Leveraged Finance · AI Build-Out

AI's $32 Billion IOU: The Junk-Bond Table That Should Keep You Up At Night

Filed under: things our model squawked at before we'd had our coffee.

By Santro AI · June 2026 · Source: @zerohedge on X

Every morning our model listens to the wire like a trader with one ear on the squawk box. Most of it is noise. Then, every so often, something cuts through — and this week it was a single screenshot doing the rounds on X.

It came from ZeroHedge which, in our book, is about the sharpest account in the game when it comes to catching the data that actually matters, then describing it and tearing it apart. As traders, we'll say it plainly: we don't read anyone better for this. That's not a throwaway compliment — it's why our model is wired directly into ZeroHedge's feed, and why, if they ever open up an API, we'll be first in line to plug into it. When they squawk, we listen.

So when this table popped up, the model flagged it as an interesting new squawk. We pulled it apart. Here's what we found.

It's fine. I'm fine.
Everything's fine.— the bond market, May 2026

APLD9.25% CORZ7.75% SECMOS8.875% EDGCOM7.5%

First, the scary part

The table — originally Morgan Stanley Research data, surfaced by ZeroHedge — tracks data-center construction deals living in the public high-yield bond market. Translation for anyone still early in the journey: these are IOUs from companies building AI data centers, and they're rated below investment grade. The street's polite word is "high yield." The honest word is junk.

Here's the number that made the model perk up: at the end of 2025 there were 4 of these deals. Today there are 12. Combined, that's roughly $32.4 billion borrowed, funding about 3 gigawatts of computing power. The AI build-out is increasingly being paid for not with cash, but with borrowed, higher-risk money.

That alone looks scary. A whole industry leaning on the junk market in under six months. But here's the thing. The headline is scary. What's inside the table is scarier. Because once you crack open each row, you start seeing who's really on the hook — and how thin some of these promises are.

Crack open any deal

Tap a ticker to decode every column — same data, made human.

Coupons compared

Higher coupon = lenders demanded more to compensate for risk. Notice the backstop isn't the only thing that moves it.

How to read one of these

A bond name like WULF 7.750 '30 reads as three facts: WULF is the borrower, it promised 7.75% interest a year, and it has to repay the loan in 2030. Then each row tells you how much was borrowed, what the three rating agencies (S&P / Moody's / Fitch) think, how much power it builds (in megawatts — in AI, power is the product), who rents that power, and whether someone bigger is backstopping the rent. That last bit is the whole ballgame. Keep your eye on it.

The 12, ticker by ticker

Where these sit on the credit ladder

Everything above the dashed line is "investment grade." Every deal here lives just below it, in the top tier of high yield.

Now the part nobody tweets

We'd be bad analysts — and worse traders — if we just yelled "junk!" and walked off. So, honestly: most of this table is upper-tier high yield (BB / Ba). Nobody here is deep-junk CCC garbage. Several deals (BLKPRL, MERIDI, WULF, CIFR, FLASHC) have investment-grade giants — Amazon, Google — standing behind the rent, which changes the risk picture completely. And a chunk of this capacity is already live and earning (CORZ, EDGCOM, APLD), not vaporware.

The trap is the lazy story "backstop = safe, no backstop = doom." The data says otherwise. TRACTC has no backstop and the cheapest coupon — because NVIDIA and a clean rating carry it. SECMOS has SoftBank support and a near-9% coupon — because it's tiny and early. The coupon reflects the whole picture: tenant, rating, how built-out it already is, and the backstop — never one column alone. That's the discipline. Don't anchor on the scary headline or the comforting one. Read the row.

Bottom line

The scary headline — 4 deals to 12, $32bn of AI built on junk — is real. But the inside is a split screen: a handful of genuinely fragile, naked-credit, high-coupon names (looking at you, APLD and SECMOS) sitting right next to deals quietly guaranteed by some of the most creditworthy companies on earth. The risk isn't evenly spread. It's hiding in specific rows. Your job is to know which.

Not investment advice. We're traders sharing how we read a table, not telling you what to buy. All ratings, coupons and figures are as reported in the source table (bonds as of 31 May 2026) and may change.

← All posts