In May, CoreWeave syndicated a $3.1 billion delayed draw term loan backed by GPU infrastructure and the customer contracts running on it. It was the first financing of its kind broadly syndicated into the US leveraged loan market, a real test of whether a wide investor base would treat compute as a financeable asset.
The answer was emphatic. Reported orders ran to roughly $19 billion against a $3.1 billion deal, pricing tightened 50 basis points from initial talk and landed at SOFR plus 450 with a 99 issue price. A book covered six times over is a very good day.
Ten weeks later the same borrower returned with a $2.6 billion facility of the same type. Initial talk was SOFR plus 425 to 450 at 99, roughly where the previous deal had cleared.
It did not clear there. The spread flexed wider to SOFR plus 550. The issue price widened to 97. Documentation was tightened, including a debt service coverage covenant at the borrower level and a standing liquidity requirement. Final yield to maturity landed above 10.4 percent.
Same issuer. Same structure. Same asset class. Roughly 150 basis points of additional all-in cost, plus covenant concessions, in ten weeks.
The two facilities are not identical. Each is supported by a different set of customer contracts, and investors will form their own views on how those compare. That is part of a progression the market has worked through all year, from early deals anchored by investment grade counterparties toward structures supported by a wider range of contracted offtake.
The direction of the flex matters more than the level
Spreads move, and a hundred basis points in a quarter is not by itself remarkable. What is remarkable is the reversal in who was setting the price. In May the deal tightened into aheavily oversubscribed book, meaning the borrower had pricing power and investors competed for allocation. In July it flexed wide and gave up two points of issue price to get done, meaning the marginal lender dictated terms and the borrower accepted them.
That is not a repricing of risk. That is a change in who is in charge, and it happened inside a single quarter.
The documentation point deserves as much attention. When lenders take spread alone, they are pricing risk they already understand. When they take spread and covenants together, they are underwriting the asset class. A coverage test and a liquidity requirement are lenders buying the right to intervene early, which is what you ask for when you are no longer confident you can predict the downside.
The picture is consistent across the capital structure. A data center project financing leased to the same operator priced in the neighborhood of ten percent this month. High yield paper from the same issuer has traded below par at yields above eleven. Secondary bond prices have swung several points in both directions inside a single week while remaining meaningfully lower than six weeks ago. Even investment grade spreads for the hyperscalers have widened.
Why it probably does not mean prices fall tomorrow
None of this is a demand signal. Nobody in the physical market has reported that customers stopped buying. Utilization is not collapsing, rate cards are not being cut. The tape moved and the underlying business did not.
What moved is the cost and availability of the capital funding the buildout. That is a financing event, not a demand event, and the two have very different consequences.
It is also worth being skeptical that this reverses quickly. Scott Goodwin, a sharp credit investor, made the point publicly that portfolio construction in credit is governed by notional and duration in a way equity is not, because credit lacks the payout profile that compensates for concentration. The implication is that as AI related issuance keeps growing, the marginal buyer of that paper needs more spread simply to keep their book balanced, regardless of what they think of the business.
Add a heavy forward calendar of data center supply, higher base rates, and policy as a live wildcard, and the case that cost of capital snaps back tight in the next few months is weaker than it looks.
A longer cycle is not necessarily a worse one. But it is a different one to plan for.
The bridge to the physical market
So what does any of this have to do with someone buying or selling GPU hours?
Think about what sits inside a compute rate card. Hardware, amortized over some assumed life. Power. Colocation and operations. And the cost of the money that bought the hardware in the first place.
Three of those are visible and quoted. The fourth is not. It never appears on a rate card, and it just moved 150 basis points.
That tension has to resolve somewhere. It can resolve in price, and prices have not moved. It can resolve in margin, which is not sustainable. Or it can resolve in terms, which is where we would expect to see it first and where it is hardest to observe from outside.
Terms are the transmission mechanism. Not price.
If you buy compute
The first order consequence is that you should expect to be asked for more money up front. Operators fund capital expenditure from three places: debt, equity, and customer prepayments. When debt gets more expensive and equity is scarce, the third does more of the work, because it is the only one that does not require a syndicate to agree.
Which is worth sitting with, because most compute buyers have never framed it this way. When you prepay for capacity, you are extending unsecured credit to your supplier. Cash handed over today against delivery over the next one to three years, ranking behind their secured lenders, with no collateral and no covenants. That is a credit decision whether or not anyone called it one.
The second order consequence runs in your favor.
Lenders size loan tenor to contract tenor, because contracted revenue is what repays the loan. Tighter conditions make them more conservative, which makes them want longer and better contracted revenue before they will fund anything.
Which means the value of your signature just went up. A buyer willing to commit for twenty four or thirty six months is worth materially more to an operator today than six months ago, because that commitment is what makes the operator financeable at all. If you can sign longer you should be paid for it, in price and in structure, and the gap between what a one year buyer and a three year buyer can negotiate will widen from here.
If you are not willing to sign longer, it is worth knowing precisely why. Fear of being stranded on the wrong hardware generation, uncertainty about your own demand, inability to fund the upfront, or a view that prices fall. Each has a different structural answer, and only one is really about price.
If you sell compute
Two things follow.
Do not build a plan that requires refinancing cheaper. The path where cost of capital compresses fast enough to bail out a marginal structure is not the base case right now. If you can term out funding, the case for doing it sooner is stronger than it was.
Look harder at the lease channel. The blocker on a lot of middle market compute financing is not the debt, it is the sponsor equity underneath it. A lender advancing seventy percent against cost leaves thirty percent that somebody has to write a check for, and that check is what stops deals closing. Equipment leasing gets much closer to full funding, because the lessor buys the hardware and you pay rent, which changes what you need in hand on day one. That channel holds more capital than most operators realize, and it is not the same set of firms as private credit.
If you finance compute
The question this market has not answered, and the one everything else depends on, is what a fleet is worth partway through its life. Every advance rate, every recovery assumption, every lease with real residual risk in it, and every downside case in every credit model is an implicit view on that number. In the structures currently in market, the assumption is generally that decline is smooth and gradual. That is how depreciation is usually modeled and it is a reasonable convention for most equipment.
We are not sure it describes this asset. Compute has a second expiry date alongside its contractual one, which arrives when the models people actually want to run need more than the hardware can provide. That does not happen gradually, it happens on a release date. A structure that survives a smooth thirty percent annual decline is not necessarily one that survives a sixty percent quarter, even if the two average out.
Nobody has a repossession and remarketing history for this asset class yet, which means the first fleet that comes back to a lender and gets re let will set the recovery assumption everyone else marks against. That is a lot of weight on a single data point, and it is the largest unpriced variable in compute credit.
Where this leaves us
The credit market repriced this quarter and the rate cards did not follow. The adjustment is happening somewhere less visible: in upfront payments, in tenor, in structure, and in who is prepared to fund what. Anyone transacting in compute is now downstream of a credit cycle whether they think of themselves that way or not.
Our team at Pluto sits in the middle of this, between the people who buy compute, the people who own it, and the people who finance it. We come from a mix of backgrounds across credit, trading, engineering and the cloud businesses themselves, and we think the most interesting questions over the next year come from the seams between those three groups rather than any one of them.
We would genuinely like to hear from you.
If you buy compute, we are focused on the trade off between tenor, price and what you are asked for up front, and we are happy to talk through how others are approaching it.
If you own compute, we are building a picture of who is actually lending and leasing into this market, at what size, and against what. It is early, and it is a lot more useful shared than hoarded.
If you lend or lease into compute, we would like to compare notes. We are spending real time with originators and credit teams on how to get comfortable with an asset that has no recovery history, and we do not think anyone has the answer yet.
One last thought. Everything above describes risks that are real, widely held, and almost entirely unhedged. Operators are long hardware they cannot sell forward. Buyers are long a price they cannot lock. Lenders and lessors carry residual exposure with no way to lay it off. Every mature commodity market eventually builds the tools to move that risk to whoever wants it, usually through cash settled contracts that never require anyone to deliver a single machine. We expect compute to follow the same path, and building that layer is a large part of what we are working on.
This is the first of what will be a regular note. We will keep writing about what we see across the physical market, the financing market, and the risk transfer layer starting to form between them.
Reach us at stanley@pluto.trade and ronit@pluto.trade, or just reply to this.