A rival CEO just called CoreWeave's 15% bond yield "insane." He's half right. It isn't insane. It's the credit market finally doing the math that equity markets have spent two years avoiding.
OpenAI GMI Cloud's chief looked at a double-digit yield on a listed AI cloud and called it detached from reality. But yields don't come from nowhere. They come from risk. And the risk sitting under every neocloud balance sheet — the GPU depreciation curve — has never been this expensive to ignore.
Here's the number that matters. If that 15% figure is real — and I want to be blunt, it has not been cross-verified against issuance documents — then CoreWeave's cost of new debt sits roughly 2.5 to 3 times above what Microsoft, Google, or Amazon pay to borrow. In a business where compute is turning into a commodity, a 3x capital cost gap isn't a headwind. It's a verdict.
Context: How a Rental Business Got This Fragile
Let me set the table for anyone living under a rock. Or a rock with better yields.
CoreWeave listed in March 2025. It's the poster child of the "neocloud" category — companies built to rent out GPU capacity, usually NVIDIA hardware, at scale. Unlike AWS or Azure, they don't have a search engine or an ad business quietly subsidizing the capex. They borrow money, buy GPUs, and lease those GPUs on multi-year contracts. That's the entire model.
The mechanics are simple and brutal. You take on debt. You buy an asset that loses value fast. You rent it out. Your profit is the spread between rental income and the combination of interest cost plus depreciation. Everything hinges on three variables: how much you can charge, how long the hardware stays valuable, and how cheaply you can borrow.
For two years, all three cooperated. AI demand exploded. H100s were gold. Money was everywhere. That party is ending, and a bear market has a way of stripping the paint off a business model.
GMI Cloud sits inside NVIDIA's Cloud Partner program. It's smaller, it's regional — US and Asia-Pacific, including Taiwan — and it leans hard into sovereign AI and localized compute. Its CEO looked at CoreWeave's borrowing cost and publicly called it detached from reality. That comment is the news. But the comment is also a data point about the commenter, and I'll get to that. First, the technical guts, because that's where the real story lives.
Core: Depreciation Is the Whole Ballgame
Start with depreciation, because it's the assumption everything else rests on.
CoreWeave depreciates its GPUs over roughly six years. The industry norm runs four to six. NVIDIA's architecture cadence — A100, H100, H200, B200/GB200, then Rubin — turns over every 12 to 18 months. Read those two sentences together. You're depreciating an asset on a six-year schedule while its replacement ships every year and a half. That's not a rounding error. That's a structural mismatch, and the credit market is now pricing it.
The rental price is already telling you the depreciation assumption is too optimistic. Public market data shows H100 rental rates have fallen hard from their 2023 peak. Some channels have slid from the $8 per GPU-hour range down toward $2 to $3. That's not demand disappearing. That's demand rotating. New capacity flows to the newest silicon. Old silicon keeps working — it just earns less. Cash flows decay faster than a straight-line depreciation curve assumes.
This is where most bull narratives quietly cheat. They say "AI compute demand is growing, therefore GPU owners are fine." Wrong. Demand growth and asset value decay coexist all the time. Total compute demand can rise while the unit economics of last-generation hardware collapse. Incremental demand flows to B200s and GB200s, not to the H100s sitting on a 2024-vintage balance sheet.
When I model this out, the sensitivity is violent. Shorten the depreciation window from six years to three or four — which is arguably more honest given the cadence — and CoreWeave's EBITDA and leverage ratios get repriced in a hurry. That repricing is very likely the true source of bond investors' anxiety. It isn't that they think management is incompetent. It's that they can't verify the residual value of the collateral.
And collateral is exactly what this debt rides on. Lenders are secured against GPUs. But GPUs are a terrible collateral class in one specific way: there's no mature secondary market, no standardized valuation benchmark, and — critically — no single global price. Export controls fragment the market geographically. An H100 inside China is worth something completely different from an H100 inside the United States. You can't liquidate a global GPU pool at a global price, because there is no global price. There are two or three regional prices, separated by law.
That's why a 15% yield may be rational, not crazy. It's not a verdict on CoreWeave's operators. It's a price on the uncertainty of the residual-value curve of the collateral. The market is saying: we don't know what these GPUs are worth in three years, so we're charging you for that ignorance. When you frame it that way, the yield stops looking like a panic and starts looking like a hedge.
Now layer the business model on top.
CoreWeave runs a heavy-asset, long-contract cloud. Revenue is concentrated in take-or-pay agreements — Microsoft, OpenAI, NVIDIA, Mistral, among others. In 2024, Microsoft alone reportedly accounted for more than 60% of revenue. That concentration has come down into 2025, but it's still the defining feature. Long contracts are the valuation base. They're also counterparty credit risk in a trench coat. If your biggest customer can renegotiate or walk, your "locked-in" revenue isn't locked in at all.
Capex intensity makes it worse. 2025 capex guidance sits in the tens of billions — far above revenue. Free cash flow is deeply negative. A company with negative free cash flow and a heavy capex program must keep returning to capital markets. That's the fatal intersection with a high cost of debt. The damage from a 15% yield isn't mainly the interest on existing debt, which is largely already priced on fixed or floating loans. The damage is the marginal cost of new debt and refinancing. And this company is a frequent flyer in the capital markets.
So what does a neocloud do when its cost of capital spikes? It gets creative. Watch for asset-backed securitization of GPU cash flows, sale-leaseback deals, and off-balance-sheet SPVs. These structures don't make the risk vanish. They move it — onto insurers, pension funds, and private credit funds. The visible risk on one company's books goes down. The systemic contagion risk goes up. That's the trade nobody advertises on the earnings call.
Here's the question that should keep CoreWeave bondholders up at night: what does the maturity wall look like? If a cluster of debt matures in 2028 to 2031, the company faces a refinancing cliff exactly when its oldest GPUs are worth the least. And does the average remaining life of its contracts match the life of its debt? If contracts run three years and debt runs seven, you've built a re-contracting risk into the core of the capital structure. Sign a new customer at 2028 spot prices — which, if rental decay continues, will be lower — and the spread that justified the borrowing evaporates.
Zoom out. This single quote is a micro-signal inside a macro shift. AI compute financing is moving from equity-led to debt-led, and debt markets price asset quality earlier and more coldly than equity markets ever do. In 2024 and 2025, AI data center debt and structured financing ballooned — hyperscaler SPV joint ventures, AI data center ABS, GPU-backed loans. The borrowers include CoreWeave, Oracle, Meta-linked JVs, xAI, Vantage. The scale is now systemic. AI credit spreads can transmit to the broader fixed-income market.
The buyers of that debt are private credit funds, insurers, pensions, ABS investors. Widen AI credit spreads by 200 to 300 basis points and the shock travels through all of them, then feeds back to suppress AI capex appetite. That's the transmission chain, and it runs straight through the exact institutions that spent the last decade hunting yield.
Then there's the crypto crossover, and this is where it gets personal for my readers. This story was reported by a crypto outlet — Crypto Briefing. That's not incidental. Bitcoin miners pivoting to AI — Core Scientific, IREN, TeraWulf, Hut 8 — have welded crypto capital to AI compute assets. If you hold miner equity, if you hold the tokens of projects financing compute buildouts, you are already exposed to AI infrastructure credit risk whether you logged that exposure or not.
Rank the damage. High impact: pure compute rental neoclouds, second-tier data center REITs, new entrants dependent on project finance. Medium impact: the AI application layer — which actually benefits if compute oversupply cuts inference costs — and power and energy suppliers facing downgraded demand expectations. Low impact: hyperscalers with cash flow covering their own capex, and top-tier model labs.
The competitive picture is now a capital-cost war, not a technology war. Hyperscalers borrow at investment-grade rates, 4 to 6%. Neoclouds borrow in the double digits. When compute unit economics converge, a 2x to 3x capital cost gap decides the long-run winner. CoreWeave's 15% yield, if real, is the quantified size of that gap. You can't out-operate a structural cost disadvantage of that magnitude. You can only out-grow it, and growth is exactly what's getting more expensive.
CoreWeave is fighting back with real moves — first-mover advantage, a strategic NVIDIA relationship where NVIDIA is simultaneously supplier, investor, and customer, software acquisitions like Weights & Biases and OpenPipe to build MLOps stickiness, and a Core Scientific acquisition to grab power and land. Good moves. But customer concentration is a double-edged blade. Microsoft and OpenAI long contracts secure revenue while handing pricing power to the customer. When your revenue depends on a handful of giants, the giants set the terms.
Contrarian: Audit the Messenger
Now the part the headlines skipped.
GMI Cloud's CEO is a competitor. He runs a smaller AI cloud inside the NVIDIA partner ecosystem, on a regional and sovereign-compute strategy. He has an interest in CoreWeave's financing costs rising. If CoreWeave's credit gets more expensive, GMI's relative position improves in the fight for GPU allocations, customers, and capital. This is a competitor publicly commenting on the credit health of a listed rival. That's a sensitive act, and it should be read as one.
And here's the inversion nobody wants to say out loud: GMI Cloud's own cost of capital is very likely higher than CoreWeave's. Smaller scale, thinner collateral pool, lower or nonexistent credit rating. The critic may be the first one the credit cycle reaches. The company warning you about the flood may be standing in the lowest part of the valley. That doesn't make the warning wrong. It makes the messenger worth auditing.
The deeper contrarian point: the real industry risk isn't "does CoreWeave survive." It's "has the marginal financing cost curve for AI compute already shifted up." That curve sets the speed limit for AI capex growth through 2026 and 2027. If compute hits a phase of oversupply, the winners are the application layer and model startups — inference gets cheaper — and the losers are the compute asset holders. This is a redistribution of profit along the chain, not a collapse in total demand. Read it as reallocation, not apocalypse.
And watch the small neoclouds. In a credit contraction, they're more fragile than CoreWeave, because they lack scale, anchor customers, and collateral depth. The most exposed players are often the loudest. We didn't learn that yesterday. We learned it in every credit cycle since forever. Speed isn't the pulse of the market — solvency is.
Takeaway
Exchange leads see the wave before it breaks, and this one is forming offshore, in the fixed-income market, not on the token charts. The next thing to watch isn't a headline about CoreWeave. It's the AI data center ABS pipeline — whether deals price, whether spreads widen, whether insurers and pensions keep buying. If that market seizes, the capex ceiling for 2026 gets set by bond desks, not by product roadmaps. Regulation doesn't change that. Capital does. From chaos to clarity: tracking the summer of 2025 is tracking where the credit stops flowing — because that's exactly where the compute stops growing.