CoreWeave’s bonds have paid a yield set by a junk rating since the company first borrowed against its chips. The loan it began shopping this week is expected to pay roughly half as much.
The AI cloud provider is seeking at least $4 billion through a loan marketed as investment grade, according to people familiar with the matter. Morgan Stanley and Mitsubishi UFJ Financial Group are leading the financing, which would be syndicated to investors in phases and could launch as soon as next week. Terms remain subject to change.
Early pricing talks point to a yield of 7% to 7.5%, against a prevailing yield on the company’s existing high-yield debt that is about twice as high. For a borrower that has spent two years paying for growth with expensive capital, that gap matters more than the size of the deal.
The structure is what makes the lower rate possible. CoreWeave carries a rating below investment grade, but this borrowing is secured against its computing hardware and the customer contracts attached to it, which lets lenders underwrite cash flows rather than a balance sheet. The company used the same approach in March, when it completed an $8.5 billion facility backed by long-term contracts with Meta Platforms. That deal carried a floating-rate tranche priced 2.25 percentage points over the secured overnight financing rate and a fixed-rate tranche near 5.9%.
The financing follows a strong quarter. CoreWeave reported revenue of $3.66 billion for the three months ended September, up 168% from a year earlier, and on Nov. 10 raised its full-year forecast to $13 billion to $13.3 billion, ahead of the $12.9 billion analysts had expected. The shares rose 24.7% after the release and then gave back 7.8% as the wider market weakened and reports of turmoil at Nvidia, its largest supplier and an anchor investor, unsettled holders.
Interest is the line the loan addresses. Net interest expense reached $846 million in the quarter, against $310 million a year earlier, an increase of 173% that ran slightly ahead of revenue growth. The net loss was $709 million, or $1.27 a share on an adjusted basis.
A GPU cloud buys chips, installs them in leased buildings and sells the resulting compute on contracts that usually run several years. The hardware depreciates on a schedule the company chooses; the debt matures on a schedule its lenders choose. When those two calendars diverge, a business can grow quickly and still be short of cash.
Secured borrowing also rearranges the queue. Lenders with a claim on specific equipment are paid from that equipment before unsecured bondholders are paid from anything, which is why they accept a lower coupon. Existing noteholders are left with the residual, and they have already accepted a yield that assumes the residual is thin.
The collateral has a price, and the price moves. Accelerators do not hold value the way buildings do. Each generation makes the last one cheaper on the secondary market, and the buyers are often the same competitors who would be hurt by a glut. A lender that had to seize racks of four-year-old chips would be selling into whatever demand existed that month.
Lenders accepted the earlier version of this trade because the contracts sit underneath it. Take-or-pay agreements with large customers, several of them investment grade, give a lender a claim on revenue that survives a downturn in spot pricing. That is the strongest evidence available that the capacity is worth what CoreWeave says it is. It is also evidence about a handful of customers, not about the market.
The competitive picture has not changed with the financing. CoreWeave is the largest of the neoclouds, a group that rents Nvidia accelerators by the hour, and smaller rivals have listed or are preparing to. The big cloud providers keep building capacity of their own. At the margin, price is set by whichever operator happens to have idle machines that week.
Management’s own forecast hints at where it thinks the buildout goes. Capital spending of $37 billion to $39 billion for 2027 is enormous in absolute terms and close to flat against the current pace. A company that expected demand to compound for another three years would be unlikely to describe a ceiling.
Cheaper debt does not change the arithmetic of the business, only the cost of financing it. It buys time to show that contracted revenue can cover depreciation and interest in the same quarter, and it lowers the chance that a bad month in the credit markets forces a decision the company would rather not make.
The question left open is who absorbs the loss if the price of rented compute falls while the debt stays where it is. This week the answer moved a little further from CoreWeave’s shareholders and a little closer to the lenders who agreed to be paid first.
