Allan Ali — AI Debt Junk-Bond Crossover (video commentary)

Investment-Grade AI Debt Is Pricing Like Junk — and That's the Signal Everyone's Missing

I've been running hosting infrastructure for over a decade, and I've spent the last two years watching the AI buildout learn how to finance itself in real time. It started with cash. Then it moved to investment-grade bonds. Then to private credit. Now it's moved somewhere I genuinely didn't expect to see it this fast — the junk bond market is financing investment-grade AI debt, and blue-chip borrowers are paying prices that middle-tier junk issuers would blush at. If you run any kind of infrastructure business, stop what you're doing and pay attention to this, because this is the chapter where the bill starts getting priced.

Bloomberg put the story out Friday: companies financing data center projects are increasingly turning to junk bond investors to raise billions of dollars — even for debt that is technically investment grade. They called the buyers "tourists." That word is doing a lot of work, and I'll get to why in a minute. The headline number that stopped me cold: BlackRock is paying a 7.53% yield on blue-chip securities sold in July for a Texas data center project. Let me say that again. BlackRock. Blue-chip paper. Seven and a half percent. That is a higher rate than middle-tier junk bonds would often pay — and it's what a top-tier asset manager had to offer the market to get a data center deal done.

The Setup — Junk Bond Tourists Just Crossed Into High-Grade AI Debt

Here's what's actually happening, stripped of the Wall Street language. The AI buildout needs trillions. The traditional buyers of top-grade corporate debt have limits on how much AI paper they'll absorb at reasonable prices, so issuers are going down-market to find the money. And the investors down there — the junk bond crowd, the yield hunters, the "tourists" — are the only ones willing to pay up for what they're being offered. That's the definition of a market that has run out of natural buyers.

This isn't a niche phenomenon. Morgan Stanley estimated roughly $600 billion in AI-related bond issuance globally this year. The AI buildout itself is pegged at around $3 trillion. And the credit market has already started to show the strain — the New York Times reported Thursday that Big Tech's AI borrowing binge is one of the forces driving bond yields to 20-year highs, to the point where the Treasury had to step into the market Wednesday and expand how much of its own debt it can buy just to keep yields from running away. When the government has to intervene in its own bond market because corporations are borrowing too much to build data centers, that's not a footnote. That's a structural event.

The Numbers — the Crossover in Action

Let me lay out the actual deals, because the pattern is unmistakable.

BlackRock's July offering — roughly $12.5 billion tied to the Meta-BlackRock Texas data center complex — priced at 7.53% after weak demand from global investors. Blue-chip asset manager, marquee hyperscaler tenant, and the market still said "pay me like I'm taking real risk."

Galaxy Digital paid a 10% yield on a $3.5 billion debt offering backed by a CoreWeave lease last month just to attract demand. Ten percent. On a deal tied to one of the biggest names in AI compute. That's not financing, that's hazard pay.

This week, Goldman Sachs started sounding out investors on a roughly $1.15 billion junk bond to finance a data center near Richmond, Virginia — developed by American Real Estate Partners' PowerHouse Data Centers and Chirisa Technology Parks, leased to CoreWeave, with equity from Blue Owl Capital funds. The offering is expected to hit the market in September. CoreWeave, by the way, has now supported billions of dollars in junk debt issuance through its long-term leases — but because its own credit rating is speculative grade, its borrowing costs run well above projects backed by the Alphabet-class hyperscalers.

And here's the most telling detail in the whole story: one investor told Bloomberg they passed on a $14 billion offering for an Oracle data center in Michigan — partly because the bonds are callable. Callable means the borrower can refinance at a lower rate once construction is done, and the creditor eats the risk. An institutional investor looked at a $14 billion Oracle deal, the crown jewel of the AI buildout, and said "no thanks, the fine print isn't worth it." That did not happen in 2024. That did not happen last year.

The Two Readings — Discipline Arrives, or the Market Breaks

Now, there are two ways to read all of this, and like most things in this buildout, both of them are true at the same time.

Reading one: discipline is finally arriving. Investors are asking real questions. David Kinsley at Impax Asset Management said it plainly — the questions now are "do you have execution issues? Supply chain or construction delays? Is it a lower-quality tenant?" Borrowers are agreeing to amortization clauses — paying principal back before maturity, which removes risk down the line. More issuers, especially in the junk market, are getting backstops from Alphabet's Google — a covenant guaranteeing a data center lease gets paid even if the tenant defaults. There are provisions capping building costs. That's the market doing its job: saying no to bad terms, demanding protection, pricing risk. If you're a believer in markets, this is the healthy part.

Reading two: the credit signal is breaking. When investment-grade paper prices like junk, it means the market can no longer tell the difference between good risk and bad risk — or worse, it doesn't trust the difference. The last time blue-chip borrowers paid junk-level rates at scale was 2008. I'm not saying this is 2008 — the assets are real, the tenants are real, the power contracts are real. But the pricing mechanism is behaving the way it behaves at the end of a credit cycle, not the beginning. Robert Tipp, head of global bonds at PGIM, called it "a wall of worry piled up before it." That's not the language of a healthy market.

The Part Nobody's Talking About — the Covenant Kitchen Sink

Here's what's actually under-reported in all of this: the market is inventing new financial structures in real time because there is no convention for AI debt yet. Amortization clauses. Hyperscaler backstops. Construction cost ceilings. JPMorgan's John Servidea said the quiet part out loud — "we're seeing what different investors value when it comes to these financings and how they're evaluating risk and return... as supply increases, we expect deal terms and structures to continue to evolve."

Translate that from banker-speak: nobody knows what this paper is worth. There's no 30-year history of data center bonds to price against. There's no default data. There's no cycle to reference. So the market is writing the rulebook while the deals are being sold — which means every structure is being tested by people who have never seen a downturn in this asset class. The protective covenants aren't a sign of confidence. They're a sign that sophisticated money is already pricing in tenant default, construction failure, and refinancing risk on an asset class that has never been through a real credit cycle.

The Secondary Bottleneck — When Investment-Grade Prices Like Junk, Every Signal Distorts

For me, the deeper problem is the pricing signal. In the last year I've written about how the AI market sends contradictory signals — Google stock dropping on a $205 billion capex plan while Supermicro surges on a $60 billion backlog, and every supplier in the chain getting mixed messages about whether to build. This is that same distortion, now in the credit market, and it's worse because credit is the thing that actually funds the buildout.

Think about what a 7.53% blue-chip rate does to every other borrower in the chain. If BlackRock has to pay 7.53% for a Texas data center, what does a regional colocation operator pay for a $20 million expansion? What does a power developer pay to build the substation? What does a smaller cloud provider pay to buy the racks? The AI borrower at the top of the credit stack is bidding up the cost of capital for everyone below them — and every one of those smaller borrowers is calibrating their business plan off a rate that's been inflated by the most aggressive borrower on Earth. When the top of the market prices like junk, the bottom of the market doesn't just get more expensive. It gets distorted. It gets mispriced. It gets dangerous.

What This Means for Independent Hosting Providers

If you run an independent hosting, colo, or infrastructure business, here's what I'd tell you to do with this information.

First — watch AI credit spreads, not AI stock prices. The stock market is a sentiment barometer. The credit market is the actual funding valve. When investment-grade AI debt starts trading at junk levels — which is what 7.53% means — the valve is already constricting. Capacity financing dries up six to nine months before you feel it in hardware availability or colo pricing. The spreads are your early warning system.

Second — lock your own financing now. If you need to refinance, expand, or buy equipment, do it while the market still has appetite for non-AI borrowers. Every dollar of AI debt issued at these yields pushes your borrowing costs up with it. The window for cheap capital is closing, and it's closing because of the buildout, not because of you.

Third — position as the capital-light alternative. The giants are paying 7.5% to 10% to build. Your incremental expansion is a fraction of that cost. That's your competitive edge, and it's widening every quarter. Sell it. A customer fleeing a $200 million hyperscaler bill can't necessarily afford your premium — but a customer who's tired of watching their provider's cost of capital climb will listen.

Fourth — read the covenant structure of any deal you touch. If a data center lease needs Alphabet to backstop the rent, the tenant's credit isn't doing the work — the backstop is. Sophisticated money is hedging tenant default. Before you sign a long-term lease with an AI-adjacent tenant, ask yourself who's really guaranteeing the check. If the answer is "nobody, they're just big," you're now the junk bond tourist.

Fifth — don't confuse this with the end of the buildout. This is not the crash. This is the market starting to price the risk that's been there all along. The buildout continues — but it's going to continue at prices that reflect reality, and that changes the math for everyone who built a business plan on cheap capital.

The Bottom Line

I've been saying for months that the AI buildout's real bottleneck isn't chips, power, or cooling — it's the bill. This week, the bill started getting priced. Investment-grade AI debt at 7.53%. CoreWeave-backed paper at 10%. A $14 billion Oracle deal turned down over a callable clause. Covenant structures being invented on the fly because there's no history to price against.

That's not panic. That's the market doing what markets do when they realize they don't know what something is worth. And in a credit market that has never priced this asset class through a downturn, "we don't know what it's worth" is the most dangerous sentence in finance. The tourists are in the building, folks. Check your exits — and your leases.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Sources: Bloomberg, Reuters, The New York Times, Crypto Briefing, LiveMint, Janus Henderson, PGIM via Bloomberg, Morgan Stanley, Bloomberg Television.