CoreWeave Capex and Debt 2026: Guidance, Bonds and Customers
· Updated
CoreWeave now expects $37 billion to $39 billion of 2026 capex, about $3 for every $1 of its $13.0 billion to $13.3 billion revenue guidance. It spent $14.5 billion in Q3 alone, and says it will fund 2027 without new unsecured bonds. How it pays.
CoreWeave guides to $13.0 billion to $13.3 billion of revenue in 2026. Circles are sized by area.
Its capex guidance for the same year is $37 billion to $39 billion.
Shrink both to coins. The revenue becomes one dollar of sales...
...and the capex splits into about $3 of building for it. The highest ratio among the large AI builders we track.
CoreWeave rents out NVIDIA GPUs in data centers it leases and equips. It's often called a "neocloud," a cloud company built for AI alone. Customers sign multi-year contracts for capacity, and CoreWeave borrows against those contracts to buy the chips. When it works, revenue grows very fast: Q3 revenue was $3.66 billion, up 168% on a year earlier. The cost is debt that more than tripled in a year and interest that takes about 23 cents of every revenue dollar.
Customers sign multi-year contracts for capacity.
CoreWeave borrows against those contracts to buy the chips.
A year ago that debt was $14.0 billion. By September 30, 2026, it was $46.9 billion.
Now picture one dollar of CoreWeave revenue as 100 dots.
Net interest expense was $846 million in Q3: 23% of revenue, flowing back to lenders.
The short version
- 2026 capex guidance: $37 billion to $39 billion, raised three times this year from $30 billion to $35 billion in February.
- Spent so far: $30.7 billion in the first nine months by CoreWeave's own measure, including $14.5 billion in Q3. That leaves $6.3 billion to $8.3 billion for Q4.
- Funding: after October's bond sell-off, no new unsecured bonds through 2027 at today's yields. Loans secured on customer contracts and about $9 billion of customer prepayments due in the next three months do most of the work.
- 2027: if it builds only for contracts already signed, CoreWeave expects 2027 capex "in line with" 2026.
- Backlog: about $132.5 billion of contracted revenue at September 30, up from about $104 billion in June and roughly ten times 2026 revenue guidance.
- Customers: three customers made up 72% of Q2 revenue, down from one customer making up 71% a year earlier. Meta, Microsoft, OpenAI and Jane Street are named in its filings. The Q3 split comes with the 10-Q.
A year earlier, one customer made up 71% of CoreWeave's revenue. The pale disc is all revenue.
By Q2 2026 that block had broken up: three customers made up 72%.
Their shares: 36%, 26% and 10%. CoreWeave doesn't say which customer is which.
Meta, Microsoft, OpenAI and Jane Street are all named in its filings.
- Debt: $46.9 billion on the balance sheet at September 30, 2026, up from $14.0 billion a year earlier. Net interest expense was $846 million in Q3, 23% of revenue, and is guided to $1.05 billion to $1.15 billion in Q4.
- The October sell-off: the stock fell 36% from September 30 to $55.60 on October 22 as its bond yields passed 15%, then kept sliding to a 2026 closing low of $43.66 on November 5. Q3 was CoreWeave's answer: no new unsecured bonds at these yields, and less than $5 billion of new money raised at the company level through 2027. What happened.
Company figures are from CoreWeave's filings and earnings calls through its Q3 2026 report on November 10, 2026; the Q3 10-Q, with customer shares and debt maturities, wasn't yet filed, so those still show June 30. Market figures run to the November 11, 2026 close. We update after each quarterly report.
CoreWeave capex guidance: every change
| Date | What CoreWeave said | For | Source |
|---|---|---|---|
| Early 2025 | $20 billion to $23 billion | 2025 | As reported when it was cut, |
| November 10, 2025 | Cut to $12 billion to $14 billion after a third-party data center developer fell behind schedule | 2025 | |
| February 26, 2026 | 2025 actual: $14.9 billion. 2026 guidance: $30 billion to $35 billion | 2025 and 2026 | |
| May 7, 2026 | $31 billion to $35 billion; low end raised on component pricing | 2026 | |
| August 11, 2026 | $35 billion to $39 billion | 2026 | |
| November 10, 2026 | $37 billion to $39 billion; low end raised because it expects to deliver capacity to customers sooner than planned. 2027 "in line with" 2026 if it builds only for signed contracts | 2026 and 2027 |
The pattern matters. In 2025, CoreWeave's spending was limited by how fast others could build the buildings it rents, and it ended the year $5 billion to $8 billion below its original plan. In 2026, the guidance has gone only up, by $4 billion to $7 billion so far. The latest range sits just under the $37.69 billion analysts expected (Visible Alpha, via CNBC).
- Guided to $20B to $23B early in the year
- Cut to $12B to $14B in November after a developer fell behind
- Landed at $14.9B
$5B to $8B below the original plan
- Started at $30B to $35B in February
- Low end raised to $31B in May on component pricing
- Raised to $35B to $39B in August
- Low end raised to $37B in November
Up $4B to $7B so far
Two capex numbers: CoreWeave's vs the cash-flow statement
CoreWeave quotes "CapEx" on its calls, which counts equipment when it's received. The cash-flow statement counts cash paid for property and equipment, which can come months earlier or later.
| Period | CoreWeave's "CapEx" | Cash paid for property and equipment |
|---|---|---|
| 2025 | $14.9bn | $10.3bn |
| Q1 2026 | $6.8bn | $7.7bn |
| Q2 2026 | $9.4bn | $6.4bn |
| Q3 2026 | $14.5bn | $9.4bn |
| First nine months of 2026 | $30.7bn | $23.6bn |
The nine-month cash figure is from the cash-flow statement; the quarters add to $23.5bn because of rounding.
Cash figures are from the 2025 Form 10-K, the
Q2 2026 Form 10-Q and the
Q3 2026 release. Q3's $14.5 billion is from the CFO's remarks on the
Q3 call. The guidance uses CoreWeave's own measure, so we use it in the worked example below.
Worked example: what Q4 is left with
| Step | Low end | High end |
|---|---|---|
| 1. 2026 capex guidance | $37bn | $39bn |
| 2. Spent in the first nine months ($6.8bn + $9.4bn + $14.5bn) | $30.7bn | $30.7bn |
| 3. Implied Q4 (1 − 2) | $6.3bn | $8.3bn |
Pick a full-year total
Last quarter we wrote that a big Q3 followed by another raise would mean the 2026 range was still too low. That's what happened: Q3 came in $1 billion above the top of CoreWeave's $11.5 billion to $13.5 billion guidance, and the low end of the year went up by $2 billion. But the top end didn't move. Taken literally, the guidance says Q4 capex drops to $6.3 billion to $8.3 billion, 43% to 57% below Q3. CoreWeave says it pulled deliveries forward, so part of what would have been Q4 spending landed in Q3. A Q4 well above $8.3 billion would mean the range was still too low.
Capex per megawatt
Data centers are measured in megawatts (MW) of power, and "capex per MW" is a common shortcut for what AI capacity costs. CoreWeave had more than 850 MW of active power at the end of 2025, passed 1.7 GW at September 30, and still expects more than 1.85 GW by the end of 2026, about 1 GW added. Only about 150 MW of that is left for Q4, while Q4 capex is still guided at $6.3 billion to $8.3 billion, so a growing share of the spending is for capacity that goes live in 2027.
Dividing $37 billion to $39 billion of 2026 capex by about 1,000 MW added gives roughly $37 million to $39 million per MW. Treat that as a rough ceiling, not a precise cost. Some 2026 capex is for capacity that goes live in 2027, and the power figures are "more than" targets. It does show where the money goes. A shell and power hookup cost a fraction of that per MW. Most of it is GPUs, servers and networking, which wear out in years, not decades.
Power status, from CoreWeave's releases:
| Date | Active power | Contracted power |
|---|---|---|
| December 31, 2025 | More than 850 MW | About 3.1 GW |
| June 30, 2026 | 1.5 GW | About 3.7 GW |
| September 30, 2026 | More than 1.7 GW | More than 4.6 GW, of which more than 1.3 GW outside the US |
| End of 2026, guidance | More than 1.85 GW |
The backlog, and who it's from
| Date | Revenue backlog |
|---|---|
| December 31, 2025 | $66.8bn |
| March 31, 2026 | $99.4bn |
| June 30, 2026 | About $104bn |
| September 30, 2026 | About $132.5bn |
Sources: CoreWeave's Q4 2025,
Q1 2026,
Q2 2026 and
Q3 2026 releases. The
hyperscaler backlog tracker puts CoreWeave's RPO next to Oracle's, Microsoft's, Alphabet's and Amazon's.
Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was 10×.
Roughly ten times: about $132.5 billion at September 30, up from about $104 billion in June.
Backlog is revenue customers have committed to but CoreWeave hasn't earned yet. At about $132.5 billion, it's roughly ten times CoreWeave's 2026 revenue guidance. On the Q3 call, the CFO said more than half of it is on contracts where delivery to the customer should start by the end of this year. That matters because backlog only turns into revenue once capacity goes live. At June 30, the 10-Q put remaining performance obligations at $103.7 billion, with 41% expected within 24 months.
The customer list has widened, but it's still short. From the Q2 10-Q:
- Three customers made up 36%, 26% and 10% of Q2 2026 revenue. A year earlier, one customer made up 71%. CoreWeave doesn't say which customer is which.
- Named significant customers include Microsoft and OpenAI. In September 2025, OpenAI committed to pay up to about $6.5 billion through May 2031 under one order. In March 2026, Meta committed up to about $21.0 billion, and in April 2026 Jane Street committed about $6.0 billion.
That's real progress: a year ago, one customer leaving would have taken most of CoreWeave's revenue with it. It's still a business that depends on a handful of AI labs and cloud giants keeping their plans.
How CoreWeave pays for it
| September 30, 2025 | June 30, 2026 | September 30, 2026 | |
|---|---|---|---|
| Debt on the balance sheet | $14.0bn | $35.1bn | $46.9bn |
| Cash and cash equivalents | $1.9bn | $5.5bn | $4.1bn |
| Debt minus cash | $12.1bn | $29.6bn | $42.8bn |
| Property and equipment, net | $20.7bn | $46.7bn | $56.8bn |
| Quarterly revenue | $1.37bn | $2.58bn | $3.66bn |
| Quarterly net interest expense | $0.31bn | $0.64bn | $0.85bn |
| Interest as a share of revenue | 23% | 25% | 23% |
| Quarterly operating income (loss) | $52m | −$49m | $207m |
| Quarterly net loss | −$110m | −$626m | −$709m |
Source: CoreWeave's Q3 2026 release, its Q2 2026 release and Forms 10-Q, via SEC XBRL. Debt is the balance-sheet value of recourse and non-recourse debt, excluding leases; cash excludes restricted cash.
In the first nine months of 2026, CoreWeave's operations brought in $5.6 billion of cash, and it spent $23.6 billion on property and equipment. New debt covered most of the gap: $27.0 billion borrowed and $7.6 billion repaid. It also sold $3.0 billion of stock privately, $2 billion to Nvidia in January and $1 billion to Jane Street. In Q3 alone, free cash flow (operating cash minus equipment) was about −$7.5 billion, less negative than the −$10.1 billion analysts had expected (FactSet, via CNBC). During the quarter it sold
$4.2 billion of 2.875% convertible notes due 2033, closed a $2.6 billion loan facility for shorter customer contracts, and set up its first at-the-market program to sell up to 35 million new shares.
The funding plan for 2027
On the Q3 call, CFO Nitin Agrawal answered October's question directly (transcript,
CNBC):
- No new unsecured bonds: "At current yields, we do not intend to issue additional parent-level unsecured debt through 2027."
- Less than $5 billion of company-level money: it expects "less than 5 billion in incremental parent level capital through 2027." As we read it, that includes any new stock or convertibles.
- Secured loans and prepayments do the rest: "asset level debt" (loans secured on specific customer contracts and the GPUs that serve them) should "cover the significant majority" of what it still needs, alongside about $9 billion of customer prepayments and deposits expected in the next three months.
That's a big change from the $30 billion Baird's Rob Oliver thought CoreWeave would need to raise in 2027 (via CNBC). It's also a cheaper mix: CEO Mike Intrator put CoreWeave's weighted average cost of debt at 7.9% at the end of Q3, about half the 15.78% its unsecured 2031 notes yielded in October. The catch is that asset-level loans are tied to specific contracts, so they only work for capacity a customer has already signed for. That's why the 2027 capex outlook is framed as "committed backlog only."
Nebius, its closest listed rival, leans on cheaper money too: convertible notes paying about 2% and customer prepayments. See Nebius capex and debt.
Tap the line you think is the tell.
Spotted it.Not that line. Borrowing did the work. CoreWeave's debt went from $14.0 billion to $46.9 billion in a year, and net interest takes about 23% of revenue.
On CoreWeave's preferred measure, the business looks healthy: adjusted EBITDA (earnings before interest, taxes, depreciation and some other costs) was $2.26 billion in Q3, a 62% margin. The gap between that and a $709 million net loss is mostly $1.89 billion of depreciation on all those GPUs and $846 million of interest on the debt that paid for them. Q3 interest came in just under the $860 million to $940 million CoreWeave had guided. For Q4 it guides $1.05 billion to $1.15 billion. Full-year revenue guidance implies Q4 revenue of $4.69 billion to $4.99 billion, so interest would take about 21% to 25% of it.
Why CoreWeave stock and bonds fell in October 2026
CoreWeave needs lenders to keep refinancing it at rates it can afford, and in October they balked. With interest rates rising and doubts growing about who pays for the AI build-out, CNBC reports that many investors see debt-funded GPU renters like CoreWeave as the most vulnerable part of the AI trade. The stock closed at $55.60 on October 22, its lowest close of the year to that point. That's 36% below the September 30 close, 38% below where it closed on the day of its Q2 report and 70% below its June 2025 closing high. Nothing in CoreWeave's reported numbers changed. The price of new money did. The pressure didn't end in October: on November 4, the day after the midterm elections, CoreWeave fell 9.9% as data center stocks sold off over a Texas vote to extend a freeze on new data center approvals (
StockTitan;
our midterms analysis). It closed at $43.66 on November 5, its lowest close of 2026 and 50% below the September 30 close. After its Q3 report on November 10, the stock rose in extended trading (
CNBC) and closed 24.7% higher at $60.77 on November 11 (Yahoo Finance closing prices).
| Date | What happened | Source |
|---|---|---|
| October 14 | Shares fell 15.5%. The yield on the 9.625% notes due 2032 had briefly touched 15% the day before and stood at 14.8%; the 9.75% notes due 2031 yielded about 14.9%. S&P rates CoreWeave B+, which is junk. | |
| October 15 | The 2031 notes yielded 15.78%. Five-year credit default swaps went above 1,140 basis points before settling near 1,038, up 16% in a week. Shares closed 5.9% higher. | |
| October 16 | Shares fell 11.9% to $61.39, ending the week down 25.2%. | Yahoo Finance closing prices |
| October 19 | Shares touched a 52-week low and closed slightly lower. The 2032 notes yielded 14.9%, and the five-year CDS spread was about 1,082 basis points. Oracle's own CDS spread widened to 315. | |
| October 21 | Shares closed at $61.82. Credit desks told CNBC the high-yield market is "effectively on hold" until after the midterm elections. | |
| October 22 | Shares fell 10.1% to $55.60, the lowest close of the year to that point, on a day the S&P 500 lost 1.2%, the day after Meta said it would create a new class of non-voting shares to "finance its growth", a sign that even the richest AI spenders are looking for new money. Other borrowers that rent out AI chips fell too: Nebius 7.2%, IREN 8.4%. After the drop, CNBC put the five-year CDS spread at 1,145 basis points, up 39% in October. | Yahoo Finance closing prices; |
A bond's yield is the return a buyer locks in at today's price. CoreWeave sold those 2031 notes in April at a 9.75% coupon. A 15.78% yield means buyers now demand about 6 points more a year to hold the same debt, so the price of the notes has fallen. A credit default swap is insurance against a default, and its spread is the yearly cost in basis points (hundredths of a percentage point). S&P Global Market Intelligence's Gavan Nolan told CNBC that 1,000 or more is generally a sign of credit distress.
The latest maturity schedule, from the Q2 10-Q, shows why the price of new money matters: this is a company that has to keep borrowing to repay and grow. The Q3 10-Q will update it.
| Principal due | Amount |
|---|---|
| Rest of 2026 (from June 30) | $4.4bn |
| 2027 | $6.2bn |
| 2028 | $4.4bn |
| 2029 | $2.4bn |
| 2030 | $3.2bn |
| Later | $14.9bn |
| Total | $35.6bn |
Source: Q2 2026 Form 10-Q, note 10. Rows don't add exactly because of rounding. Principal, the $35.6 billion CoreWeave has to repay, is higher than the $35.1 billion balance-sheet figure in the table above, which is net of unamortized discounts and issuance costs.
About $10.6 billion, 30% of the total, is due within 18 months of June 30. For scale, CoreWeave's operations brought in $5.6 billion of cash in the first nine months of 2026. Most of it is amortizing term loans secured on customer contracts, plus equipment financing, rather than bonds. The bonds don't mature until 2030 to 2033. So the sell-off didn't threaten a near-term default. It made the next round of unsecured borrowing more expensive, which is why CoreWeave now says it won't sell more at these yields:
| New unsecured borrowing | At the 9.75% coupon it paid in April | At the 15.78% yield of October 15 |
|---|---|---|
| Yearly interest per $1 billion | $97.5m | $157.8m |
| Yearly interest per $10 billion | $0.98bn | $1.58bn |
An extra $600 million a year on $10 billion of new bonds would be about 5% of 2026 revenue guidance. CoreWeave borrowed $27.0 billion in the first nine months alone, so the price of new money matters more to it than to almost any other AI builder.
CEO Mike Intrator answered in a memo quoted by CNBC: "We are confident in our backlog which more than supports all of our debt obligations." He has a point. The backlog, now about $132.5 billion, is roughly 2.8 times the $46.9 billion of debt on the balance sheet, and Deutsche Bank said around two-thirds of it is with investment-grade customers. But backlog turns into cash only as capacity goes live, while interest is due every quarter. The gap between those two clocks is what the bond market is pricing.
What a share sale would cost shareholders
The other source of money is new stock. In October, Intrator's memo said, according to CNBC, that CoreWeave has "additional flexibility to raise equity over time as market conditions and our capital needs warrant." In Q3 it built the tool for that: an at-the-market program, which lets it sell new shares into the market a little at a time, for up to 35 million shares. CoreWeave had about 559 million shares in Q3 (weighted average,
Q3 release), so selling all 35 million would add 6.3% to the share count and shrink each existing holder's stake by 5.9%, whatever the price.
The CFO's cap of less than $5 billion of company-level money through 2027 limits how far that goes. Here's what raises of different sizes would mean at the November 11 close of $60.77 and at the 2026 closing low of $43.66 on November 5, before any discount to the market price:
| Raised | New shares at $60.77 | Stake shrinks by, at $60.77 | New shares at $43.66 | Stake shrinks by, at $43.66 |
|---|---|---|---|---|
| $1 billion | 16 million | 2.9% | 23 million | 3.9% |
| $5 billion | 82 million | 13% | 115 million | 17% |
| $10 billion | 165 million | 23% | 229 million | 29% |
For comparison, Nvidia, which supplies CoreWeave's chips, paid $87.20 a share when it invested $2 billion in January (CNBC). Money raised at the November 11 close would buy 43% more of the company per dollar than Nvidia's did, and at the November 5 low almost twice as much, so the price at which CoreWeave uses its share program matters as much as the size.
CoreWeave vs Oracle and the hyperscalers
| CoreWeave | Oracle | Microsoft, Amazon, Alphabet, Meta | |
|---|---|---|---|
| Latest capex guidance | $37bn to $39bn (2026) | $90bn to $95bn (FY ending May 2027) | $720bn to $740bn combined (2026), after Meta narrowed its range to $135bn to $145bn ( |
| Capex vs revenue | About 280% to 300% of 2026 revenue guidance | Up to about 106% of FY27 revenue guidance | 27% to 43% of trailing revenue |
| Free cash flow | Negative | Negative | $61.5bn combined over 12 months, negative in the latest quarter ( |
| Main funding | Debt secured on contracts, customer prepayments | Debt, new shares, customer prepayments | Operating cash flow, bonds, and new shares at Alphabet and Meta |
More detail: Oracle capex 2026 and 2027, the
hyperscaler capex tracker, the
hyperscaler debt tracker and
AI spending vs the dot-com bubble.
Our read
CoreWeave is the purest bet on AI compute demand in the stock market, and the most leveraged. The growth is real: Q3 revenue was up 168% on a year earlier, backlog is about ten times this year's revenue, and the customer list now includes Meta and Jane Street alongside Microsoft and OpenAI.
But the model only works if three things hold at once. Customers have to keep paying through contracts that run five years or more. GPUs have to earn their cost back before newer chips make them cheaper to rent. And lenders have to keep refinancing at rates CoreWeave can afford. Interest takes about 23% of revenue, and the $207 million Q3 operating profit was 6% of sales. Q3 did ease the third worry: CoreWeave plans to fund 2027 with loans secured on signed contracts and customer prepayments rather than 15% bonds. That's cheaper, but it ties growth to contracts already in hand. Of all the large AI builders, CoreWeave still has the least room for any of those three to slip.
- Customers keep payingThrough contracts of five years or more
- GPUs earn back their costBefore newer chips make them cheaper to rent
- Lenders keep refinancingInterest takes about 23% of revenue
What we'd watch next: whether Q4 capex really drops to the $6.3 billion to $8.3 billion the guidance implies or the range goes up again, whether the roughly $9 billion of prepayments arrives on time, how many shares CoreWeave sells through its new program, and, in the Q3 10-Q, whether the share of revenue from the top three customers keeps falling. Another delay from a data center developer, like the one that pushed 2025 spending $5 billion to $8 billion below plan, would be the first sign of trouble.
Frequently asked questions
What is CoreWeave's capex guidance for 2026?
CoreWeave guides to $37 billion to $39 billion of capital expenditures in 2026. It raised the low end from $35 billion on November 10, 2026, after starting the year at $30 billion to $35 billion. It spent $30.7 billion in the first nine months on its own measure, including $14.5 billion in Q3, which leaves $6.3 billion to $8.3 billion for Q4. Its 2026 revenue guidance is $13.0 billion to $13.3 billion, so planned capex is about three times revenue.
How much did CoreWeave spend on capex in 2025?
$14.9 billion by CoreWeave's own measure, and $10.3 billion in cash paid for property and equipment. It had originally guided to $20 billion to $23 billion, then cut the range to $12 billion to $14 billion in November 2025 after a third-party data center developer fell behind schedule.
What is CoreWeave's capex per MW?
There's no official figure. Dividing 2026 capex guidance of $37 billion to $39 billion by the roughly 1 GW of active power CoreWeave expects to add during 2026 gives about $37 million to $39 million per MW. That overstates the true per-MW cost somewhat, because some 2026 spending is for capacity that goes live later. Most of the cost is GPUs, servers and networking, not buildings.
How much debt does CoreWeave have?
CoreWeave had $46.9 billion of debt on its balance sheet at September 30, 2026, up from $14.0 billion a year earlier, against $4.1 billion of cash. Net interest expense was $846 million in Q3 2026, and the company guides Q4 interest expense to $1.05 billion to $1.15 billion. At June 30, about $10.6 billion of principal was due by the end of 2027, mostly term loans and equipment financing; its bonds mature from 2030 to 2033. It says it won't issue more unsecured bonds through 2027 at current yields.
Why is CoreWeave stock falling?
In October, because its borrowing costs jumped as interest rates rose and investors began to doubt who will pay for the AI build-out. The stock fell 36% from September 30 to $55.60 on October 22 and reached its 2026 closing low of $43.66 on November 5, before rising to $60.77 the day after its Q3 report. Its notes due 2031, sold in April at a 9.75% coupon, reached a 15.78% yield, and its five-year credit default swap spread rose to 1,145 basis points, past the 1,000 that S&P Global Market Intelligence calls a general sign of distress. Nothing in its reported results changed; the cost of new money did. In its Q3 report on November 10, CoreWeave answered with a funding plan that avoids new unsecured bonds through 2027.
What are CoreWeave's bond yields?
On October 15, 2026, CNBC reported a 15.78% yield to maturity on CoreWeave's 9.75% senior notes due 2031. A day earlier its 9.625% notes due 2032 yielded 14.8%, and they were still at 14.9% on October 19. CoreWeave also has $2.0 billion of notes due 2030, 9.00% notes due 2031, euro 8.5% notes due 2032, 1.75% convertible notes due 2031 and 2032, and the $4.2 billion of 2.875% convertible notes due 2033 sold in September 2026. S&P rates the company B+.
Who are CoreWeave's biggest customers?
CoreWeave's latest 10-Q shows three customers made up 36%, 26% and 10% of Q2 2026 revenue, without naming which is which. The filing names Microsoft and OpenAI as significant customers and describes commitments from Meta (up to about $21.0 billion), OpenAI (up to about $6.5 billion under a September 2025 order) and Jane Street (about $6.0 billion).
Limits of this data
- CoreWeave's "CapEx" and its cash-flow capex differ, sometimes by billions in a quarter.
- Some guidance figures come from earnings call transcripts, and the original 2025 range comes from news reports.
- The capex-per-MW figure is our rough calculation, not a CoreWeave disclosure.
- Bond yields and credit default swap spreads are press-reported snapshots and move every day.
- Q3 debt is the balance-sheet value from the results release; the 10-Q with principal amounts, maturities and customer shares wasn't filed when we updated, so those figures are still from June 30.
- Funding plans and Q3 capex are from the CFO's remarks on the earnings call, as transcribed.
- Backlog depends on CoreWeave delivering capacity on time. Contracts can be delayed or extended, as one was in 2025.
Go deeper with member research
CoreWeave's numbers show how far AI building now runs on borrowed money. Member reports follow that chain:
Issue 07: The Vote That Knocked Almost a Quarter off Data Center Stocks: which data center campuses CoreWeave rents in Texas, what the freeze on new approvals means for the landlords behind its capacity, and whether the post-election sell-off overshot.
Issue 05: NVIDIA Is Building the Credit Market for AI: how GPU capacity is becoming collateral for private credit, insurance balance sheets and securitized debt, the market CoreWeave borrows from.
Comfort Systems and the scarce capacity behind AI construction: why construction and mechanical capacity, the bottleneck that cut CoreWeave's 2025 spending, has pricing power.
Issue 02: The AI Capex Shock: where the next infrastructure dollar goes, who keeps the profit, and what could break first.
Companies in this article
Keep reading
Members-only research
Members are reading this
- Issue 05: NVIDIA Is Building the Credit Market for AI
NVIDIA has found a way to turn knowledge of its own roadmap into a lower cost of capital for its customers. The result could move AI infrastructure from venture equity into private credit, insurance balance sheets and securitized debt.
- Issue 09: Is Nvidia Stock a Buy at 10x Earnings Without Huang?
At $155.46, Nvidia trades at 9.8 times next year's expected earnings, about half the S&P 500's multiple. What the price assumes, and our three calls.
The thesis, the numbers behind it, and what would break it. Full access is $39 a month.
Read the full researchGet a full issue free
Enter your email and we'll send you a complete Yield Theory issue, then the opening of each new one. Unsubscribe anytime.