The claim
The 10-year closed 5.18% on 2026-09-24, the highest since July 2007. The natural story writes itself: the AI buildout is the most capital-hungry project in corporate history, rates are at two-decade highs, so the financing must be getting squeezed and the levered names must be in trouble.
Three separate tests of that story. It fails all three for the companies doing most of the spending, and passes all three for exactly one name, which is not the one being written about.
Test one: is there a refinancing wall?
No, and not remotely.
Nebius has zero borrowings maturing within 24 months. The earliest convertible matures in June 2029, and after an August exchange it is roughly $100M of original principal. The first material maturity is February 2030. No series carries a scheduled investor put, only a change-of-control repurchase right, which is not a date.
The runway is real but it is not a long one. Nebius has nothing for two years and then it arrives quickly: 24% of principal by 2030 and 61% by 2031, because the February 2030 and March 2031 notes are large. Over five years that is more of its stack coming due than CoreWeave's 58%, the opposite of the impression "no near-term maturities" leaves. The two are mirror images: CoreWeave owes early and then flattens, Nebius owes nothing and then faces a cliff. What softens the cliff is that the debt is convertible and deep in the money, carried at $8.5B, redeeming at $10.0B accreted, marked at $20.8B fair value. If the equity holds, those convert instead of being repaid in cash and the wall is never a cash event. That is a genuine mitigant and a conditional one.
IREN has 82.6% of principal maturing beyond FY2031, against $7.6B of cash and restricted cash. Everything due through FY2028 is $447M, 5.8% of the stack.
Oracle has $17.4B maturing over 24 months against $37.1B of cash.
A wall is where a rate move actually bites: you repay old cheap money and replace it at today's price. None of these companies has to do that for years.
Test two: is the existing debt even rate-sensitive?
Mostly no. It is fixed and it is long. Floating-rate share of debt: Meta 0%, Microsoft 0%, Alphabet about 1.4%, Amazon about 3.5%, Oracle about 6% through commercial paper and bank term loans. IREN and Nebius each added one floating facility in mid-2026, both small against their stacks. CoreWeave is about 34%.
Meta's and Amazon's filings both state they are subject to no financial covenants. The coupons are locked out to 2065, 2066, 2075 and 2076, and in Alphabet's case a hundred-year bond maturing in 2126.
A move in the 10-year today does not raise the existing interest cost of Meta, Alphabet, Amazon or Microsoft by a single dollar. And the floating debt that does exist prices off short-term SOFR, not the 10-year.
Interest expense at these companies is rising sharply. Meta's went from $232M to $754M in a year, Alphabet's from $261M to $1,278M, Amazon's from $516M to $1,314M. That is a volume effect, not a rate effect, and the filings say so: Meta attributes it to "higher long-term debt balances."
Test three: does the market trade them as rate-sensitive?
Not really. Correlation of daily returns against the daily change in the 10-year yield, 2026 year-to-date: QQQ -0.32, Oracle -0.18, IREN -0.12, CoreWeave -0.06, Nebius -0.005.
The broad Nasdaq is more rate-sensitive than three of the four names supposedly being punished by rates. Nebius is indistinguishable from zero.
So where did the rate move actually land?
On the next dollar, not the last one. The cleanest proof is the same company selling the same paper twice, in the table below. Same issuer, same credit, same tenor, months apart. Nothing about the company changed. The price of money did.
That matters only because of what comes next: Meta guides to roughly $130-145B of capex in 2026, Amazon spent $96.3B in a single half and expects to spend more, and Microsoft has signed $329.1B of leases that have not yet commenced, up from $92.7B a year earlier.
The biggest exception: CoreWeave
One name fails the tests above in the other direction, and by a wide margin. Oracle, IREN and Nebius each carry a little floating debt; CoreWeave carries a third of its stack that way. If you want the company where the rate story is real, it is not Oracle and it is not Nebius.
A third of CoreWeave's debt floats. Roughly $12.2B of its $35.6B gross principal is SOFR-linked. That is not an estimate: the company's own rate-sensitivity disclosure says that for every 100 basis points, six-month interest expense moves about $61M, which implies a floating balance of about $12.2B and matches a facility-by-facility build-up almost exactly.
It pays what a levered borrower pays. Effective rates by facility, as filed: 15% on its oldest term loan, 12% on the Magnetar loan, 11% on two more. Its senior notes carry 8.5% to 9.75% coupons, against Meta's 4-6%.
Its cost of new money is visibly repricing inside a single year. The spread on successive delayed-draw facilities went SOFR+2.25% in March, to SOFR+4.50% in May, to SOFR+5.50% in August. Same borrower, same year, same instrument type, 325 basis points wider.
And the near-term maturities are real. $7,513M of debt is classified current against $5,524M of unrestricted cash, so current debt exceeds cash on hand. Somewhere between 30% and 42% of total principal falls due within roughly 24 months. A debt-service-coverage covenant of 1.35x on one facility begins testing from 2026-09-30.
Two things cut the other way and belong in the same breath. CoreWeave hedges: interest-rate swaps of $4,661M notional, under covenants requiring it to hedge at least 95% of anticipated floating borrowings on specific facilities, which takes economic floating exposure closer to $7.5B than $12.2B. And it holds $103.7B of contracted remaining performance obligations against all of this.
The oddity worth sitting with: CoreWeave has by far the most rate-exposed balance sheet of the five and the second-lowest correlation to the 10-year, at -0.06. Whatever is moving that stock day to day, the market is not trading it as a rates instrument.
Three things that surprised me
Oracle is not the outlier borrower. Over 2025-26 Oracle issued $43B of senior notes. Amazon issued roughly $105B equivalent across five currencies, Alphabet took face debt from $12B to about $126B, Meta issued $55B. Amazon now carries more gross debt than Oracle. Oracle is distinctive for being deeply net debt, about -$88B while the others still hold securities cushions, and for being the only megacap leaning on commercial paper and bank term loans.
Amazon flipped from +$54.2B net cash to -$9.6B net debt in two quarters.
Microsoft has issued no bonds in two fiscal years, repaid $3.0B, and moved its obligations into leases instead: finance-lease liabilities went from $14.9B in FY2022 to $66.6B in FY2026, and its finance-lease interest of $2.5B now exceeds the cash interest on all of its bonds at $1.5B.
What would change my mind
Any of the megacaps issuing floating-rate or short-dated paper at scale, which is precisely what separates them from CoreWeave today.
A convertible-heavy name whose stock falls far enough that conversion stops being the exit and cash repayment becomes real. Nebius's converts are carried at $8.5B but redeem at $10.0B accreted and are marked at $20.8B fair value precisely because conversion is deep in the money.
The capex commitments themselves being funded with debt that reprices. IREN has $13.6B of committed capex due within 12 months against $7.6B of cash, and its newest money, an August 2026 facility, is fixed at 9.00%, by far the most expensive in its stack.