The number arrived in a press cycle, not a filing. "Fifteen percent." That is how GMI Cloud's CEO described CoreWeave's bond yield, and then he called it insane.
I have spent three days trying to verify it. I cannot. The coupon, the maturity, the secured-versus-unsecured distinction, the loan-to-value covenants — none of it exists in the public record at the resolution a credit analyst actually needs. What I can verify is stranger than the number itself. Every GPU-collateralized loan trading right now is priced against an asset whose residual value nobody has honestly modeled. The 15% is not a verdict on CoreWeave's competence. It is the bond market admitting it cannot price a depreciating chip inside a jurisdictionally fractured secondary market. That admission is the story. The headline is just the noise wrapped around it.
Context — CoreWeave listed in March 2025 after converting a GPU-rental business into what the industry now calls a neocloud. The model is simple to state and brutal to finance: borrow against GPUs, sign long take-or-pay contracts with hyperscalers and labs, and rent the hardware out at a spread. Microsoft once accounted for more than 60% of revenue. Capital expenditure guidance for 2025 sits in the tens of billions, far above revenue, which means free cash flow is deeply negative and the company must keep returning to capital markets. That last sentence is the whole thesis. A business that must refinance constantly is exquisitely sensitive to its marginal cost of debt — not to the interest on debt already issued, but to the price of the next dollar.
This is where crypto readers should stop scrolling past an AI cloud story. The capital pools have merged. Core Scientific, IREN, TeraWulf, Hut 8 — miners who spent years converting electricity into hashes are now converting it into H100-hours. When a Bitcoin miner pivots to AI hosting, its balance sheet becomes an AI infrastructure credit. When an AI neocloud defaults or reprices, the contagion runs through the same private credit funds, the same insurance books, the same structured products that crypto lenders fed during 2021. The wall between "crypto risk" and "AI risk" is a spreadsheet line, not a firewall.
Core — Start with the depreciation mismatch, because everything downstream flows from it. CoreWeave uses roughly a six-year useful life for its GPUs. Nvidia's architecture cadence runs twelve to eighteen months: A100, H100, H200, B200, GB200, Rubin. The accounting horizon and the silicon horizon do not overlap cleanly. That gap is not an accounting trick; it is the central technical-financial coupling point of the entire neocloud sector.
Now watch the rent. H100 on-demand pricing has fallen hard from its 2023 peak — some channels moved from the eight-dollar-per-GPU-hour range toward the two-to-three-dollar range. That is a cash-flow decay curve steeper than any linear depreciation schedule. The bulls will tell you total AI compute demand keeps growing. They are right. But incremental demand flows to the newest silicon, and older hardware loses unit economics even as aggregate demand rises. Asset depreciation and demand growth are not opposites. They coexist, and conflating them is the single most common error in the neocloud bull case.
This is precisely the terrain I mapped after The DAO, when I spent four weeks reverse-engineering EVM opcode differences with three independent auditors. The lesson then was the same lesson now: the headline "hack" or "collapse" is almost never the mechanism. The mechanism lives in the assumptions the model treats as constants. For CoreWeave's lenders, the constant is the GPU residual curve. If that curve bends — if six years becomes four — EBITDA and leverage ratios get restated, and the 15% stops looking insane and starts looking like the market simply got there first.
Here is the part the coverage missed. Lenders take GPUs as collateral, but GPUs have no mature secondary market, no standardized valuation benchmark, and no fungibility across borders. An H100 inside mainland China is a different asset from an H100 in Virginia, because export controls segment the market. A collateral that cannot be liquidated at a known price in a known jurisdiction is not collateral in the way a bond covenant assumes. It is a promise with a serial number.
So the questions I cannot answer from the outside matter more than the answers I can. Does the debt carry loan-to-value covenants that trigger margin calls when GPU marks fall? If yes, a price decline becomes a liquidity spiral, not a gradual repricing. Is the 15% a coupon or a secondary-market yield-to-maturity? Those are different animals wearing the same number. Is the instrument secured or unsecured? Secured against what, at what haircut, verified by whom? The code didn't fail here — the disclosure did.
Then there is the maturity wall. CoreWeave's refinancing calendar and the average remaining life of its take-or-pay contracts must roughly match, or the company carries re-contracting risk on top of refinancing risk. If contracts run three years and debt runs seven, the spread that looked safe at signing evaporates at renewal. Nobody outside the deal book can see that mismatch, which is exactly why a competitor's public comment lands with such force. It is not analysis. It is a signal fired into a fog.
And the fog has a structure. The likely escape routes for a neocloud under funding pressure are asset-backed securitization, sale-leaseback, and off-balance-sheet SPVs. Every one of those moves risk off one company's books and onto insurance books, pension books, and private credit books. The visible risk to CoreWeave falls while the systemic contagion surface rises. This is how a single company's credit problem becomes a fixed-income market problem, and it is how a crypto-adjacent asset class quietly embeds itself in retirement savings.
Contrarian — Everyone is treating the GMI Cloud CEO as a whistleblower. Read the incentives instead. GMI Cloud is a smaller player inside Nvidia's partner ecosystem, running regional and sovereign-AI compute, with a thinner collateral pool and a lower — or absent — credit rating. Its own funding cost is almost certainly higher than CoreWeave's, not lower. When a competitor with worse access to capital publicly questions a rival's creditworthiness, that is not disinterested forensics. That is a competitor bidding for GPU allocations, customers, and capital in the same auction.
The deeper contrarian read: the person shouting the warning may be the one most exposed. If AI compute financing enters a tightening cycle, the marginal borrower — the small neocloud without long-term contracts or collateral depth — gets squeezed first. Arbitrage isn't risk-free. It's a stress test, and the small players are the ones who fail it. The 15% yield may not be a warning about CoreWeave at all. It may be the first honest price the market has put on anyone's GPU debt, and the loudest objection came from the party that number implicates most.
Takeaway — Stop asking whether CoreWeave survives. Ask whether the marginal financing cost curve for AI compute has permanently shifted up. If it has, the 2026–2027 capex ceiling for the entire sector is already set, and the profit reallocation runs from compute-asset holders to application-layer builders enjoying cheaper inference. Truth is not mined; it is verified on-chain — and the chain here is the credit spread. Watch the AI data-center ABS prints, watch miner hashprice against AI hosting revenue, and watch who refinances next. Code is law, but logic is justice. The logic says the cheap-money era for GPU debt just ended.