AI's financial machine is running into a wall of physics

A data center goes up in eighteen months. The power plant and transmission lines to run it take five years or more ^[1]. That gap is not a scheduling detail; it is the shape of the entire AI buildout. Eaton, which makes the electrical equipment data centers run on, reports a U.S. backlog of 307 gigawatts: fifteen years of work at last year's build rate ^[2]. The binding constraint on AI is not money. It is copper, transformers, memory chips, and the years it takes to string wire. Into that gap, Wall Street has built a financial machine. Nvidia has partnered with six of the largest asset managers, including Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, to mobilize more than $500 billion for AI infrastructure, packaging chips as a long-lived, investable asset class ^[3]. Around that sits a vendor-financing platform where outside partners hold the risk while Nvidia books 40 to 50 percent of committed capital as revenue ^[4], and a revenue-sharing program that hands customers GPUs without upfront cash in exchange for a cut of future revenue ^[5]. Nvidia has also put more than $100 billion into OpenAI, Anthropic, Intel, and others: companies whose chip purchases are Nvidia's own revenue ^[6]. The machine was built to bridge the gap between infrastructure spending and the earnings that are supposed to justify it. But the gap it spans is physical, not financial. The physics is now showing up inside the instruments. Memory shortages pushed Nvidia to raise server prices more than 15 percent ^[7], with DRAM contract prices projected to surge as much as 95 percent and SK hynix's parent estimating shortages persist until 2030 ^[2]. Morgan Stanley's analysis of a fully optimized data center finds it costs $25 billion a year to rent and generates $23 billion in output, and the bank expects AI output prices to keep falling ^[8]. The cost of capital is absorbing the same constraint. QTS Realty, an investment-grade borrower, sold $3.9 billion of debt at 7.23 percent, a junk-level yield ^[9]. Applied Digital paid 9.25 percent, a rate approaching distressed territory ^[10]. Meanwhile, the earnings the bridge was meant to reach are not materializing. Goldman Sachs finds no meaningful relationship between AI adoption and productivity across the economy, with a GDP impact of 0.1 to 0.2 percentage points ^[11]. The earnings that do exist are concentrated in the infrastructure firms themselves, which account for roughly half of S&P 500 earnings growth ^[12]. So the loop closes: Nvidia's own investments generate demand for its own chips, and the profits that show up belong to the companies selling the infrastructure, not the ones using it. The bulls are not blind to the physics. BlackRock and Bank of America reject the bubble label, conceding only an "air pocket" where spending temporarily outruns revenue, but even they concede data centers could consume 20 percent of U.S. electricity by the end of the decade ^[13]. That is a physical concession, not a financial one. The financialization was built to bridge a gap that is physical, and the physical gap is now degrading the bridge. That is not a prediction that the boom ends. It is a description of the mechanism by which it is already paying for physics with borrowed money.