Hey Reader,

CoreWeave isn’t the flashiest AI company, and unless you read the business pages, you may not have heard of it. But this month, Wall Street started betting against it.

As we dig into that, we’ll also answer a question that the AI hype crowd doesn’t like to talk about, because it’s complicated and uncomfortable:

How, exactly, are we paying for all the data centers?

CoreWeave borrows money to buy AI chips. Lenders now consider that a riskier bet.

CoreWeave is an important piece of the AI puzzle because it’s what’s called a “neocloud”: a newer company building data centers for AI training and inference as fast as it can. The “neo” sets it apart from established companies like Google, Microsoft and Meta, which are known as “hyperscalers.”

(I don’t know why they name everything like it’s right out of The Matrix. 🙄)

Investors can buy insurance against a company defaulting, called a credit default swap (CDS). This month, the cost of insuring CoreWeave hit a record high, implying roughly a 60% chance CoreWeave defaults on its debt within five years.

Nobody charges that much to insure a company they think is safe. As a result of the souring of the market for CoreWeave’s bonds, the stock is down about 70% from its all-time high.

Because CoreWeave is a young company without a pile of cash, it finances its buildout mostly with debt. Lenders give CoreWeave a loan to buy GPUs (graphics processing units, the chips that run modern AI) and build data centers. It then pays the loan back over time with the money it makes renting out that computing power.

Lately, lenders have gotten nervous. Basically, they’re asking:

“What if investment in all these data centers happens in 2027, but revenue from increased AI adoption doesn’t come for decades?”

In that scenario, it might be hard for CoreWeave to pay its bonds in 2031, even if the company is ultimately making a prescient bet. The challenge is that their bet needs to pay off relatively soon, and even if they turn out to be right about AI demand, that demand might not come fast enough for their current financing situation.

The average bond investor has stopped believing CoreWeave’s story and thinks it’s more likely than not that the timing of AI growth will work against the company – they’ll build too much, too soon. Few investors want to buy CoreWeave bonds anymore, so the company has to find other ways to raise cash.

Again, this doesn’t require anything to go “wrong” with AI adoption. It just requires AI adoption to happen more slowly than CoreWeave expects.

The 4 ways to pay for a data center

Think of it like real estate (except your house is full of lots of advanced computers).

1. Pay cash. Until recently, Google, Microsoft and Meta bought data centers with cash from the profits of their existing businesses. That’s changing, and they are leaning more on debt now. FactSet calculates that new debt went from 9% of hyperscaler spending in 2024 to 32% by mid-2026.

2. Take out a cheap mortgage. Giants like Google and Microsoft now borrow at enormous scale. The IMF says hyperscalers went from less than 1% of global investment-grade bond issuance in 2024 to 10% this year. Lenders give them cheap loans because they are giant, reliably profitable businesses.

3. Let someone else own the building, then rent it. Meta’s Hyperion campus in Louisiana is owned 80% by the investment firm Blue Owl and 20% by Meta. (Again with the Matrix-y names!) That joint company borrowed $27.3 billion at relatively low rates, and Meta starts paying rent in 2029. Meta becomes the tenant in its own data center, and most of the debt stays off its balance sheet.

4. Borrow against the furniture and the lease. This is the neocloud model. CoreWeave buys Nvidia GPUs, rents them to “hyperscalers”, and borrows against the chips plus those signed contracts. When Meta’s contracts back the loan, lenders treat it almost like Meta debt: one $8.5 billion CoreWeave loan earned an investment-grade A3 rating from Moody’s. Add it all up and CoreWeave carries about $35 billion of debt plus $16 billion in lease commitments, against a $104 billion contracted backlog. On one hand, this is great – money is coming in the future! The question is just whether you believe the future money will arrive in time to pay the debt.

Are Anthropic and OpenAI “subprime” borrowers?

Business debt isn’t necessarily bad. But any time you see a company with a lot of complex debt, there are echoes of the 2008 financial crisis.

Jefferies strategist Chris Wood made that comparison this month. By his estimate, about half of the roughly $2.3 trillion in contracted backlog at Microsoft, Alphabet, Amazon and Oracle is tied to two companies, OpenAI and Anthropic, and neither one is profitable yet. (Anthropic is losing $13 billion a month.) It’s a mountain of debt that depends on two borrowers paying later.

But in 2008, the losses landed on banks that leveraged credit default swaps, which is why the whole system froze. Today, a lot of the riskiest AI debt sits with private credit funds, insurers and pension funds (bond giant PIMCO anchored the Hyperion deal, for example). The IMF’s stress tests found big banks have enough capital to absorb an AI slowdown, even as it warns a sharp drop in AI investment could cut U.S. economic output by 1.8% by 2028.

If data center debt gets more expensive, prices could rise downstream.

If the landlord’s mortgage gets more expensive, the rent goes up. According to Goldman, CoreWeave raised its prices 25% in July and another 10% in October. If your business plan assumes AI gets cheaper every quarter, build a version where it doesn’t, and stay platform-agnostic so you can easily switch providers as prices fluctuate. (We have lots more about cheaper open models coming soon.)

And if you want to know whether the AI boom is healthy, watch the bond market, not the stock market. Stock investors buy stories. Bond investors want their money back.

Until next time,

Rob Howard
Founder, Innovating with AI

👋 Thanks for being a part of Innovating with AI. You’re in good company – some of our 175,000 readers work at Google, Apple, Microsoft and IBM… plus Zoom, Deloitte, and Meta.

❤️ Did you love (or hate) today’s newsletter? Reply to this email to tell me. I personally read every single message (without AI).

📬 Share this post (no paywall). If a friend forwarded you this email, subscribe here.