Opinion: Wall Street's AI panic vanished in three days. The debt behind it didn't.

For about two weeks in late October and early November, the stock market did something it had not done since the dot-com unwind of 2000: it forced investors to ask, seriously, whether the most expensive companies in the index were worth what the index said they were worth. The Nasdaq 100 set a record close of 31,224.47 on October 6, then fell more than 10 percent by November 2, meeting the standard definition of a correction. On November 5 alone, estimates of the one-day loss in Nasdaq market value ranged from $2.3 trillion to $2.5 trillion, among the steepest single-session drops on record for the index. Then, on November 9, the Nasdaq 100 jumped 3.3 percent as Morgan Stanley and UBS told clients the rout had gone too far. Nvidia, tellingly, kept falling even on the rally day, down 4.6 percent, as some investors grew more selective about which AI names they trusted.

The speed of the reversal is being read on trading desks as proof that the panic was overdone. That reading is premature. What the past month has actually shown is that a handful of companies — Nvidia, Microsoft, Oracle, OpenAI and the "neocloud" data-center operators that rent them capacity, chief among them CoreWeave — have built a financing structure complicated and leveraged enough that a single bad week in sentiment can erase trillions of dollars, and that the structure itself has not changed at all since the rally resumed.

The debt behind the buildout

The clearest evidence sits in Nvidia's own regulatory filings. In its quarterly report covering the period through July 26, 2026, filed with the Securities and Exchange Commission, Nvidia disclosed that its supply and purchase commitments to chipmakers and component suppliers had more than doubled in a single quarter, from $119 billion to $279 billion, as part of $366 billion in total future commitments. The same filing disclosed a subsequent event: guarantees, capped at $105 billion, that Nvidia agreed in August to provide on data-center leases covering roughly 4.25 gigawatts of capacity being built by SB Energy on behalf of an OpenAI affiliate, bringing Nvidia's total guarantee exposure to $108.5 billion. In plain terms, the company that sells the chips is now backstopping the real-estate and power infrastructure of the company that buys them, a structure critics have compared to the vendor financing that inflated, and then helped sink, telecom valuations a generation ago.

Credit markets are pricing that arrangement with increasing unease. By early November, the cost of insuring CoreWeave's debt against default had risen to roughly 1,155 basis points, implying a five-year default probability that has more than doubled over the past year, while Oracle's five-year credit-default-swap spread sat near 298 basis points, up sharply from earlier in the year after S&P Global Ratings cut Oracle's credit grade and cited uncertainty over how it would pay for its AI commitments. Nvidia's own CDS spread, by contrast, remained far lower, near 139 basis points — a sign that bond investors are already distinguishing between the company sitting atop the AI supply chain and the companies borrowing heavily to build it out beneath it.

The case that this is different

The strongest argument against alarm is that, unlike the dot-com-era companies, today's AI leaders have real and growing revenue. Nvidia's data-center sales have scaled into the hundreds of billions of dollars; OpenAI has tens of millions of paying users; enterprise software firms are reporting measurable productivity gains. The analysts who told clients last week's sell-off was overdone were making a specific claim: that AI demand is not speculative in the way late-1990s fiber-optic capacity or e-commerce traffic projections were, and that a market correcting for excessive multiples is healthy repricing, not the unwinding of a fraud. There is also a fair governance point in Nvidia's own disclosures: the company explicitly warns investors that it offers "no assurance" the OpenAI arrangements will be completed as structured, language that at least puts the risk on the record rather than hiding it.

Why the rally is not the answer

None of that resolves the deeper problem, which is less about whether AI products generate revenue and more about how concentrated, debt-financed and economically load-bearing the buildout has become. The Federal Reserve's own Financial Stability Report, published in May, found that the market contacts it surveys most often cited risks from artificial intelligence as the most salient threat to financial stability, ahead of geopolitical shocks and persistent inflation, and that equity valuations "have remained high" by historical measures even before the autumn's volatility. The Federal Open Market Committee, meeting on October 28, voted 10-2 to hold its benchmark rate at 3.75 to 4 percent, with two governors dissenting in favor of a cut rather than a hike — a split that the Fed's own statement attributed partly to inflation that "remains elevated," leaving little room for the kind of rate relief that bailed out leveraged growth stocks in past downturns.

Meanwhile, AI-related investment is no longer a side bet for the economy; it is a substantial share of what is keeping the economy growing at all. In its advance estimate of third-quarter output, the Bureau of Economic Analysis reported that real GDP grew at a 3.4 percent annual rate. Independent economists parsing the bureau's own contribution tables have calculated that business spending on information-processing equipment and software — a category dominated by data-center and AI infrastructure, though not exclusively AI — added roughly 1.09 percentage points to that growth, close to a third of the headline number. When a category this narrow accounts for that much of the country's growth, and when the companies inside it are leaning on guarantees and vendor financing rather than retained earnings to fund it, a sentiment shock in equities is not a contained event. It is a stress test of the mechanism now doing a disproportionate share of the work of keeping the broader economy expanding.

That is the argument for treating the November 9 rebound as a reprieve rather than a verdict. Oil prices falling and short covering can move an index 3.3 percent in a session; they cannot retire $108.5 billion in guarantees, reverse a credit-rating downgrade, or change the fact that Nvidia's supply commitments more than doubled in three months. Regulators do not need to decide today whether AI spending is a bubble in the strict sense. They do need to start treating the credit exposures running through non-bank lenders, private credit funds and vendor-financed leases with the same scrutiny the Fed already applies to bank balance sheets, and investors pricing AI-adjacent debt and equity together need to stop reading a one-week rally as proof the underlying structure has been tested and passed. Nvidia reports its next quarterly results on November 17, five days from now, and that report — not the relief rally that preceded it — will be the first real evidence of whether the demand underneath this buildout is as solid as its boosters insist.