On one side, chipmakers are posting record profits; on the other, compute operators are borrowing at costs of around 15%. This split in financing is more striking than earnings in the current AI compute boom. Strong results from TSMC and ASML show that upstream companies selling the picks and shovels are indeed making money. But downstream, CoreWeave has $100 billion in orders while its debt costs have surged to around 15%, making it difficult for cash flow to cover investment.

The gap in financing costs comes down to collateral and credit backing. Investment-grade financing costs around 5.9%, while unsecured bonds have to offer yields in the high double digits. Anthropic is also raising money, with yields on its subordinated debt exceeding 10%. It is the same compute story: upstream companies get low-cost funding, while downstream players rely on high-interest debt to finance investment. The spread in between is the market’s price for cash-flow certainty.

Looking ahead, two factors will determine whether compute operators can still turn orders into profits: whether operating cash flow can cover equipment spending, and whether financing costs will continue to rise. Record profits upstream do not mean things are going well downstream; the cost of capital has already spelled out the answer in bond yields.

#AI算力 #融资成本 #chips