What was actually being sold
Start with the instrument, because the name is doing a lot of hiding. CoreWeave was selling a first-lien delayed-draw term loan of $2.6 billion, maturing September 2031.
Translated: it is closer to a corporate credit card than a mortgage. The lenders commit the full $2.6 billion up front, but CoreWeave only draws it as it needs it, and only pays interest on what it has actually drawn. On the undrawn balance it pays a standby fee of 50 basis points — half a percent a year for the right to call the money later. The drawing window closes in December 2026.
First-lien means these lenders stand at the front of the queue against the pledged assets if things go wrong. That queue position is the whole argument, and we will come back to it.
The money buys GPUs and the equipment around them, to serve compute contracts with Anthropic, Jane Street, Midjourney, Hudson River Trading and Anysphere. Those contracts are structured take-or-pay: the customer pays for the reserved capacity whether or not it ends up using it.
That word matters more than any hardware specification in the deal. A lender is not really lending against graphics cards, which are depreciating, mobile, and hard to repossess profitably. It is lending against a stream of payments that five named companies have contractually promised to make. The GPUs are the security. The contracts are the repayment source.
The nine days
Syndicated loans are not sold at a fixed price. The arranger — here JPMorgan as administrative agent and joint lead arranger, alongside MUFG — publishes indicative terms, holds a lender call, and takes orders. If the book fills easily, terms tighten in the borrower's favour. If it does not, they widen. The market sets the clearing level in public, over days.
CoreWeave launched on 17 July. The lender call was Monday 20 July. Commitments were due at noon on Thursday 30 July.
| Term | Opening talk, 17 July | Final, 30 July |
|---|---|---|
| Spread over benchmark | SOFR + 425–450 bp | SOFR + 550 bp |
| Issue price | 99 | 97 |
| Yield to maturity | 8.53% – 8.79% | 10.44% |
Two levers moved, and they compound.
The spread is the margin over the floating benchmark rate. It went up by 100 to 125 basis points — a basis point being one hundredth of a percentage point, so this is between one and one and a quarter percentage points of additional annual interest, every year, on everything drawn.
The issue price is the subtler one, and it is where most readers lose the thread. A loan issued at 97 means lenders hand over 97 cents for every dollar they are owed. Put $970 million in, be repaid $1 billion. That three-cent discount is not a fee; it is extra return, earned at maturity, and it lifts the effective yield above the headline spread. Moving from 99 to 97 doubled the discount.
Together they took the yield from roughly 8.5% to 10.44%. On the full $2.6 billion, if it is entirely drawn, the difference is on the order of $50 million a year — SGC arithmetic on the reported yields, not a company disclosure, and lower in practice while the facility is only partly drawn.
The same borrower, three prices
The interesting comparison is not CoreWeave against the market. It is CoreWeave against CoreWeave, ten weeks earlier.
Exhibit 1
What CoreWeave paid to borrow, 2026
Spread over the benchmark rate, in basis points. One borrower, four deals, ten weeks.
In May, CoreWeave reported an $8.5 billion non-recourse, investment-grade-rated facility whose floating tranche priced at SOFR plus 225. Also in May, it closed a separate $3.1 billion facility at SOFR plus 450 — described at the time as the first publicly syndicated financing vehicle backed by HPC infrastructure. That deal reportedly drew $19 billion of orders. It was roughly seven times oversubscribed.
Ten weeks later, a structurally comparable GPU-backed loan needed 550 and a two-point discount to clear.
Read the ladder carefully and it is not one number. It is three different promises, priced accordingly. The 225 deal was non-recourse and rated investment grade — a ring-fenced structure lenders were willing to treat as high quality on its own terms. The 450 deal was GPU-backed and rated below investment grade. The 550 deal was the same species as the 450, sold into a market that had changed its mind in between.
What the lenders wrote into the documents
Price is the visible concession. The documentation is the durable one, and it usually tells you more about what lenders are actually afraid of.
Three protections are on the public record. There is a 1.35× debt service coverage ratio covenant at the borrower level: for every dollar of principal and interest due, the structure must generate $1.35 of cash available to pay it. Breach it and the lenders get rights they did not previously have. There is a $112.5 million minimum liquidity requirement, running until twelve months before the September 2031 maturity. And the financial covenants remain in force until the parent reaches an investment-grade rating — the protection relaxes only when the credit improves, not when the borrower would like it to.
A covenant is simply a promise with a tripwire attached. Without one, a lender who becomes worried can only wait. With one, deterioration triggers a conversation while there is still something to negotiate over. The presence of a maintenance covenant on a large syndicated loan in 2026 is itself the signal: for most of the past decade, borrowers in strong markets have not had to grant them.
Fitch assigned the facility BB+ with a recovery rating of 2, while affirming CoreWeave's corporate rating at BB−. The facility is rated two notches above the company that owes the money. That gap is the collateral and queue position doing visible work: the agency's view is that if CoreWeave failed, these particular lenders would recover more than the company's general creditors. It is a precise statement about structure, not optimism about the borrower.
The credit market was making the same point elsewhere
While the loan was being sold, the cost of insuring CoreWeave's existing debt was climbing.
Exhibit 2
The cost of insuring CoreWeave's debt, July 2026
Five-year credit default swap, in basis points per year.
Several outlets converted that 855 into an implied probability of default of roughly 50% over five years. Treat that number with care. Converting a swap price into a default probability requires an assumption about how much lenders would recover after a default, and the standard convention is a rough market default, not a company-specific forecast. Change the recovery assumption and the headline probability moves substantially. What the price reliably tells you is the cost of transferring the risk, and that had gone up by more than half in a month.
Nor was this one company's problem.
Exhibit 3
This was not only CoreWeave
Five-year CDS levels reported in late July 2026. Each row states its own base period, because the reported comparisons do not share one.
| Borrower | Base level | Base period | Late July 2026 |
|---|---|---|---|
| CoreWeave | ≈550 (derived) | 1 July 2026 | ≈855 |
| Oracle | ≈145 | end 2025 | 215 (record) |
| Meta | ≈56 | end 2025 | ≈95 |
| Amazon | 36 | start 2026 | ≈68 |
Oracle's five-year protection reached a record. Amazon's and Meta's roughly doubled from their respective starting points. Moody's warned in July that the scale of AI capital spending was pressuring credit quality across the largest technology borrowers. On 29 July, CoreWeave shares fell about 9% and Nebius about 10% as the swap move fed into equity. Trading in AI-linked credit protection reached around $650 million in the second quarter, reportedly up about 600% year over year — a derivatives market forming around a risk that did not previously need one.
Why this borrower, and why now
Scale explains the sensitivity. CoreWeave is expected to spend more than $34 billion on AI infrastructure this year. In the first quarter of 2026 its interest expense doubled to $536 million and free cash flow was negative $4.71 billion. None of that is hidden or surprising: it is what building capacity ahead of contracted demand looks like on a cash-flow statement.
But it does mean the company must return to the debt markets repeatedly, and a borrower who must come back frequently is exposed to the market's mood on each visit. In May the mood was $19 billion of orders. In July it was a 125 basis point flex and a maintenance covenant.
It is worth being precise about what did not change. The offtakers are the same names. The contracts are still take-or-pay. The GPUs are the same equipment. Investors did not receive new adverse information about Anthropic, Jane Street, Midjourney, Hudson River Trading or Anysphere. What changed was the quantity of similar paper competing for the same pool of money, and the price at which that pool was willing to take concentrated exposure to a single sector.
What we would need to see to call a turn
A single repriced loan is a data point, not a trend. Six indicators would tell us which way this is actually going, and each is observable without private information:
- Whether the next GPU-backed syndication clears inside 550, and at what issue price.
- Whether maintenance covenants persist once spreads stabilise, or are the first concession borrowers claw back.
- Whether pricing begins to separate by offtaker credit quality — an investment-grade counterparty and a venture-funded one should not clear at the same spread indefinitely.
- Whether ratings agencies keep notching facilities above the corporate rating, or narrow the gap as recovery assumptions are tested.
- Whether CDS and primary loan spreads reconverge, or the insurance market continues to price a risk the loan market will not.
- Whether any borrower discloses a covenant test result — the first real evidence of whether 1.35× was conservative or optimistic.
SGC tracks the first, third and fifth of these directly. The gap this transaction exposes is not a shortage of prices; it is the absence of a comparable, auditable record of what a compute credit actually promises — the contract term, the assignability of the receivable, the collateral control, and the remedy — held to the same standard as a quoted GPU-hour. That is the layer we are building toward.
This note separates reported facts from SGC analysis and marks derived figures as derived. It is market-structure research, not investment advice, a credit rating, a solicitation, or an offer to lend, arrange financing or trade any instrument. SGC has no position in, and no commercial relationship with, any company named. Sources were checked on 6 August 2026. Send corrections to info@sovereignglobalcompute.com.