The AI Buildout Is Being Borrowed Into Existence

Share

For a while, the story of AI money has been a venture story. Investors back the labs the way they used to back startups, placing a bet that one model will pull ahead and win a market that looks too big to stay shared. To most people, that is still what is happening, a repeat of every platform race we have seen, a crop of richly funded challengers fighting to become the default. Then you look at where the money actually coming out this week is coming from, and it starts to look less like venture and more like a railroad being financed in the nineteenth century, on borrowed money, on commitments, and on the credit of rivals. The AI buildout has crossed from equity into credit, and that changes more than how it is paid for.

The cleanest place to see it is in Anthropic's revised IPO filing, published this week. Revenue for the latest quarter came to roughly $17 billion, up more than tenfold from a year earlier, a number that would be extraordinary for almost any company in any industry. The same filing showed something humbler underneath. Anthropic reported a loss again for the quarter and disclosed about $895 billion in committed purchase and lease obligations, almost all of it tied to compute and infrastructure. So the company growing revenue faster than almost anyone in recorded business history is simultaneously on the hook for close to a ninth of a trillion dollars of future spending that it has already promised to make.

That second number is so large it stops being about one company. Nearly $900 billion of committed compute is more than a single market can absorb without bending, and it tells you something about the whole industry rather than about Anthropic. The frontier is no longer funded by venture rounds that hand out a valuation in exchange for equity. It is funded by commitments so big they have to be financed like infrastructure, which means borrowed money, long leases, and counterparties strong enough to underwrite them. When a buildout is that large, the financing stops being a detail of the strategy and becomes the strategy itself.

And this is where the irony lands. The company standing behind the financing for Anthropic's data center is Google, its most direct rival in frontier models. TeraWulf, the data center developer, secured credit support from Google to borrow against a roughly $19 billion, 20 year lease to house Anthropic at a campus in Kentucky. Google and Anthropic spend every product cycle trying to beat each other. Yet when it came time to make a data center financeable, Anthropic's biggest competitor lent it the creditworthiness to do it. You would be hard pressed to design a clearer sign that competition and credit have become so entangled that a clean rivalry is hard to find.

The same pattern runs through the other side of the frontier. OpenAI reported closing a round of about $122 billion at an $852 billion valuation, the largest private round ever recorded. But it is worth noticing who wrote the checks. Amazon put in roughly $50 billion and is named the exclusive cloud partner. Nvidia put in about $30 billion, selling into the very buildout it helps to create. These are not outsiders placing a detached bet on a likely winner. They are the giants who want to sit inside the model's orbit, and they do not arrive with a check alone. They arrive with commercial arrangements attached, the cloud contract, the chip supply, the guarantee of capacity, and those arrangements are part of why the number can be that large.

Strip out the brand names and a different picture emerges. The people financing the frontier are mostly the people who want to sell it something: the cloud, the chips, the data center space, the power. That is not a venture dynamic. It is a supplier dynamic, and it quietly shifts who holds the leverage. A vendor does not set the terms of the relationship the way an equity investor does, and a lender does not set the terms the way either of them does. The shift is visible across the sector. Oracle filed to sell about $20 billion of stock and notes to fund its buildout this week, and Meta borrowed another $15 billion in debt. The whole industry is crossing from letting investors own a share of the upside to owing the money markets a return on principal regardless of how things go.

This deepens the dependency problem I have written about before rather than simplifying it. The old version of that problem was clean. A startup that builds on one model risks that provider changing its prices, its product, or its terms, and the startup has nowhere else to go. The new version is stranger, because now even the providers are dependent, on each other and on borrowed money. Anthropic depends on Google's credit to stand up its own capacity. OpenAI's biggest investor is also the cloud it runs on and one of its biggest suppliers. The rivalry everyone talks about, Google versus Anthropic, Amazon head to head with OpenAI, is also a web of credit and commercial dependence so tangled that neither one can act without leaning on the other.

To be fair, this is not obviously a bad thing, and it is not unusual. Capital heavy businesses have always leaned on debt, because when the growth is real the returns can compound faster than the cost of the money. Railroads, telecom, and the first phase of the cloud all went through a stretch of being financed by people who wanted to be near them. And there is a genuine case that this is strength rather than weakness. If Amazon is willing to put $50 billion into OpenAI and back its infrastructure, that is private conviction expressed in the most expensive way available. It is not obviously wrong to read the lending and the guarantees as confidence instead of panic. The people closest to the deal may simply believe the demand is as big as it looks.

But it is worth being honest about what changes when a sector turns to credit. Leverage does not care how good your story is. A company financed by debt has to keep paying whether the bet wins or not, whether the next model ships on time or a cheaper rival undercuts it. In the very quarter where Anthropic's revenue grew more than tenfold, it still ran a loss on top of roughly $900 billion of commitments ahead of it. There is a version of this decade where the frontier works out and the whole thing pays for itself. There is also a version where a miss, a regulatory setback, or a cheaper competitor turns those commitments from a moat into a millstone. Credit is what lets you build the future faster than the present would otherwise allow, and credit is also what makes the future unforgiving when it arrives late.

The pattern confessing itself here is not specific to AI. A generation ago the same rhythm ran through the first internet boom and the cloud buildout. Infrastructure so necessary that the people supplying it ended up financing it, and then the people who financed it ended up owning a great deal of the eventual outcome. The durable winners of those booms were rarely the companies that borrowed the most. They were the ones that borrowed at the right time and had something lasting, a relationship, a distribution, an owner of the customer, sitting on the other side of the ephemeral thing, the website, the app, the model, the raw capacity. The thing being funded was necessary. It was not, by itself, the thing that was worth owning.

I am not predicting a reckoning, and I do not think the frontier is about to blow up. The revenue is real, far more real than it was a year ago, and the demand for what these systems do is not manufactured. But it is worth noticing where the money is coming from now, because it tells you what the people closest to the deal actually believe. They are lending against the infrastructure and the relationship, not buying a share of the intelligence. The fate of the frontier is still in dispute. The party that gets paid whether anyone wins or not is not, and when your biggest rival is willing to guarantee your rent, the competition has been priced one way and the credit another.