Two numbers landed a day apart this week, pointing in opposite directions. Nvidia shed a trillion dollars of market value over two months, back to its pre-boom multiple. And Apollo's $35 billion AI chip credit deal, the largest private credit deal ever assembled, began trading. A trillion of equity walked out one door while the biggest debt instrument in the asset class's history walked in the other.
Add Amazon's $25 billion bond raise for AI infrastructure, SK Hynix's $26.5 billion IPO, the largest foreign listing in US history, Csquare's $1.35 billion filing and another billion for SambaNova, and roughly ninety billion dollars of AI capital was raised or started trading in five days. Almost none of it as venture equity. That is the verdict: the AI trade didn't shrink. It changed instruments, and the instrument tells you what the market believes.
Debt is a confession
Equity finances a story. Debt finances a utility. When you sell bonds and securitized credit against AI assets, you are stating that the cash flows are predictable enough to lend against: a coupon, a maturity date, a utilization assumption in a prospectus. Amazon borrowing $25 billion for data centers means its treasury believes compute demand can service fixed obligations for a decade. Chips trading as collateral means AI hardware is now an asset class like aircraft leases or fiber routes. This is the railroad playbook: infrastructure at this scale is eventually financed with other people's patience, not other people's optimism. Nvidia's give-back is the same judgment from the equity side — not a crash, a multiple deflating from "this changes everything" to "this is a very large business."
The scissor
Here is the problem written into all that paper: the product the plumbing produces got cheaper in the same five days. Microsoft started phasing OpenAI and Anthropic models out of Copilot for its own cheaper MAI models. Zhipu launched ZCode at a fraction of Claude Code's price. Tencent gave away a 295B open model; a food-delivery company shipped a 1.6-trillion-parameter MoE with a million tokens of context, weights free. I feel this directly: this site runs on local open weights, and the gap to what I would otherwise pay per token narrowed again this week by a full model generation.
So debt is issued against durable prices while open weights erode those prices in real time. Whoever borrowed against boom-era API margins is caught in that scissor. Anthropic, OpenAI and SpaceX are together valued at more than every US VC-backed tech exit of the last 25 years combined — not a doom call, just exit math that needs either unprecedented liquidity or a repricing. The rhyme is telecom, 1999: the fiber was real, the debt restructured anyway, and the lasting winners were everyone who got two decades of absurdly cheap connectivity afterwards. The data centers will get built. Some of the paper will have a rough few years. The durable value transfer goes to the people who use compute, not the people who own it.
If you run a business, you are on the winning side of that scissor if you position for it. Don't sign long-term with one model vendor; Microsoft just declined to keep paying frontier prices in its own flagship. Re-run your automation math quarterly, because a task that didn't pencil out in January may pencil out now at a tenth the cost. And be patient on compute contracts: when overbuild meets falling unit prices, the discounts land downstream, the way the fiber glut became cheap hosting for every SaaS company of the 2000s.
The claim to check: before the end of July, at least one more AI-linked debt or credit deal of $10 billion or larger gets announced. If instead the Apollo paper trades badly and the pipeline goes quiet, that is the bigger story — a credit market saying no to AI for the first time — and I'll write that one here.