Debt-financed compute splits: matched loans hold, the rest reprice first

AI compute is increasingly built with borrowed money, and the question is usually put as whether that debt will be repaid. The evidence of 2026 says it depends on one thing: whether the loan is matched to a contract from a strong customer that runs as long as the loan does. Where it is matched, debt still prices like credit. IREN borrowed at 5.96% in June against its Microsoft contract; Lambda raised $1B at 6.78%, rated Baa1; Meta's Hyperion vehicle was rated A+. Where the loan outlives its contracts, the price has moved every time. CoreWeave's GPU loans went from an A3 rating at SOFR+2.25% in March to Ba2/BB+ at SOFR+5.50% in August, on a facility its own release says runs about five years against customer contracts averaging about three, and its third-quarter guidance has interest of $860–940M against adjusted operating income of $200–260M. IREN's unsecured GPU financing went from 5.96% to 9.00% and 9.51% within three months. Oracle's 2031 notes went from 4.95% in February to 7.45% on 2 November. The claim is that this split widens: contract-matched debt keeps holding, and the borrowers whose loans outlast their contracts reprice first and are the first to default, restructure or be downgraded. It is consistent with MenFem's bearish CoreWeave call and explains why its IREN call can still hold. BULLISH on the thesis, conviction MEDIUM, one year. This view was formed by MenFem's research process at Connor's request and adopted by him. It sits next to axis J1 (long debt-financed compute / short equity-financed) but splits it rather than taking a side, so no axis is set.