably undramatic. Demand for computing capacity is real, it has been paid for, and it is scarcer than supply. What remains unsettled is whether the companies serving that demand can survive financially long enough to collect on it. The backlog answers the first question. The interest expense answers the second.
Try TradingView Free for 30 Days
Plus get a $15 discount on your first subscription through this link.
