Premium: The Hater's Guide To AI Debt (Part 2)

The year is 2026, and you are a hyperscaler CEO. You zip up your Patagonia Vest, type UDPATE CALENDER WHERE WHY to your Muse agent, and it tells you that your CFO has sent you an email about something called a “critical finance meeting,” and you roll your eyes.

You were up until 2AM talking to your 38 GPT-6 agents that were vibe coding a dashboard of “company efficiency wins,” and if anything it’s kind of rude that your CFO is interrupting your “mindfulness hour” where you listen to Andrew Hubermann and do something called an “elevated ab crunch” that hurts your neck every time, somehow.

Behind you, your horribly-trained Shiba Inu (called “Basis Points”) angrily humps your Eames chair, and when you tell it to stop it only seems to hump it harder. Your CFO, who has been waiting for 15 minutes, appears on the video with a grim look in their eyes. “What is this? What is the need for this interruption?” you snap. “You’re ruining my mindfulness! Have you any idea how important my mindfulness is? It’s so early in the day, and I’ve barely had any mindfulness!”

Your CFO stops themselves from saying that it’s 12:15PM, and decides to cut to the chase. “Hey, so, remember our conversation last week?”

You begin to shake uncontrollably. “...no. I. don’t. How d-”

Your CFO interrupts, and seems more stern than usual. “Listen. You wanted us to buy a bunch of GPUs, and we bought a bunch of GPUs. That’s fine. But we’ve had to raise tons of debt to do so, and it turns out that the debt that we’ve raised isn’t enough to build all of them, and they’re taking years more than we expect to-”

You begin shaking again. “You…you made a mistake. You messed up. This is on you.” Basis Points is now staring at the wall and growling at it for some reason.

The CFO frowns. “No, these are entirely separate externalities — the war in Iran, the Fed hiking interest rates, the concerns around AI data center debt, the whole supply chain is screwed, everything’s getting more expensive, and we have a bunch of debt-”

You roll your eyes. “Just make it go viral, I don’t know what to tell you,” you say as you hang up. With your mindfulness hour ruined, your week is effectively washed, so you decide to book a trip to Hawaii to recover.

After all, all that boring shit is someone else’s problem!

…except it really, really isn’t.

In last week’s newsletter (and part one of the Hater’s Guide To AI Debt series), I went into the core issues with the AI bubble’s debt spree, which can be simmered down to a few major points (and these are all helpful links to the specific part of the newsletter for easy reference!):

• That the AI data center buildout needs hundreds of billions of dollars’ worth of debt, and hyperscalers need to have pristine earnings growth in the next two years to avoid it eating any and all profits alive.
• AI data centers are generally funded by project financing (IE: loans that are repaid using customer lease payments), convertible notes (loans that can become company stock), and bonds.
• Trillions of dollars’ worth of data center debt is held off-balance-sheet.
• The more debt it raises, the more expensive the debt becomes, and the more money spent on AI capex, the more expensive that capex becomes.
• That the entire industry is functionally dependent on Anthropic and OpenAI’s compute spend.

Put another way, the first part of The Hater’s Guide To AI Debt was about the form of debt — how it’s raised, how it functions, and where it might be going — and today’s about the function.

The biggest worry I have about the AI bubble right now is that the debt is priced for perfection, yet said debt is being invested in thousands of the most-ambitious, intricate, and fragile infrastructure projects in the world, requiring specialist talent and materials and access to power at a scale unheard of in the history of society.

And in many, many cases, the companies building them don’t even have any experience building AI data centers, securing power, or, in many cases, doing much of anything.

You see, despite the AI bubble being inflated by some of the largest and most powerful companies in the world, the actual AI data center buildout is being handled in a way that borders on the lackadaisical “Move Fast And Break Things” model of Silicon Valley.

While they may have an idea of what they’re building and the money to do so and lots of well-credentialed experts, every data center project is its own unique monster with ever-expanding problems caused by everything from geography to the weather to the mechanical issues you find from condensing the power of an entire city into a space a thousandth of the size full of expensive AI chips that need bespoke cooling.

And it’s this gaggle of choices — thousands of chaotic, problematic and ultra-challenging infrastructure projects — that the finance industry has fed somewhere between $100 billion and $150 billion of debt (without including hyperscalers), and plans to feed hundreds of billions of dollars more, all under the assumption that “everything will be alright” and that these are, functionally-speaking, no different from building a regular building.

In part two, I’m going to talk about the grisly truth of the AI data center buildout — the doom loop of debt, delays and doubt that will compound the costs of getting these things built, and how the whole thing has become so unfathomably expensive that it makes the economics of building an AI data center — and paying off the underlying debt — near-impossible for the majority of projects.

This is The Hater’s Guide To AI Debt Part 2, or Gross Profit Unlikely.