The Bond Market Is Doing the AI Trade's Math, and CoreWeave Is the Worked Example

The box math at 15%

The numbers from the All-In conversation were stark. CoreWeave's debt yields have reached 15 percent as investors question its cash flows, its credit-default swaps imply a 16 percent annual probability of default, and a specific GPU loan — DDTL 1.0 — "spiked to 15%." The secondary market has discounted the company's unsecured bonds by 20 percent, lifting the implied coupon from 9.6 percent to roughly 15 percent. That is the price the market now charges for the neocloud model: borrow against five-year GPU contracts and hope the rates cooperate. m:…1487

The "box" economics, as the hosts explained them, are genuinely attractive on paper: CoreWeave packages GPUs and contracts into financed units such that "within two and a half years of a five year deal, we have paid for everything." The company has raised more than $35 billion and needs roughly another $35 billion to finance signed contracts — but it has built only about half of the capacity it has contracted for, and fixed-price contracts cannot capture spot-market upside, leaving what the hosts estimated as $20 to $30 billion of revenue on the table. Revenue is concentrated, with three customers accounting for 75 percent, and the structural backstop is an opco-propco arrangement in which individual data center assets stand alone. m:…1487

The credit desk is leading the equity desk

CNBC's Fast Money panel put the equity side in plain terms: CoreWeave has lost a quarter of its value since Monday, 25 percent of every dollar it takes in goes to paying debt, and the SMH semiconductor index just logged its ninth straight day of losses — its longest streak in 25 years. One panelist's verdict: "I think the credit market is currently leading the equity market and rightfully so." Each incremental dollar of revenue growth, another argued, is being "eaten and eroded away by increased cost of capital" — even as CoreWeave's earnings rose 110 percent sequentially last quarter. The growth is real; the financing arithmetic is the problem. m:…1735

The rate sensitivity is the trap. AI infrastructure businesses finance in a variable-rate model — "a spread to the risk-free rate" — which the All-In hosts called "the big problem in these businesses." Their rule of thumb: if the 10-year Treasury reaches 5.5 or 6 percent, business models built on financing spreads face severe distress. That is not a CoreWeave story anymore; it is a statement about every levered AI-infrastructure vehicle priced off the same curve. m:…1487

The whole bond market repriced

Morgan Stanley's Carolyn Campbell, speaking on Thoughts on the Market, showed how the repricing extends beyond the neoclouds. Highest-rated data center bonds now yield close to 7 percent, up from below 5 percent at the start of the year, and her analysis finds that "when borrowing costs approach 8%, issuing these bonds becomes significantly more challenging under these current financing terms." Five-year Treasury rates have climbed more than a percentage point since January. Operators have issued about $15 billion in ABS and $7 billion in CMBS so far this year — and Morgan Stanley expects $120 to $150 billion of data center supply over the next two years. The demand side is not the question: U.S. data center power capacity is projected to grow from 60 to 150 gigawatts by 2028. The question is who can fund it at these rates. m:…9553

Power is the second constraint behaving like a bond market. Campbell warned that availability bottlenecks could leave capacity 30 to 40 percent below projections — which is why she argued co-location facilities offer better value than hyperscale campuses, with shorter leases and pricing flexibility. Mayfield's weekly note made the same point from the supply side: "The bottleneck keeps moving. Power, compute, memory, silicon, and land each take their turn." m:…9553 m:…9590

The spread between the funded and the financed

The contrast inside the numbers is the story. TSMC just reported a 69 percent jump in quarterly net profit and is holding to $60 to $64 billion of 2026 capital spending; Microsoft plans multi-gigawatt deployments of its in-house Maia 300 inference systems starting in 2027; Modal raised $750 million at a $15.75 billion valuation. The funded — hyperscalers and profitable silicon — keep building. The financed — levered vehicles like CoreWeave — are discovering that operational excellence does not hedge the cost of money; the All-In hosts themselves described CoreWeave as excellent at standing up compute capacity. GPUs now last six to eight years and even older H100s rent at premiums, which helps the box math — but only if the debt side stays quiet. m:…9590 m:…1487