The AI Buildout's Shortage Is Now Money — and the Sellers Are Supplying It

The companies that sell the machinery of the AI buildout have started arranging the money to buy it. Nvidia, the largest of them, is negotiating a $250 billion backstop for OpenAI's data center in Ohio, a guarantee that would stand behind the project's financing ^[1]. Broadcom is arranging chip financing for OpenAI and Anthropic through special purpose vehicles, the legal shells that keep a debt off a company's own books ^[2]. The sellers of the buildout have become its financiers. Broadcom's chief executive Hock Tan was blunt about what most threatens the buildout now.

[What] I think might be the biggest constraint is capital. — Hock E. Tan

He is describing a wall that has moved from the physical to the financial: chip stocks erased more than a trillion dollars in a six-session slide driven by borrowing costs and a revenue miss, not by any shortfall of chips or electricity ^[3]. He still insists demand for compute infrastructure outstrips supply ^[2]. Intel's chief executive Lip-Bu Tan names the same constraint from the other side of the table.

If there's any challenge in the front, it's capital and the cost of capital, and then the other part is in the longer run, people want to see the ROI. — Lip-Bu Tan

OpenAI's finance chief Sarah Friar warned the company may struggle to fund roughly $600 billion in committed computing capacity without accelerated growth ^[4]. The sellers say money is the wall, the buyers warn they may not be able to cash the orders. So the sellers finance. Watch how the financing walks outward, from the seller to everyone else's portfolio. Broadcom is arranging credit for named customers like OpenAI and Anthropic while its own credit-default swaps, insurance against a default, hit records, a sign bondholders are nervous about the very debts being arranged ^[2]. Nvidia goes a step further, negotiating a $250 billion backstop of an entire customer's data center in Ohio and pairing it with a $500 billion partnership with SK Group; critics including Jim Cramer and Moody's Analytics say such arrangements may hide the real level of demand ^[1]. Then comes the last step, so far only a proposal: Larry Fink, head of BlackRock, the world's largest asset manager, wants to securitize the buildout — bundle the data-center loans into securities that can be sold to pension funds and insurers — so the financing no longer sits with the chipmakers and their customers but inside ordinary capital pools ^[5]. The exposure walks outward, from chipmaker to customer to the bond portfolio of anyone with a retirement account. The borrowing is already touching people who have nothing to do with AI. A widening swap spread means the market thinks default is more likely. Since the hyperscalers — the giant cloud companies — began selling debt in volume to fund their buildouts, the spreads on unrelated, high-quality companies have crept up with them: the cost of insuring against default at LVMH, Sanofi, and BAE Systems rose more than 10% ^[6]. BNP Paribas's Josh Farberow put the mechanism in one line.

All high-quality credit competes with hyperscalers for capital. — Josh Farberow

The giants' own default-insurance costs have widened in step. European Central Bank chief economist Philip Lane linked the global borrowing to higher long-term interest rates.

The spillover to the euro area bond market means that the increase in long-term interest rates constitutes a material tightening of financial conditions for the euro area. — Laney Stewart

^[7] The investment firm August counts some $150 billion of the riskiest corporate debt maturing in 2027 and 2028, and Goldman Sachs's Amanda Lynam flags 2028 walls in loans and private credit — a calendar for when this repricing turns into a reckoning ^[8][9]. The bulls have a real answer, and it deserves full weight. Amazon's chief executive Andy Jassy says the new equipment pays for itself in under three years, and the contracted cloud backlog has swollen to $2.4 trillion, roughly a full year of existing cloud revenue ^[10][11]. On the resulting economics, he was unambiguous.

The resulting revenue, free cash flow and return on invested capital is very compelling. — Andy Jassy

Google Cloud grew 82% in a year; Microsoft's Azure passed $100 billion with a $678 billion backlog ^[11]. Demand, in other words, is no longer the dispute. But the same filings carry the tension. In the quarter Alphabet posted that record cloud growth, it also recorded its first negative free cash flow as a public company — $5.9 billion in the red — while raising its 2026 capital-spending guidance to as much as $205 billion ^[11]. Meta's free cash flow collapsed 91% in the same stretch ^[1]. The two largest buyers of AI computing are burning cash at the same moment their revenue hits records. Real demand and unpayable capex are not opposites; they sit on the same ledger, in the same quarter. And the one number that could settle which is winning — how much revenue the AI itself actually produces — is precisely the number no one can see. Two reports on OpenAI's revenue came in $20 billion apart, one near $50 billion and one near $70 billion, and Hock Tan compared pricing such a market to running a casino without reliable data ^[12]. Michael Burry's charge that trillions in data-center debt sit hidden off the books remains his claim and the critics', not established fact ^[13]. The industry's answer to the opacity is telling: Apollo Global Management, a funder of exactly this kind of private credit, moved to daily valuations on its $850 billion book after the SEC warned that a lack of market data is no excuse for sloppy pricing ^[14]. The machine that finances the buildout is now pricing the day it has to show its work.