The AI trade's real divide is a balance sheet

In July, Meta broke the industry's favorite story. The story was simple: own the chips and win. Meta had spent somewhere between $130 billion and $145 billion on GPUs this year, more than anyone else ^[1]. Then it started selling the surplus. Meta Compute launched as a direct competitor to AWS, Azure, Google Cloud, and the neoclouds that had built their entire pitch on scarcity. CoreWeave and Nebius fell 17% on the news ^[2]. The read from analysts was blunt.

We haven't done that yet, because we think that we have a use for the compute. But obviously, if we get to a point where we feel that we have overbuilt, then that is an option that we have, and that is partially what gives us confidence in investing in building this out. — Pavel Durov

The biggest buyer of AI hardware in the world had overbuilt, and was now competing with its own suppliers. The strange part is that the physical scarcity was real. Micron's high-bandwidth memory is sold out through 2027 and into 2028, locked in by non-cancellable agreements that run to 2030 ^[3]. Infrastructure stocks fell on July 4 anyway. The market was pricing a spending plateau: that the buyers would stop buying before the sellers ran out. At the other end of the stack, the model makers can't turn usage into profit. Leaked OpenAI documents show inference costs climbing from $3.8 billion in 2024 to $8.65 billion in the first nine months of 2025 ^[4]. The price of a token fell 80% in a year ^[5]. Even at a billion dollars a month in revenue, OpenAI is unprofitable and, by its CFO's own account, perpetually short of compute ^[6]. The financing that connects the two layers is where it gets uncomfortable. Bill Gurley has flagged deals in which tech giants fund startups that then spend the money on the investor's own cloud services ^[7]. Steve Eisman puts the entire hyperscaler position in one sentence.

The futures of these massive companies, in a sense, are a bet that OpenAI and Anthropic are going to succeed. — Steve Eisman

Goldman Sachs, for its part, found no meaningful relationship between AI adoption and economy-wide productivity ^[8]. The inversion isn't sorting the industry into hardware winners and software losers. It's sorting by balance-sheet thickness. Amazon raised capital spending 72% to nearly $120 billion, and it can do that because e-commerce cash flow pays the bill ^[9]. Google funds its build with advertising. Oracle sits on the other side of the ledger. Its stock is down 57% from its peak, and it carries $129.5 billion in debt after $55.7 billion in capital spending. Analysts now question its $300 billion OpenAI contract, since OpenAI reported a $38.5 billion net loss and would need to pay roughly $60 billion a year ^[10]. OpenAI's CFO has named the trap herself. The company needs its own chips to keep its models out of competitors' hands, but it can't afford them. Renting from a cloud provider means teaching that provider how to build the very thing OpenAI sells.

If all we do is buy from others, all we’re doing is giving them our IP because they’re learning how to build AI infrastructure. — Sarah Friar

She is describing the cage from inside it.