The story Wall Street wants to tell itself today is one of infrastructure triumphant. Nvidia has signed memorandums of understanding with six of the largest financial firms—Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR—to create financing platforms that could mobilize more than $500 billion in third-party capital for AI compute 2. The pitch is elegant: treat data centers like toll roads, long-duration assets with predictable cash flows, and let the capital markets do what they did for pipelines and power plants. It is a narrative of maturation, of an industry moving from speculative venture to bankable utility.
The number that does not fit is not Nvidia’s. It is Anthropic’s. The AI startup behind Claude is reportedly preparing for an October IPO at a valuation of at least $2 trillion 10, a figure that would surpass SpaceX and make it the largest debut in history. To justify that price, investors are leaning on a forecast that second-quarter revenue exceeded $11.5 billion—more than 14 times the $787 million from a year earlier—with positive adjusted operating income 9. That growth is real, but it is the kind of number that demands a specific question: who is buying, and with whose money?
Follow the contradiction. Nvidia’s plan is to treat AI compute as an asset class, enabling customers to finance their purchases rather than pay upfront 2. SK Hynix has approved a $38.3 billion investment in two new memory chip plants, betting on sustained AI-driven demand for DRAM and HBM, with first cleanrooms opening in 2029 3. The supply chain is borrowing against a future that is itself being financed by the same lenders underwriting the demand. The accounting logic is circular but not fraudulent: Nvidia books the sale, the financial firm books the asset, and the customer books the compute. Everyone’s balance sheet looks better, provided the underlying workloads generate the cash flows to service the debt.
The incentives are the problem. When a chipmaker helps its customers borrow money to buy its chips, the sales cycle becomes a financing cycle. The risk is not that Nvidia is lying—it is that the company has every incentive to keep the machine running at full capacity, and the financial firms have every incentive to keep collecting fees on structured products they may not fully understand. This is not a novel pattern. It is the same architecture that gave us commercial real estate debt and, before that, collateralized debt obligations. The asset class changes; the leverage does not.
The defenses are predictable and not unreasonable. The firms involved are sophisticated, the assets are physical and identifiable, and the demand for AI compute is not a fiction—Anthropic’s revenue growth is evidence of real customers spending real money 9. The Fed’s dilemma adds context: Kevin Warsh faces a September meeting with inflation still running at 3.4% and wholesale prices slowing, but three regional presidents already dissented in July in favor of a hike 8. If rates rise, the cost of carrying $500 billion in new AI debt rises with them.
The unresolved risk is not a crash. It is a quiet repricing. If the IPO market stumbles, or if one major AI customer consolidates its spending, the gap between the book value of financed compute and its market value will close slowly, not suddenly. The consequence for the reader is a choice: treat the Nvidia financing platforms as evidence of AI’s arrival as a mature asset class, or as a sign that the industry’s largest player is now its own most enthusiastic lender. The tradeoff is between growth and transparency. The question that matters is not whether the $500 billion gets deployed, but whether the debt behind it is priced for a world where AI demand grows at 14x every year—or merely at 3x.