Ed Zitron 长文剖析 AI 行业依赖垃圾债融资的信用风险

Ladies and gentlemen, we are in a glorious, splendiferous, outrageous era of crap! 

Some call it “high-yield,” some call it “speculative-grade,” but to those of us living in the real world, we know them as junk bonds, the formerly-sleepy world of below-investment grade debt that’s exploded in the last six years. 

Since the beginning of 2011, junk bond issuance has increased by seventy-five times, from a mere $3 billion issued against $60.8 billion in investment-grade bonds in the first half of 2011 (around 5%) to an astonishing $229 billion of junk against the $805.9 billion in the first half of 2026 — making the high-yield market about 28.4% the size of the investment-grade market.

The high-yield market has ballooned over the last few years thanks to the vagaries of the global economy and, in recent years, the arrival of an entirely new category of crap. 

And “junk,” in the case of today’s newsletter, is going to extend so much further than simply “high-yield,” and into the murky world of leveraged loans that have been used to fund dodgy software deals and an increasing amount of data center debt deals.

So why should you give a shit about junk? 

The AI Industry Loves Junk Because Its Businesses Are Crap

As I covered in this week’s free newsletter, Anthropic and OpenAI are two of the single-worst businesses in the history of capitalism, with negative EBITDA cashflows and massive customer concentrations, making them likely to be rated toward the lowest tiers of junk — CCC, or the equivalent rating in Moody’s system, Caa:

A CCC credit rating indicates that an entity is currently vulnerable and dependent on favorable business, financial, and economic conditions to meet its financial commitments. This rating suggests a significant risk of default, meaning that the issuer's ability to fulfill its obligations is highly uncertain and contingent on external factors.

While you probably know this, “default” refers to when a company has missed a payment on its debt and the credit agencies don’t believe that it will ever make that payment. 

The AI labs will require $50 billion to $100 billion of annual issuance across the worlds of high-yield, private credit, leveraged loans (high-interest loans with floating rates issued to companies with a ton of existing debt), and revolving credit lines that, according to a source in fixed-income, are the kind of thing you only use as a last resort.

They’d also be the first parts of the AI industry to truly touch the world of junk other than the disgraceful AI neoclouds, companies that raise oodles of debt with the vague promise of building AI data centers some time in an indeterminate future, but serve the more immediate benefit of pumping NVIDIA’s revenue numbers by buying GPUs years before they’ll ever dance with the power grid. 

Regardless of how shitty it is, this industry needs hundreds of billions of dollars to fund the theoretical buildout of these potentially-possible data centers — CoreWeave alone, per UBS, needs to raise $102 billion in debt through 2030 — at a rate with no historical comparison. Even the railroad bubble of the 1800s had, based on rough calculations of inflation, only around $250 billion in total investment compared to the $400 billion in issuance expected from hyperscalers alone in 2027.

Yet the vast majority of companies building AI data centers range from low credit to no credit, with whatever creditworthiness they may have based near-entirely on their proximity to either NVIDIA or one of the hyperscalers that is invariably going to be their customer.

The problem, you see, isn’t really the AI of it all, but the uniquely stinky business of being in AI. Even under the best possible circumstances, an AI data center will need billions of dollars up front, years in advance of any potential return on investment, buoyed only by the hype around the potential payoff at the other end. 

All of this exists in a time in history when more and more companies are demanding more and more debt from the markets — including the large, wet son of Larry Ellison and his $52 billion bond/loan sale to fund their takeover of Warner Brothers Discovery, $12.4 billion of which was funded with junk due to the sheer scale of the buyout, with Paramount/Skydance/Whatever paying as much as 9.1% on the 10-year end of the deal. 

And as I’ve discussed in the past, Larry Ellison’s Oracle now sits on the very last rung of investment-grade debt, with any further downgrades dropping it into the junk markets and leading to investment and pension funds legally having to dump its debt en masse, which would be a big problem considering the $30 billion or more of personal loans Ellison has collateralized using Oracle stock will face brutal margin calls in the event the stock dumps…which it will absolutely do in the event of a downgrade.

Its last downgrade came from the most obvious of places, per S&P Global:

OpenAI remains a key credit risk. We estimate that OpenAI makes up roughly half of the $638 billion in RPO. OpenAI’s ability to meet its contractual obligations and raise external financing will be contingent upon AI tailwinds continuing and its models being market leaders. If OpenAI were unable to pay Oracle, we believe Oracle could be left with massive data center leases that it might be unable to exit or have to re-lease to new tenants under less-favorable terms. As a proxy for OpenAI’s future prospects, we’re tracking OpenAI’s financial commitments to data center operators and chip makers to gauge its overall financial exposure and its market share among enterprise and consumers.

Today’s exploration of junk is a culmination of multiple different threads of premium research, ranging from the AI data center bubble to the SaaSpocalypse, with massive amounts of leverage powering some of the worst deals in history, creating a setup for private credit and other investors to lose billions of dollars as a result.

Welcome to the Hater’s Guide To Junk, or Enter The Crapverse.