Last week a competitor called a bond yield insane. That is the part of the story worth keeping.
The number was 15% โ the yield GMI Cloud's CEO attached to CoreWeave's debt, in a quote that traveled from Crypto Briefing into every AI-infrastructure channel I monitor. The timeline argued about whether 15% was high. Almost nobody argued about the number sitting one layer beneath it: six years.
CoreWeave depreciates its GPUs over roughly six years. NVIDIA ships a new architecture every twelve to eighteen months โ A100, H100, H200, B200/GB200, Rubin. The book life of the asset runs four to six times the cadence of the thing that makes it obsolete. That gap, not the yield, is the credit event. A 15% coupon on a six-year schedule is not a lender being emotional. It is a lender pricing a residual value curve it cannot see. The bond market is not questioning CoreWeave's execution. It is questioning the denominator.
Strip the branding and CoreWeave is a leasing finance business. Borrow, buy assets that depreciate fast, rent them out on medium-term take-or-pay contracts, refinance when the debt matures. Three facts define the risk.
Revenue is concentrated. Public disclosures put Microsoft above 60% of 2024 revenue; that share fell in 2025, but the concentration is structural. Long contracts are simultaneously the valuation base and the counterparty credit exposure.
Capex is enormous. The 2025 guidance sits in the tens of billions against revenue an order of magnitude smaller. Free cash flow is deeply negative. The company cannot self-fund. It must return to the capital markets, repeatedly, on whatever terms are offered.
The asset is the collateral. Lenders take GPU-backed security. That is the entire thesis and the entire fragility.
Scale the backdrop. Across 2024โ2025, AI data-center debt and structured finance expanded into a systemic register: hyperscaler SPV joint ventures, AI data-center ABS, GPU-collateralized loans, with issuers spanning CoreWeave, Oracle-linked ventures, xAI, Vantage, and others. The buyer base is not equity. It is private credit funds, insurers, pensions, and ABS investors. That is the audience the 15% is speaking to, and it is the audience that will transmit any repricing into the broader fixed-income market.
GMI Cloud sits on the other side of the table. It is an NVIDIA Cloud Partner in the small-to-mid tier, US and Asia-Pacific footprint including Taiwan, oriented toward regional and sovereign AI. Smaller balance sheet, thinner collateral pool, lower or absent rating. Its own funding cost is almost certainly higher than CoreWeave's. When a smaller, more fragile competitor calls a larger one's yield insane on the record, that is not analysis. It is positioning. Read the quote as a data point about GMI Cloud's capital anxiety. Read the yield as a separate, more useful signal.
Start with the mechanics. Under a six-year straight-line schedule, a $30,000 accelerator books roughly $5,000 of depreciation a year, about $0.57 per GPU-hour across continuous runtime. Every dollar of rental revenue above that line reads as EBITDA. Shorten the schedule to three years and the same card must earn $1.14 per hour before it contributes anything. Nothing about the hardware changed. Only the assumption did.
The assumption matters now because the rental price is falling. H100 hourly rates have compressed from the ~$8 range at the 2023 peak toward $2โ3 on spot channels. The market is telling you cash flows decay faster than a linear schedule implies. The GPU still works. It just earns less, because the marginal buyer of compute routes to the newest silicon.
This is the confusion the bull case exploits. Demand growth and asset depreciation are not opposites. Total AI compute demand can rise indefinitely while the residual value of any given generation collapses. Incremental demand goes to B200, not to the H100 pledged as collateral. A bull saying "AI demand is exploding" and a lender saying "that card is worth less every quarter" are both correct, and they are describing different objects.
Run the re-rating. Take a simplified operator: 10,000 accelerators at $30,000 each is $300M of hardware. On a six-year schedule, annual depreciation is $50M. On four years, $75M. On three years, $100M. If that operator generates $120M of rental revenue a year, EBITDA reads $70M under the six-year assumption, $45M under four years, and $20M under three. Now divide the same debt by those numbers. The leverage ratio moves from comfortable to distressed without a single dollar of revenue changing. The depreciation schedule is the highest-leverage accounting assumption in the entire neocloud sector, and it is set by management, not by the market. That is precisely what a bond investor prices when demanding a double-digit yield. They are not betting against CoreWeave. They are betting against the schedule.
GPUs are poor collateral. State it plainly.
No mature secondary market. No standardized valuation benchmark. No auction history from which to model a recovery rate. And a jurisdictional split most credit models ignore: an H100 inside China and an H100 outside China are economically different assets because of export controls. One serial number, two prices, and one of them is not realizable. A lender marking collateral to a single global price is marking to a fiction.
Which raises the covenant question nobody has answered on the record. Does the debt carry loan-to-value maintenance covenants tied to collateral appraisal? If it does, a GPU price decline forces additional collateral or cash. That is a margin call on a depreciating asset. The spiral writes itself: prices fall, LTV breaches, the borrower posts more collateral or sells, prices fall further.
In 2022 I dissected Mercurial Finance for a post-mortem. Different mechanism โ improper risk parameterization in a lending pool โ identical failure shape. A parameter set by people who assumed an asset's price was stable, stress-tested against a world where it was not. The protocol did not fail because the code was wrong. It failed because the assumption beneath the code was. The code doesn't depreciate. The accountant does.
So what does 15% actually price? Not incompetence. Residual-value uncertainty. The lender is saying: I will lend against this collateral, but I need to be paid for the possibility that recovery value is half what the schedule claims. A 15% yield on a secured GPU-backed instrument is the market's estimate of how wrong six years might be.
Two questions decide everything before anyone quotes the number. Secured or unsecured? Coupon or yield-to-maturity in the secondary market? A distressed bond trading at a 15% YTM is a different object from a new issue carrying a 15% coupon. The first says the market repriced an existing claim. The second says the issuer had no cheaper option. Same figure, opposite diagnosis. I have not seen the indenture. Neither, I suspect, have most of the accounts quoting it.
Here is where the yield actually bites. Most existing debt is already priced โ fixed or floating loans struck in friendlier windows. The legacy interest bill does not move because a secondary quote moved. The damage is at the margin: the cost of the next dollar, and the dollar after that. For a company that must fund tens of billions in capex and cannot self-fund, marginal debt cost is the business model. Raise it 300 basis points and the arithmetic of every future contract changes. The code doesn't care about the narrative. It cares about the marginal cost of the next dollar.
Which is why the escape hatches matter more than the headline. Asset-backed securitization of GPU cash flows. Sale-leaseback. Off-balance-sheet SPVs that hold hardware and lease it back. These are not exotic. They are standard. But understand what they do. They move residual-value risk off the issuer's balance sheet and onto insurers, pension allocators, and private credit funds. Single-name visibility drops. Systemic contagion rises. The risk does not vanish. It is redistributed to holders with less ability to mark it.
The unglamorous question: what does the maturity wall look like? If debt matures across 2028โ2031 in a cluster, the company refinances into the market that exists then, not the one that exists now. Then check the term match. If the average take-or-pay contract runs three years and the debt runs seven, there are four years of financing an asset against a customer who may not renew. Re-contracting risk is the quiet killer in every leasing model. It never appears in EBITDA. It appears at renewal. The code doesn't renegotiate the covenant. The calendar does.
Zoom out. Competition in AI compute is no longer about who engineers best. It is about who borrows cheapest. Microsoft, Google, Amazon, and Oracle fund at investment-grade spreads โ call it 4โ6%. Neoclouds fund in the double digits. Once GPU-hour unit economics converge, and they are converging because the silicon is identical, a two-to-three-times capital cost differential decides the outcome. That is a structural disadvantage. No operational efficiency closes a gap that lives on the liability side of the balance sheet. CoreWeave's differentiation is real: first-mover position, a strategic NVIDIA relationship where the supplier is also investor and customer, software acquisitions for MLOps stickiness, the Core Scientific deal for power and land. None of it changes the cost of capital. It only buys time against it. The gap is not a management failure. It is the price of being small in a market where scale is the only input that lowers your cost of money.
Note where this story was reported. A crypto outlet covering an AI-cloud credit event. That is not incidental. Bitcoin miners pivoted to AI โ Core Scientific, IREN, TeraWulf, Hut 8 โ and welded crypto capital to AI compute assets. If you hold miner equity, or a fund with miner exposure, you hold AI infrastructure credit risk whether you underwrote it or not. The correlation is not theoretical. It is on your statement.
The consensus read is that CoreWeave is the fragile one. Flip it.
The entity that publicly flagged the credit risk is the one with the thinnest cushion. GMI Cloud is smaller, holds a shallower collateral pool, and pays more to borrow โ probably more than 15%, without the scale or the contracted backlog to negotiate it down. In a credit contraction, the first failures are not the largest leveraged players. They are the ones with no access to the capital markets at all. The warning-sender is often the first casualty. It simply arrives as a warning rather than a filing. CoreWeave may well survive. The neocloud tier beneath it may not.
And the real industry question is not whether one company survives. It is whether the marginal financing cost curve for AI compute has shifted up permanently. If it has, the ceiling on 2026โ2027 capex growth is already set, and no demand forecast revises it upward. If compute runs into temporary surplus, the profit does not disappear โ it moves. Asset holders absorb the depreciation; the application layer and model startups get cheaper inference. That is redistribution inside the stack, not the end of the trade. Most of the people arguing about 15% are arguing about the wrong actor.
The next signal will not come from a CEO quote. It will come from a rating action on a GPU-backed securitization, or a disclosed LTV breach buried in a filing nobody reads. Watch the covenant language. Watch the maturity wall. The number everyone argued about was the loudest one; the six-year schedule underneath it was the one that will actually be tested โ and when it is, the market will not ask whether the yield was insane. It will ask what the collateral was worth.