The CoreWeave 2032 Bond Yielded Above 16 Percent. The Coupon Is 9.625.

In short
Key takeaways
- CoreWeave's 2032 bond carries a 9.625 percent coupon and yielded above 16 percent on 6 November 2026, about 640 basis points over its own coupon.
- The company's five-year credit default swaps spiked above 1,289 basis points on 5 November, past the October peak, then eased 8 percent to around 1,155.
- Five-year CDS rose 76 percent on Nvidia, 49 percent on Broadcom and 44 percent on Amazon in the month to 6 November.
- Absolute spreads were roughly 139 basis points for Nvidia, 298 for Oracle and 1,100 for CoreWeave, so the percentage moves and the risk levels rank the names differently.
- Five AI bond deals totalling 67 billion dollars were issued from 29 October, against 296 billion dollars of hyperscaler debt issuance for the year to date.
On 6 November 2026 the yield on CoreWeave's 2032 bond climbed above 16 percent. The bond carries a 9.625 percent coupon, so buyers were asking for roughly 640 basis points more than the interest the company had already committed to pay. The move did not arrive alone. The previous day, CoreWeave's five-year credit default swaps spiked above 1,289 basis points, widening past their October peak, before easing about 8 percent to around 1,155 on the Friday.
A bond yield and a swap spread are different instruments. One is the return an investor requires to hold a fixed coupon to 2032. The other is the annual cost of insuring against non-payment over five years. On this occasion they told the same story, and the story was set by dealers quoting into a week when the conditions around the underlying business changed.
What sixteen percent is measuring
The coupon on the bond is 9.625 percent and does not move. A yield above 16 percent means the price has fallen far enough that the remaining payments return more than 16 percent a year to whoever buys now, if the company pays as promised. The 640 basis points between coupon and yield is the discount the market applied to that promise.
Whether 640 basis points is a lot depends entirely on the comparison. Against a Treasury curve it is a number without much meaning, because the risk being priced is not interest rates. Against the same instrument's coupon it says the market repriced this specific credit in a short window. Against other names in the same sector it says something more useful, which is where in the capital structure of the AI buildout the repricing started.
The spread across the sector is the interesting number
Over the month to 6 November, credit default swaps rose 76 percent on Nvidia, 49 percent on Broadcom and 44 percent on Amazon. The absolute levels were roughly 139 basis points for Nvidia, 298 for Oracle and 1,100 for CoreWeave. The report made the qualification itself: a large increase from a low starting point does not imply the same credit risk.
Read the two sets of numbers together and the ordering separates. Nvidia's 76 percent is the largest percentage move and the smallest absolute one, a rise from a base consistent with a large investment grade issuer. CoreWeave's move is both the highest level and, at roughly eight times Nvidia's spread, a different category of price. Amazon's 44 percent is the smallest percentage move and lands somewhere in between on level.
That pattern is what a sector repricing looks like when it is applied across a credit spectrum rather than to one name. The market did not decide that Nvidia had become a risky company. It raised the cost of protection on the sector, and the existing ranking of balance sheets was preserved through the move. The weakest name carries the widest spread and its spread widened most in absolute terms. Nvidia's widened least in absolute terms while widening most in percentage terms, which is what happens when you start from 139.
There is a version of this story that is only about CoreWeave. The cross-sector numbers are what argue against it, and they are also the thinnest part of the evidence, because three names in one sector is a small sample and two of them are the largest companies in the industry.
Why the neocloud is where it shows up first
CoreWeave's revenue comes from long-term compute contracts, and the neocloud model leans on customers pre-paying for capacity. That structure matters more than the headline spread suggests. A customer that stops pre-paying does not simply reduce revenue; it removes the cash that services the debt raised to build the capacity in the first place.
The reported mechanism runs the other way through the same pipe. Any sign of consolidation or retrenchment in capital spending across the industry could mean fewer or shorter data centre compute contracts, or delays in prepayments. Delays, not cancellations, are enough to matter, because the debt has a schedule and the prepayment does not.
The maturity does some of the work too. A 2032 bond is roughly six years of credit risk, and the equity and the contracts can move a long way inside six years without any default. That is why a yield on one instrument is a weak summary of a company's whole funding position. CoreWeave's other debt, its maturities, and how much of the stack is secured or floating are not in the reporting, and a single bond's yield cannot stand in for them.
Four pressures that arrived together
The first was political. The 3 November midterms put Democrats in both chambers, and data centre opposition became visible on election day, including a protest in Leavenworth County, Kansas. Goldman's estimate, reported by CNBC, was that moratoriums could affect 45 percent of planned US capacity, with much of that in Texas. The market was then working through Democratic governors in Ohio, Texas and Georgia, the three states holding much of the planned capacity. Texas governor-elect Gina Hinojosa has said she is not opposed to data centres but wants a moratorium until the legislature passes protections for communities.
The second was on the demand side. Goldman reported that hyperscalers collectively ran negative free cash flow in the third quarter, with capital spending exceeding operating cash flow. That is not a crisis on its own; it is the arithmetic of a buildout in progress. The sharper point in the same note is that growth in purchase commitments is a separate and sizable use of cash from capital expenditure. If the customers funding their own construction are also pre-paying suppliers, the prepayment channel is fed by the same balance sheets under strain.
The third was the price of issuance. Five AI bond deals totalling 67 billion dollars began issuing from 29 October, against 296 billion dollars of hyperscaler debt issuance this year. When that much paper arrives in a fortnight, the marginal buyer sets the price for everyone, and the sector's protection costs rose together.
The fourth was the response. National Economic Council director Kevin Hassett told CNBC on 6 November that he expects data centre growth to continue as forecast despite the election result. The same day, the President endorsed the BAAJA permitting bill in a social media post, highlighting its data centre provisions; Reuters reported on 9 November that the bill would require large data centre operators to bear more of the cost of the grid upgrades needed to serve them. Both are statements about the pipeline. Neither is an instrument in the credit market, and the instruments moved anyway.
Two notes from the desks
Two of the quoted notes are worth recording, because both describe the same mechanism the numbers do. Deutsche Bank's Jim Reid wrote that there had been plenty of warnings from the credit space, and that debt and equities were now moving more in sync. Jefferies' Christopher Wood put the end of the capex race this way: the most plausible way it ends is that the credit card is taken away.
Both are opinions from named desks, quoted in the report and not evidence in themselves. The first is at least a measurable claim, since correlation between a company's bonds and its shares can be computed. The second is a description of a funding structure, and the structure is the part that can be checked: a buildout financed with debt against contracts with counterparties whose own credit is repricing does not need to fail commercially in order to stop. It needs only to become unfundable at the price on offer.
Two forecasts that disagree on timing
Goldman's numbers were 440 billion dollars of hyperscaler debt issuance next year, plus 300 billion dollars of project level issuance across investment grade and high yield, against 296 billion dollars this year. Bank of America's were a 31 percent rise in AI related credit issuance from 2026 to 2027 to 685 billion dollars including public and private debt, then a 23 percent fall to 531 billion dollars in 2028.
These are not the same series. One counts hyperscaler debt and adds a separate project level figure; the other counts AI related credit broadly. The comparison cannot be made cleanly, and a reader should not try. What is comparable is the mechanism both describe: issuance depends on the credit market's willingness to keep buying, and Goldman's wording names the risks to further issuance as delayed monetisation and growing purchase commitments, subject to market conditions.
The BofA 2028 decline is the part worth flagging precisely, because it is a published expectation of a peak rather than a record of one. It is one bank's call, and its own stated downside is a slowdown in AI infrastructure projects after the Blue Wave, which would reduce capital needs and therefore debt market funding. The forecast was struck eleven days before the vote that produced the outcome its downside case describes.
What has been priced and what has not happened
No moratorium has been enacted in Texas, Ohio or Georgia. The 45 percent figure is the share of planned capacity that could be affected, a conditional, and Hinojosa's position is a stated preference pending legislation rather than an executive action. The credit market is pricing a probability, not recording an event.
That distinction is the reason the political and financial halves of this story should not be merged. The bond yield is a fact about 6 November. The moratorium is a scenario with an author. A reader who treats the second as though it had the status of the first has skipped the step where the evidence actually sits.
The administration's claim and the market's claim are also not in contradiction. Saying growth continues as forecast is a claim about construction. A bond yielding 16 percent is a claim about the cost of financing it. Both can be accurate at once, and the second one constrains the first without refuting it.
One absence is worth noting. The Data Center Coalition has not endorsed the permitting bill, and no industry body has publicly objected to the spread widening. A sector that believed the repricing was wrong had an obvious reason to say so, and did not.
Where this could be wrong
There is no document. A bond yield is not filed anywhere; it is a dealer quote aggregated for clients and reported second hand. Every figure here comes from one CNBC report on 6 November 2026, which attributed its data to Deutsche Bank, Goldman Sachs, Jefferies and Bank of America. I did not pull a pricing source, a swap data repository, or a CoreWeave filing. Anyone who needs the level should check a live pricing source rather than a week-old news figure.
The 16 percent applies to one instrument. CoreWeave's total debt, its maturity ladder, and how much is secured or floating were not established, and a debt stack dominated by short dated secured borrowing would make this bond a much weaker signal about the company's overall cost of funds than it appears.
The claim that this is a sector repricing rests on three names. If other neocloud and data centre real estate credit did not move, the CoreWeave widening was idiosyncratic and the political frame is the wrong one. Three large caps in related businesses is suggestive, not decisive.
A spread of roughly 1,100 basis points implies a meaningful probability of a credit event, but a spread cannot tell you whether that event is a restructuring, a downgrade, or a large contractual dispute. Nobody has published the decomposition, and it is not derivable from the price.
The four sources in the ledger are the four documents that exist: one CNBC report, one CNBC interview, and two Reuters pieces on the political side. There is no primary record of the bond terms, no comparison set of neocloud spreads, and no reading of CoreWeave's own disclosures. That is a short ledger for a post of this length, and it is short because the topic has no paper behind it.
On this page
Sources
- The AI credit card is getting costlier as CoreWeave bond yield tops 16 percentNick Wells and Michael Bloom, 2026
- US Congress gets first post-Blue Wave crack at data centers with permitting billReuters, 2026
- Trump's midterm drubbing demonstrates power, data centre opponentsReuters, 2026
- Data center growth to continue as forecast, Hassett saysCNBC, 2026
Frequently asked questions
Why does a bond with a 9.625 percent coupon yield more than 16 percent?
Because the price fell. The coupon is fixed at issue, so the only way the yield rises to 16 percent is for buyers to pay well below par for the remaining payments. The extra return is compensation for the chance those payments are not made as scheduled, and for holding a fixed coupon while the market's assessment of that chance changed.
What is the difference between the bond yield and the credit default swaps?
The bond yield is the return an investor requires to hold the debt to maturity in 2032. The five-year swap spread is the annual cost of protection against non-payment over five years. They are different instruments with different maturities, and on 6 November 2026 they moved together: the swaps spiked above 1,289 basis points on 5 November, then eased to about 1,155 while the bond still yielded above 16 percent.
Does a 76 percent rise in Nvidia credit default swaps mean Nvidia is in trouble?
No. Nvidia's absolute five-year spread was roughly 139 basis points, which is consistent with a large investment grade issuer. A 76 percent rise from that base is a small number of basis points in absolute terms. CNBC made the point directly: a large increase from a low starting point does not imply the same credit risk. The direction is informative; the level is what ranks the names.
Is this a forecast that the AI buildout will stop?
No. It is a report of prices observed on 6 November 2026. The data centre buildout is a physical programme and the bond market is a financing cost. Higher financing cost changes the terms on which the programme can continue, and the size of that change is what the spreads measure.
Where do these figures come from?
From CNBC's report of 6 November 2026 by Nick Wells and Michael Bloom, which attributed the market data to Deutsche Bank, Goldman Sachs, Jefferies and Bank of America. No filing or exchange record carries a bond yield; these are dealer quotes aggregated for clients, so the figures here are second hand by construction.