Who holds the bill for the AI buildout

The Nasdaq-100 shed a trillion dollars this week, and in the middle of it Microsoft sold $25 billion of new bonds that drew more buyers than it had paper to give them. The same days brought Alphabet its worst single-day fall in more than two decades — 12.6 percent, on a revenue miss and negative free cash flow ^[1]. One market, two verdicts. October did not grade whether AI works; it graded who holds what. The demand side kept saying the same thing it has said all year. In August, the neocloud operator Nebius reported that demand for its compute still outstrips the supply it can offer ^[2], and lab revenue is roughly doubling ^[3]. What moved the tape in October was cash flow, not customers. Amazon's free cash flow hit negative forty-two billion dollars as equipment purchases ballooned ^[4], and its chief executive made clear the spending is not about to pause.

We still expect 2026 cash [capital expenditures] of approximately $220 billion, the majority for AWS, and expect to spend meaningfully more in 2027. — Andy Jassy

None of this was a demand problem. The head of Nvidia, the company whose chips the whole buildout runs on, named the real constraint ^[5].

Because there's just so much demand across so many sectors of AI, there's a scarcity of capital. — Jensen Huang

Not too few customers — too little money to pay for the machines those customers want. That gap is the story of October. The bill did not disappear; it moved. The first place it moved was onto the people selling the machines. In a single October week the industry lined up roughly $150 billion of debt backed by pledged chips, anchored by Broadcom arranging $50 billion for OpenAI's custom silicon and syndicating a $60 billion package for Anthropic — and Nvidia signed on to backstop more than $100 billion for OpenAI's Ohio campus ^[3]. To get those deals across the line, the suppliers began paying the labs with their own paper. AMD handed OpenAI and Meta warrants — the right to buy up to 320 million shares later at a penny apiece — and Broadcom wrote residual-value support into Anthropic's $60 billion package, a promise to cover part of the loss if the pledged chips ever resell for less than the debt they secure ^[3]. These are the same instruments the auto industry once used to move cars nobody had finished paying for, applied now to chips. And the chipmaker that sells the machines is now trying to insure the ones it already sold: Nvidia has been passing its depreciation data to insurers and reinsurers, working with the broker Howden Re, so lenders are protected if a borrower defaults and the pledged chips cannot be resold for the full value of the debt ^[6]. The next rung is the insurance market. So far it has not produced a price; it has produced a verdict. Greg Case, the chief executive of Aon, the world's largest insurance broker, says the clustering of equipment has made these projects too concentrated to insure, and that the industry's roughly $4.5 trillion capital base would need to double or triple to absorb the risk ^[7].

If it's more difficult to get insurance, that means the financing structure is going to change completely. You can't access the debt markets. — Greg Case

The next rung is the states, and here the jam is literal. Texas Governor Greg Abbott placed a hold on new projects while the state studies power-grid risk, and Morgan Stanley had to delay a $5-to-7 billion bond sale meant to refinance Anthropic's $15 billion Hubbard data-center loan — a decision about electricity forcing the banks to carry the debt longer than planned ^[8]. The grid itself is the meter. Upgrade costs are spread across every business and resident with an electricity meter, and Maryland has forecast up to $216 a year added to household bills ^[9]. The next rung is Washington — the only rung that has answered, and the answer was no. In September, OpenAI's chief executive Sam Altman asked the federal government to serve as an insurer of last resort, the backstop when no private insurer will write the risk ^[10].

Given the magnitude of what I expect A.I.’s economic impact to look like, the government should serve the role of "insurer of last resort." — Sam Altman

Ten months earlier he had said in public that governments should not pick winners or losers and that taxpayers should not bail out companies that lose in the market ^[11].

We believe that governments should not pick winners or losers, and that taxpayers should not bail out companies that make bad business decisions or otherwise lose in the market. — Sam Altman

And the administration had said no in November 2025 — the same month Altman disowned the idea, ten months before he asked for it anyway ^[11].

There will be no federal bailout for AI. — David Sacks

The next rung asks no one's permission. New rules let exchange-traded funds hold up to 15 percent of their assets in illiquid securities — holdings that can't be sold quickly at a fair price — through side vehicles, separate funds set up to hold what the ETF's own rules bar; one fund now holds Anthropic stakes worth 9 percent of its assets ^[12]. And the largest single handoff is already on paper. Anthropic's IPO filing targets a valuation near two trillion dollars and could raise up to $100 billion from the public ^[13]. That filing is worth reading in one piece, because it carries both halves of the story. Compute costs fell from more than 250 percent of revenue to 48 percent, and the company posted a quarterly operating profit — the product works ^[14]. The same document then warns that long-term survival depends on continued growth to meet its financial obligations: the commitments, not the customers, are what outrun it ^[14]. That sentence is printed on the same paper that asks the public for up to $100 billion ^[13].