One Week in the AI Money Machine: $25B in Bonds, a Record IPO, and a Moratorium

The AI buildout stopped being a cash-flow story this month and became a capital-markets story. In the span of two weeks: Amazon moved to raise $25 billion in bonds for AI infrastructure, SK Hynix raised $26.5 billion in the largest US listing ever by a foreign company, Meta committed to doubling its compute to 14 gigawatts, and New York became the first state to freeze large data center construction.

Any one of these is a big story. Together they describe an industry that has outgrown its own balance sheets at the exact moment the political pushback arrived.

The scoreboard

When cash-rich companies borrow

Amazon generates enormous operating cash flow and is still going to the bond market for $25 billion. That's not distress; at investment-grade rates, debt is cheap relative to the option value of compute. But it marks a regime change. The first phase of the AI buildout was funded out of profits, quietly, inside existing capex lines. This phase is externally financed and publicly priced, which means the bond market now gets a vote on whether the buildout continues.

That vote has consequences. Bond investors ask ruder questions than earnings-call analysts: what's the useful life of a GPU, what's the revenue per gigawatt, what happens to collateral value when the next chip generation lands. The moment AI infrastructure is priced as credit, AI expectations get marked to market continuously.

SK Hynix cashed the check

The demand side of that trade showed up on Nasdaq. SK Hynix, the dominant supplier of the high-bandwidth memory that sits next to every serious AI accelerator, listed ADRs at $149 and raised $26.5 billion, with the offering oversubscribed seven times. Largest foreign US listing in history, from a company that was a commodity memory maker five years ago.

The proceeds go to fabs and ASML machines, which is the tell: even the suppliers are now raising outside capital to keep up with hyperscaler demand. Everyone in the chain is leveraging into the same trade.

Meanwhile Meta is trying to partially exit that chain by putting its own Iris chip into production in September. Owning silicon is what doubling to 14 gigawatts looks like when you've done the per-unit math on someone else's margins.

And then New York said no

The same week the money was flowing, New York froze permits for data centers over 50 MW. One state, one year, plenty of exemptions. Marginal in capacity terms, not marginal in signal terms: the buildout now has organized, state-level political opposition driven by electricity prices, and other states are drafting copies.

The industry has been modeling chip supply, power availability, and interconnect queues as its constraints. Add permitting risk to the list.

The part that reaches your invoice

Capex cycles are abstract until they show up in pricing, and there are three mechanisms by which they do.

Long-context pricing tiers. Serving a very long prompt is disproportionately expensive in memory, and that cost is increasingly passed through explicitly rather than averaged into the headline rate. Every xAI Grok model doubles its rate past 200,000 tokens. Expect more of this, not less: it is the cleanest way to charge for the resource that is actually scarce.

Caching as the discount that matters. Prompt caching prices cached input at roughly a tenth of the standard rate, and it exists because serving a repeated prefix costs the provider far less than serving a new one. That gap is a direct reflection of memory bandwidth economics. A workload with a large stable prefix and a high hit rate is the cheapest thing you can run, and that is unlikely to change.

Batch as spare capacity. Batch APIs are half price because they let providers fill troughs in demand. That discount tracks how tight capacity is: it is the first thing to narrow when GPUs are scarce.

What to do about it. None of this argues for building your own infrastructure, which is almost always the wrong call for a software team. It argues for three cheap habits: structure prompts so the stable part comes first and actually caches, move anything latency-insensitive to the batch endpoint, and check your provider's pricing page on a schedule rather than assuming last quarter's rates. Our cost calculator carries a verified date on every figure for exactly that reason.

What this means if you just build software

You're downstream of all of it, in three concrete ways:

1. Token prices keep falling anyway. Competition is doing its job: GPT-5.6's Terra tier landed at half the price of its predecessor in the same news cycle. The buildout is brutal for balance sheets and great for API customers, and vendors will keep using price as a weapon while they fight for share.
2. Caching and efficiency features are being pushed hard because vendors need to bend their own cost curves, not just yours. Learn them; they're where the real discounts live now.
3. Watch the credit market, not the demos. If AI progress disappoints from here, the first visible crack won't be in a benchmark. It'll be in the spread on somebody's data center bonds.

The 2023 era of the AI boom ran on wonder. The 2026 era runs on underwriting. It's less fun to watch and much easier to measure, and the measuring has now started in earnest.

The through-line is that the buildout's economics are visible in your API bill if you know where to look. Cached input is cheap because repeated prefixes are cheap to serve; long context is expensive because memory is the constraint; batch is discounted because it fills troughs. Price your workload against those three facts rather than the headline per-token rate, and most of the volatility above stops mattering to you.