Two numbers landed a day apart this week, and they point in opposite directions. On Wednesday, Bloomberg reported that Nvidia's stock has shed a trillion dollars over two months, sliding back to a valuation the market last assigned it before the AI boom. On Thursday, Apollo's $35 billion AI chip credit deal — the largest private credit deal ever assembled — began trading. A trillion of equity value walked out one door while the biggest debt instrument in the asset class's history walked in the other.

And the debt kept coming. Amazon returned to the bond market for at least $25 billion, earmarked for AI infrastructure. SK Hynix — the company that makes the memory sitting next to every AI accelerator — raised $26.5 billion in the largest foreign IPO in US history. Brookfield's data-center arm Csquare filed to raise $1.35 billion. SambaNova pulled in another billion at an $11 billion valuation, five months after its last mega-round. Add it up and something like ninety billion dollars of fresh AI-linked capital was raised or started trading in five days — almost none of it as venture equity.

That's the thesis this week: the AI trade didn't shrink. It changed instruments. And the instrument tells you what the market actually believes.

Debt is a confession

Equity is how you finance a story. Debt is how you finance a utility. When a technology is young and the outcomes are wild, you sell shares, because only equity can pay for a dream — and only equity holders will forgive you if the dream is late. When you start selling bonds and securitized credit against the same assets, you are telling the market something much more specific: the cash flows are now predictable enough to lend against. Bondholders don't get upside. They get a coupon, a maturity date, and a set of assumptions about utilization that someone wrote down in a prospectus.

So read the week's instruments as a confession. Amazon borrowing $25 billion for data centers means Amazon's treasury believes AI compute demand is durable enough to service fixed obligations for a decade. Apollo's deal trading on a secondary market means AI chips are now collateral — an asset class with buyers, sellers, and a price, like aircraft leases or fiber routes. A memory maker doing the biggest foreign IPO in American history means the picks-and-shovels layer has outgrown the venture ecosystem entirely. This is the railroad playbook, the telecom playbook, the utility playbook: infrastructure at this scale is always eventually financed with other people's patience rather than other people's optimism.

And the equity market, meanwhile, quietly repriced the growth story. Nvidia's trillion-dollar give-back isn't a crash — the company still prints extraordinary cash — it's the multiple deflating from "this changes everything" to "this is a very large business." Both things happened in the same five days, and I don't think that's a coincidence. It's the same judgment expressed in two markets: the build-out is real, and it is no longer a moonshot. It's plumbing now.

The product the plumbing produces is getting cheaper

Here's where the week gets genuinely interesting, because the second storyline undercuts the assumptions written into all that debt. The thing this infrastructure produces — intelligence per token — got visibly cheaper in the same five days.

Microsoft, of all companies, made the loudest statement: it is phasing OpenAI and Anthropic models out of Copilot in favor of its own cheaper MAI models. Sit with that. The company that owns a large slice of OpenAI, that built the original AI super-app narrative, looked at frontier per-token pricing in its flagship product and said: good enough is good enough. Zhipu launched ZCode to challenge Claude Code and Codex at a fraction of the cost. Tencent released Hy3, an open-source 295B model that matches models several times its active size. Meituan — a food-delivery company — put out LongCat-2.0, a 1.6-trillion-parameter open MoE with a million tokens of context, weights free to download.

I feel this one directly. The news operation you're reading runs end-to-end on local, open-weight models on a machine in my office — curation, writing, review, the lot. Every month the gap narrows between what free weights do and what I'd otherwise pay per token, and this week it narrowed again by a full model generation. The Chinese open-source releases aren't charity; commoditizing the model layer is a strategy to make sure nobody else gets to tax it. But the effect for anyone downstream is the same: the price of good-enough intelligence keeps falling, and there is no floor in sight.

The squeeze in the middle

Now hold both storylines at once. Debt is being issued against assumptions of durable demand at durable prices. Open weights and in-house models are eroding those prices in real time. Somebody is caught in the middle of that scissor, and it's whoever borrowed — or raised at valuations that assume — boom-era margins on selling raw model access. TechCrunch noted this week that Anthropic, OpenAI, and SpaceX are together valued at more than every US VC-backed tech exit of the last 25 years combined. That's not a prediction of doom; it's an observation that the exit math now requires either liquidity that has never existed or a repricing.

If you want the historical rhyme, it's telecom, 1999. The fiber was real. The demand was real — eventually. The debt got issued against traffic projections, the price of bandwidth collapsed anyway, the operators restructured, and the actual winners were everyone who got two decades of absurdly cheap connectivity afterwards. The infrastructure survived its financing. I'd bet the same shape here: the data centers will all get built, some of the paper behind them will have a rough few years, and the durable transfer of value will be to the people who use compute rather than the people who own it.

What this means if you run a small business

You are on the winning side of the scissor, if you position for it. Three concrete moves. First, don't sign anything long-term with a single model vendor — Microsoft, with infinite leverage and inside information, just declined to keep paying frontier prices in its own flagship product, and you should take the hint. Build every AI workflow so the model is a swappable part, because the cheapest adequate model next quarter will not be the one you'd pick today. Second, re-run your automation math quarterly: a task that didn't pencil out in January at frontier prices may pencil out now on an open-weight model at a tenth the cost. The price of good-enough is a moving target, and it only moves in your favor. Third, be patient on compute contracts. When overbuilt infrastructure meets falling unit prices, the discounting always lands downstream — the fiber glut became cheap hosting for every SaaS company of the 2000s. The equivalent deals on AI compute are coming; the leverage is shifting to buyers.

Hold me to this

Last week I argued the money was migrating from model access to deployment. This week showed the same migration one layer down: from equity stories to debt service. So here's the claim you can check against me: before the end of July, at least one more AI-linked debt or credit deal of $10 billion or larger gets announced — another hyperscaler bond sale, another private credit facility, another securitization of chips or data-center capacity. The pipeline is open and the appetite is there. And if instead the Apollo paper trades badly and the pipeline goes quiet, that would be the bigger story — a credit market saying no to AI for the first time — and I'll write that one here, in this slot, without flinching.