AI Economics for Dummies, Annotated
A riff on another article about AI economics, this time with the receipts attached.
The structure here follows Andrew Singleton's article in McSweeney's called AI Economics for Dummies, containing a set of word problems meant to highlight the business side of AI as companies in the sector start to go public.
Every dollar in this story is riding on one bet: that AI demand is real, durable, and willing to pay more than it costs to serve, and that it holds long enough for a few hundred billion dollars of spending to earn itself back before the hardware turns to scrap. Perhaps it is. After all, the technology is extraordinary, and I'll say so more than once. But there is danger in financing a bet as though it has already won. What follows is a tour of the machinery that makes a wager look like a settled fact.
Before the marbles start moving, keep one question in your pocket. When this settles up, whose hundred marbles were on the table?
I'll keep my hands, feet, and opinions where you can see them and put everything else behind a reference.
Ground Rules
There is one rule in this playground: there are one hundred marbles. Okay, maybe a few more or less. You can keep them, move them, and even tell better stories about what they're worth. But the number of marbles stays the same. Finite. Each technique below is a way of moving the same marbles in a circle and booking the lap as growth.
How to read one of these filings, fast
If you've never opened a 10-K and want to start, the SEC has a short, plain-English guide to reading one. But keep in mind, the filing only shows you the marbles the company was required to show you. The best moves are all about the marbles that never have to appear on the page at all.
1. The grape that cost two and sold for one.
Singleton's version: Alex pays $2 billion for one grape, sells Mike a grape a month for $1 billion, collects twelve months up front, books it as infinite-growth ARR. The business editor moves into his house.
This kind of real-world arrangement is vendor financing, and the cleanest recent AI-first case is Nvidia and OpenAI. In September 2025, Nvidia agreed to invest up to $100 billion in OpenAI. In exchange, OpenAI committed to deploying millions of Nvidia chips in data centers the investment pays for. The hardware company funds the customer, enabling the customer to buy chips. The money is recorded as an "investment" and comes back as "revenue." The marbles moving circularly around the board look like demand.
None of this is hidden. Analysts named it on day one, and Nvidia's own quarterly filing quietly hedged the whole thing. After all, a press release isn't a contract, a point Nvidia's own CFO confirmed two months later, telling investors the deal was still at the letter-of-intent stage with no definitive agreement completed. It does, however, move the stock.
If you want the dot-com rhyme: Cisco did it, too, in the form of lending money to telecoms so they could buy Cisco routers. The sales looked absolutely spectacular until the credit stopped.
2. The crematorium that ran on its investor's money.
Singleton's version: John's propane company invests $20 billion in Jenny's crematorium. Jenny pays John $10 billion for propane to burn. John books $10 billion in revenue and a 5% stake in a "$100 billion business." A Forbes reporter marries them both.
Here, Singleton makes the round trip grape trick explicit. Nvidia has done dozens of these, investing or lending to the very customers who then spend money on its GPUs. AMD did the same in reverse, handing OpenAI a warrant for up to 160 million shares that vests as OpenAI deploys AMD hardware.
The clearest example, though, is Anthropic, as it sits in front of two furnaces at once. The pattern is the same both times: a cloud company invests billions in Anthropic and Anthropic agrees to spend billions right back renting that same company's chips. Amazon committed up to $25 billion in 2026, on top of an earlier $8 billion, while Anthropic committed to spend over $100 billion on Amazon's cloud and chips over the next decade. Google committed up to $40 billion and Anthropic secured access to up to a million of Google's TPUs.
Nothing here is meant to comment on the quality of any product. Just the financing side of it. Each loop is legal and real on paper. But every lap also inflates how much real, outside demand looks like it exists. The furnace is always busy with the same suppliers.
And if you doubt the round trip is the point, watch what actually moved. When AMD and OpenAI unveiled their tie-up, AMD stock jumped while Nvidia's slipped, the marbles steering toward one supplier's instead of the other's. The market priced the reshuffle, not any new demand.
3. The taxi that loses money on every ride. And rusts.
Singleton's version: Laura charges a $20/month subscription, has 40 million subscribers and $13 billion in revenue, and spends $25 billion a year on gas. Also her taxi costs $1 trillion and she has to replace it every 4-8 years.
Two very real things stacked into one cab.
The first is subscription economics underwater: a frontier lab (OpenAI, Anthropic, etc.) sells access for a flat fee while each query costs more to serve than it brings in. All of it runs on infrastructure subsidized by the circular investments above. Revenue is real, it's just no match for the utility bills.
The second is the part taught early on in Economics 1 (trust me to know, I only took ECON1): the shovels rust. GPU depreciation is brutal. This year's fortune in chips is half-priced by next year's hardware. Notably in the Nvidia-OpenAI structure, leasing the chips instead of selling them lets OpenAI avoid the depreciation charge on its own books. Meaning, Nvidia carries the rust and risk of a warehouse of GPUs nobody wants if demand ever blinks.
In other words, you don't own a collection of marbles. You own a standing appointment to keep buying shiny ones as the old ones scratch.
4. The farm that fired everyone for mules.
Singleton's version: Benjamin fires 99 of 100 farmworkers and buys 99 driverless mules expecting a 1000% (10x) productivity gain. The mules cause $50 million in plow damage. His last employee now spends 100% of his time shoveling mule shit. Goldman builds an altar to him.
This is the labor bet, and it's the one I have the most to say about, so I'll keep it short here: the bet is that you can replace living labor with a capital asset that captures the same output and pockets the wage.
The mules help. Often they produce plausible output the remaining worker now spends all day checking. The productivity gain is assumed at purchase and measured, if ever, much later. By then, the altar has already been built.
I'm glossing over who actually captures that surplus in a separate piece. For now, the important takeaway is where it sits in the structure: the justification. Every marble moved in the other entries is borrowed against one major belief: the mule works.
5. The apartments nobody wants, built with someone else's everything.
Singleton's version: Xavier loses $1 billion a month renting apartments, so he commits to building $850 billion more in places nobody wants them and convinces Ted to leverage everything he owns to help. Forbes compares them to God making the heavens and the earth.
Singleton saves this for last. I'd circle it in red. It's where the structural risk actually lives and the road that private equity and private credit walk on.
In October 2025, Meta announced a roughly $30 billion joint venture with Blue Owl Capital to build the Hyperion data center in Louisiana. Through a special purpose vehicle (SPV) arranged by Morgan Stanley, the project issued ~$27 billion in debt plus ~$2.5 billion in equity, anchored by PIMCO (~$18 billion) and BlackRock (~$3 billion). BlueOwl's funds own 80%, Meta keeps 20%. Meta then leases the finished campus back.
That last step is the trick: it turns what would have been capex on Meta's own balance sheet into operating expenses off it. Tldr; a company sitting on a lot of cash chose to build this infrastructure with debt it never has to show. That tells you it would rather the rating agencies not total it up.
Two more details earn their place. First, the deal carries a residual value guarantee. Lenders won't lend cheaply against a building full of chips that might be worthless in five years. So, Meta promised to cover any shortfall in the campus's value at the end. That promise is what earned the debt its A+ rating, and an A+ rating is worth billions in lower interest. This vehicle costs Meta nothing today, only paying out if the assets themselves crater. The accounting now says the debt and the asset aren't Meta's, while the guarantee quietly pins the risk on the hyperscaler.
Second, Google does its own quieter version: by the third quarter of 2025 its filings disclosed $42.6 billion of data center leases not yet commenced and not yet recorded on its balance sheet, up from $23.9 billion a quarter earlier. A Moody's analysis found at least $662 billion in such commitments sitting off the balance sheets of five tech giants, more than their combined total debt.
Here's where the private equity kid steps onto the playground, because this kind of structure is his specialty, not the tech industry's. He doesn't build things, he arranges them. The move is always the same: borrow money, but set it up such that the debt sits on the asset itself, so if all goes wrong, the asset is what fails. Along the way he earns real money in the form of fees for putting the deal together. The debt only comes due later, and by then he plans on changing schools.
The case where the marbles are real
I've played luddite for five entries, so let me do the harder thing and make the bet honestly. There is one and it's feasible.
The bull case is not "AI solves everything." It's narrower: that the demand is real, durable, and willing to pay more than it costs to serve. And that it stays that way long enough for the buildout to earn back before the shovels rust. If that holds, the circular financing in entries 1-2 isn't a Ponzi, it's a bridge. Vendor financing that seeds a market until real outside money arrives to carry it. Railroads were built this way. So were semiconductors.
What would actually need to go right, roughly in order:
- Unit economics have to invert. Cost per query has to fall durably below price per query. The encouraging part is that it has been falling fast and the industry has reorganized itself around tokens per watt.
- Demand has to be real, not rented. Full data centers look like proof until you look closely at who's filling them. The bet is that the seats fill with outsiders paying above cost, not the same money taking another lap.
- The mule needs to plow. The productivity has to be real and captured. This is the load-bearing assumption beneath all the others.
None of these are crazy. They're just unproven, being asked to fund themselves before the proof, at a scale usually reserved for things that are already true. So the honest position isn't "this is a scam." It's "this is a real bet financed as though it were a settled fact, and the distance between those two is exactly where people get hurt." You can think the technology is extraordinary and still think the foundation is rocky.
If we actually mean it: where the public money goes
If the bet is real and we want to win it, the binding constraint isn't chips. It's power. A company can build a data center in one to two years, but getting it connected to the power grid takes 4 to 8 years with operators ensuring wires and transformers can handle the load. Buying more chips doesn't solve this.
And the demand for electricity is enormous. Data centers use about 4% of US electricity today, and newer estimates put that at 9 to 17% by 2030. Where the clusters land, the cost doesn't stay local. Demand from data centers helped push the mid-Atlantic grid's 2025-2026 capacity auction up by more than $9 billion, a bill that gets spread across every customer in the 13 states it covers, whether or not you live anywhere near a server.
So the public money shouldn't chase the circular deals. It should build the thing we need anyway: clean generation (solar, wind, nuclear), storage to run it around the clock, transmission to move it, and the permitting reform to connect all three. Solar, wind, and batteries are the fastest, cheapest capacity to add. Nuclear and geothermal supply the firm, around-the-clock power these sites demand. The DOE itself frames the surge as a once-in-a-generation chance to build clean power.
And it's buildable. China added more than 430 GW of wind and solar in 2025 alone, with close to thirty nuclear reactors under construction, and is on track to add several times more generation capacity than the US this decade. France runs its AI ambitions on a grid that's roughly 70% nuclear.
Which makes US policy decisions look less like caution than like self-sabotage: halting offshore wind, tariffing solar, and watching projected renewable growth fall by half.
The arrangement we should reject is the one that works backwards, where the public absorbs the losses, private investors keep the gains, and the government only steps in to clean up the failure.
If the public is already carrying the risk, it should get something for it: a stake in what it underwrites, ownership of the durable infrastructure it pays for, not just the bill when it breaks. The government took equity in the banks it bailed out in 2008, and public power authorities have owned the grid for a century. The precedent cuts the other way too: Fannie and Freddie were treated as privately held right up until their obligations turned out to have been economically public, and taxpayers absorbed more than $200 billion.
One test for any AI policy that calls itself ambitious: is it building the grid, or backstopping the lap?
So... whose hundred marbles are these?
Trace it back and the money behind Blue Owl, PIMCO, and the rest comes from insurance companies, pension funds, and sovereign funds. Which is to say: the retirement account of someone who has never heard the word "neocloud," the index fund sitting under a regular person's future, the premiums on a life-insurance policy bought to assist families. And you can't easily trace the money, because a chunk of this debt is built, on purpose, to stay off the filings.
When a circular bet built on unproven returns and somebody-else's-problem debt finally settles up, it doesn't settle on the kid who took the fees to enter the game. He went home. It settles on whoever's marbles were in the fund. People who never got a vote and were told the whole time that this was safe and not their concern.