Skip to content
Personal essay·Economics·5 min read·Sep 20, 2026

Who Pays for the Engine

The technology is real, and it might be the next steam engine. The money story wrapped around it is a different matter.

Something about the AI industry keeps bugging me, and it isn't the technology.

The biggest companies in the industry invest in each other and buy from each other at the same time. Money goes out as an investment and comes back as revenue, after passing through two or three partners. Nvidia puts money into OpenAI.1 OpenAI pays Oracle for computing power. Oracle buys the chips from Nvidia.2 Everyone's numbers go up, and what holds it all together isn't cash anyone has earned yet. It's what they expect to earn, and what everyone says everyone is worth.

On the circle

AMD is my favorite example, because it's so plain. It gave OpenAI the right to buy its shares for a cent apiece, as long as OpenAI keeps buying its chips.3 The supplier hands the customer a piece of itself so the customer keeps buying. Microsoft, Amazon and Google do versions of the same thing.4 Journalists already have a name for it: circular financing.5

And the numbers don't hide it well. OpenAI told investors it plans to spend about $600 billion on computing by 2030. This year it's reportedly on track to make somewhere around $25 billion, and lose around $14 billion.6

Or look at it from another side. Nvidia, a company that makes chips, is worth more than all the farmland in Australia, and not by a little. That's land that feeds some 75 million people.7 One is where food comes from. The other makes chips. That's how far up the price of everything AI has gone.

On the discount

Right now, everything is on sale. Tokens are cheap. Pay $20 and, if you use it hard, you can burn through many times that in tokens at API prices. Nobody outside the companies knows exactly what it costs them to serve you. What we do know is that the companies as a whole are losing money, and investors are covering the difference, not you.8

Investors are paying because they were promised AGI. Someone will win the race, the winner takes everything, and the loss is the price of the ticket. But look at who made that promise, and who believed it. People. Not LLMs. The people who are supposed to run the economy heard “AGI” and kept writing checks.

On the engine

I'm not sure any of this bursts. That's the honest version.

Maybe this is a new steam engine, and this one improves itself. If so, the usual math doesn't apply. Or maybe every new model is just a bit closer to the last one, progress flattens, nobody wins, and the discount ends the day that becomes obvious. I hold both. They point in opposite directions.9

What I do know is what happens to a bubble around a real technology, because it already happened. In the 1840s Britain went crazy for railways. Shares climbed until they collapsed, investors lost their money, and about a third of the planned lines were never built. But the tracks that were built stayed, and they're more than half of Britain's network today.10 The shareholders lost. The tracks stayed.

On who pays

What makes me angry isn't the technology. I'm not an idiot, I'm not going to argue against progress. What makes me angry is the padding: people dressing up the story around it when they don't even need to, because progress keeps going without it. Sooner or later that comes back to the whole world, the way it did in 2008 with mortgages.11 I'm not saying AI is a mortgage. I'm saying the shape looks the same: something real underneath, a padded story on top, and a bill that reaches people who never signed anything.

It'll be harder to see this time. In 2008 you could point at a house. Now it's frozen hiring, smaller pensions, and a bad few years blamed on “the economy.”12

On the other bill

There's another bill, and it isn't on any balance sheet. We're getting weaker thinkers. When the model does the thinking, you stop practicing it. I said in the first essay that back then you knew every corner of your own project, and that today most people won't, because people are lazy. Someday somebody will have to check the model's work, and they'll never have had to learn how.13

We'll pay for that later, too. Nobody puts that on a slide.

Closing

None of this is about the tool. Generative AI is very good, and I use it every day. That has nothing to do with whether OpenAI or Nvidia have their books in order. The companies are building up valuation and investment and feeding people the illusion of profit. The tool is a different story.

The engine works. What I don't know is who pays for it, who owns the track when it's built, and who's left holding the bill if the story turns out to be just a story.

Same field, seen from the money side.

“AI is far more dangerous than nukes… So why do we have no regulatory oversight? This is insane.”14
— Elon Musk, SXSW, 2018. Today he runs SpaceX and its AI arm, SpaceXAI (formerly xAI): Grok, X, and, since August, Cursor.15

Notes
  1. Nvidia and OpenAI

    In September 2025, Nvidia announced that it intended to invest up to $100 billion in OpenAI, released as OpenAI deployed Nvidia systems. That was a letter of intent, not a finished deal. Reports this year say it was replaced by an equity stake of about $30 billion (CNBC; Yahoo Finance, where Jensen Huang says the $100 billion is “probably not in the cards”).

    Fortune's coverage from that September spells out the pattern: Nvidia “invests in, or lends money to, its own customers,” and the money comes back as GPU purchases. A Wall Street analyst it quotes estimates that for every $10 billion Nvidia invests in OpenAI, it sees about $35 billion of chip purchases or lease payments. That ratio is an estimate, not a reported figure.

    ↩ Back to the text
  2. Oracle

    OpenAI signed to buy $300 billion of computing power from Oracle over five years, starting in 2027 and requiring 4.5 gigawatts of capacity (Built In). Separately, in May 2025 it was reported that Oracle would spend about $40 billion on roughly 400,000 Nvidia GB200 chips for the first Stargate site in Abilene, Texas, and lease the computing power to OpenAI (Network World).

    Those are two different numbers. I use the loop as a picture, and I'm not claiming the $40 billion belongs to the $300 billion contract.

    ↩ Back to the text
  3. AMD's warrant

    In October 2025, OpenAI agreed to deploy up to 6 gigawatts of AMD chips, and AMD gave OpenAI a warrant for up to 160 million AMD shares at one cent each (AMD's announcement; Tom's Hardware).

    It isn't a gift. The warrant vests in tranches, tied to deployment milestones and to AMD's share price reaching targets that go up to $600 for the final tranche. At full vesting it would be close to a tenth of AMD's shares.

    ↩ Back to the text
  4. Microsoft, Amazon and Google

    Microsoft says it holds a stake in OpenAI “valued at approximately $135 billion, representing roughly 27 percent on an as-converted diluted basis,” and that “OpenAI has contracted to purchase an incremental $250B of Azure services.” Stake in, cloud spending back.

    In April 2026, CNBC reported that Google would put $10 billion into Anthropic now and up to $30 billion more if targets are met, up to $40 billion in all. Amazon put in $5 billion with an option for up to $20 billion more. Each reserved compute on its own cloud for Anthropic. “Up to” figures are ceilings, not money already delivered.

    ↩ Back to the text
  5. Circular financing, and what happened in July

    The term comes from the press. The Fortune piece in note 1 is a good example.

    On July 27, 2026, the Wall Street Journal reported that Nvidia was weighing a guarantee of up to $250 billion for an OpenAI data center in Ohio. A guarantee here means Nvidia would stand behind the borrowing: if OpenAI can't pay, Nvidia does, and the borrowed money gets spent on Nvidia's chips. Nvidia's stock fell nearly 5% that day. A few weeks later the guarantee was reported cut to under $120 billion for the first phase, after investors pushed back (Network World; Fortune). That part matters too: the loop isn't only growing.

    ↩ Back to the text
  6. OpenAI's spending and losses

    In February the company told investors its compute target is about $600 billion through 2030 (CNBC; PYMNTS). Before that, its CEO had talked about $1.4 trillion in infrastructure commitments; the $600 billion is a planning target, a different kind of number. Spread over five years it's roughly $120 billion a year. That division is mine.

    One analysis, drawing on the company's own disclosures and press reports, gives $13.1 billion of revenue in 2025, an annualized run rate of roughly $25 billion by mid-2026, and a 2026 loss projected around $14 billion on the company's internal measure, with some outside estimates higher. These are reported figures, not audited ones. The bet behind the loop is that revenue catches up.

    ↩ Back to the text
  7. Nvidia and Australian farmland

    Nvidia's market cap was about $5.4 trillion in September 2026 (companiesmarketcap.com); it moves daily. The National Farmers' Federation says Australian farmers “produce enough food to feed 75 million people, there are only 25 million Australians,” and that farmers manage 51 percent of Australia's land mass. That's the industry body's own claim.

    I couldn't find a published total value for all Australian farmland, so I don't state a ratio. What I can do is put a ceiling on the land. Fifty-one percent of Australia's roughly 769 million hectares is about 390 million hectares. At the 2025 national median price for traded farmland, about A$10,500 per hectare (Bendigo Bank), that comes to roughly A$4.1 trillion, or about US$2.7 trillion at an exchange rate of 0.65. That is a generous overestimate, since most of that land is low-value pastoral country and the median reflects farmland that actually changes hands. Even so, Nvidia comes out about twice as large. This is my arithmetic, not a published figure.

    ↩ Back to the text
  8. The discount

    “At API prices” means what the same usage would cost if you bought tokens through the API at list price. That isn't the same as what it costs the company to serve you, and estimates of that differ. One analysis puts a heavy user's actual compute cost at roughly $4 to $20 a month, under the $20 price: the average user is profitable, and losses concentrate among the heaviest users, who can cost three to five times more than that baseline. Another piece argues that current pricing reflects market strategy rather than cost recovery, and notes that OpenAI has acknowledged that heavy users of its $200 Pro plan can still cost it money.

    So the claim in the text is narrower than “every subscriber is subsidized.” It's that the industry as a whole spends far more than it earns (note 6), and investor money covers the gap.

    ↩ Back to the text
  9. Two readings

    Both are my own readings, not measured facts. Recursive self-improvement is an open hypothesis, not a settled result. “Each model is closer to the last one” is my impression of how much better models feel from one generation to the next, not a benchmark result. I state both as open on purpose.

    ↩ Back to the text
  10. Railway Mania

    From Wikipedia's article: in the mid-1840s, shares in new British railway companies rose as speculators bought more, which pushed prices higher, until they collapsed. At the peak, in 1846, Parliament passed 263 Acts to set up new railway companies, with proposed routes totalling 9,500 miles. About a third of the authorised railways were never built: some companies collapsed, some were bought out before they could build, some were fraudulent. Still, 6,220 miles of line were built from projects authorised between 1844 and 1846, against a modern UK network of around 11,000 miles. That's where “more than half” comes from.

    The analogy has limits. Railway track lasts for generations, while chips and data-center equipment age out in years.

    ↩ Back to the text
  11. 2008

    The story I lean on: in the mid-2000s, lenders made mortgage loans to buyers who often couldn't repay them, bundled the loans into securities rated as safe, and sold them on. When house prices fell and borrowers defaulted, those securities lost value, and the banks and funds holding them failed or needed rescue. Lehman Brothers went bankrupt in September 2008, and the damage spread to jobs, savings and credit worldwide (overview). The loans and the houses were real; the story that they were safe was not. That's the shape I mean, and only the shape. I'm not saying AI works like mortgage securities.

    ↩ Back to the text
  12. Pensions and GPU debt

    One July 2026 report argues that CoreWeave's roughly $24.9 billion of debt has been packaged partly into investment-grade instruments, including an $8.5 billion facility rated A3 by Moody's, which pension funds and insurers are allowed to hold, so part of the risk sits with them. The line in the text about smaller pensions is my prediction of how a bust would spread, not something that report says will happen.

    ↩ Back to the text
  13. What the research says, and what it doesn't

    This is my prediction. The studies point the same way but don't prove it.

    Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about real uses of generative AI at work. The more they trusted the AI, the less critical-thinking effort they reported; more confidence in their own ability went with more critical thinking. It's self-reported survey data, so it shows an association, not a cause.

    In an MIT Media Lab study, “Your Brain on ChatGPT,” people wrote essays with an LLM, with a search engine, or with no tools, and their brain activity was measured. When the LLM users later wrote without the tool, they showed weaker neural connectivity than people who had never used it. The authors call the pattern “cognitive debt.” It's a small study of one task, so it's a signal, not a verdict.

    ↩ Back to the text
  14. The quote

    Said at the South by Southwest conference in Austin in March 2018 (CNBC). In the same talk he said AI “scares the hell out of me,” and that he's “not normally an advocate of regulation and oversight,” but that here “you have a very serious danger to the public.” It's an eight-year-old quote, which is why the text gives the year.

    ↩ Back to the text
  15. Who he is now

    SpaceX acquired xAI, which owns X, in February 2026 (CNBC); xAI was rebranded SpaceXAI in July. SpaceX completed its acquisition of Cursor, a $60 billion deal, on August 14. Grok is SpaceXAI's model. I write that he “runs” SpaceX rather than “owns” it: he's its CEO, not its sole owner.

    ↩ Back to the text
Comments
  • Be the first to comment.