Thursday was the worst session on record for the Nasdaq Composite measured in points. The index fell 1,422 points, or 5.52 percent. The Philadelphia Semiconductor Index dropped more than 10 percent. Nvidia lost roughly nine percent of its value, AMD fell about 17 percent, and Qualcomm, which had reported earnings the night before, gave up 11 percent. The convenient explanation is that a bubble finally popped. The useful explanation is narrower. The market repriced the political permission that the artificial intelligence buildout has been quietly relying on, and it did so in a single afternoon.
The trigger was not a chip. It was a map.
Democrats won both chambers of Congress in the midterms, and they also flipped governorships in states that sit at the center of data center construction. Texas, Georgia and Ohio are three of the busiest fronts in the current buildout, and all three changed hands. In Texas, governor-elect Gina Hinojosa campaigned on halting new data center approvals until the legislature passes laws covering electricity, water, air quality and property rights.
That is the mechanical reason semiconductor and data center equities fell harder than software. A presidential veto can stop a federal statute. It cannot stop a state permitting decision. Investors who had priced the buildout as a technology problem learned, in one session, that it is also a land use problem, and land use is decided by people who just won elections promising to say no.
The scale involved is not marginal. A June 2026 review by the Texas Tribune counted 335 operating data centers in the state and at least 248 more planned or under development. UBS has estimated that roughly 30 percent of the new US data center capacity projected for 2028 could be affected if approvals stall. When a third of the pipeline acquires a political risk premium, the supply chain behind it gets repriced at once. That is what happened to cooling vendors, power equipment makers, transformer suppliers and the neoclouds that lease capacity by the hour.
The second leg of the selloff was credit, and it had been flashing for months. Technology companies have issued close to $500 billion of new debt this year to finance AI infrastructure. Goldman Sachs expects hyperscaler bond issuance to climb from about $250 billion in 2026 to $420 billion in 2027, with roughly 35 percent of next year capital spending funded by debt. Those are not the numbers of a sector paying for a construction cycle out of cash flow. They are the numbers of a sector that asked the bond market to underwrite projects whose revenue arrives years later.
The bond market has already answered. Oracle five-year credit default swaps have widened to around 263 basis points, and CoreWeave has set records of its own. The notes tell the same story at a glance: recent long-dated SpaceX paper near 83 cents on the dollar, Nvidia notes due 2060 near 63 cents, CoreWeave notes issued this year near 78 cents. Bond desks did not need an election result to get nervous. They needed one to reprice.
What did not change on Thursday is the earnings picture. Bloomberg Intelligence data puts third-quarter S&P 500 earnings growth at 31 percent year over year, up from a July estimate of 22 percent. By Thursday, 418 index companies had reported and 84 percent had beaten estimates. Capital spending is still being funded, orders are still being placed, and the demand that built this trade has not disappeared. That gap between a repricing and a collapse is why the dot-com comparison is lazy.
Daniela Hathorn, senior market analyst at Capital.com, framed the caution cleanly: investors remain worried about elevated technology valuations and about whether substantial AI spending can generate sufficient returns. She is right about valuation. But valuation and demand are separate questions, and Thursday mostly answered the first one.
The distribution of pain shows which question the market was actually asking. Hyperscalers fell, but they can shift a project from a state that says no to one that says yes, and they can fund construction from operating cash. The deepest losses landed on the middle layer: Applied Digital down roughly 19 percent, Nebius about 17 percent, CoreWeave about 14 percent. Those companies typically depend on a small number of sites and on uninterrupted access to capital markets. A delayed permit is an inconvenience for a hyperscaler. For a neocloud it is a covenant test.
For businesses that sell into this buildout, the lesson is about scheduling rather than sentiment. Equipment suppliers, engineering contractors, cooling and power vendors, and the freight forwarders who move transformers and servers do not watch demand evaporate in a week. They watch delivery dates move. A moratorium does not cancel an order book. It stretches it, and it shifts who carries inventory risk while the paperwork clears. Companies that model commodity prices with care still model approval timelines casually. That is the exposure Thursday exposed.
There is a rate dimension as well. Cleveland Fed President Beth Hammack said inflation has moved in the wrong direction since March and that policy will need to move to a restrictive stance. Futures markets put roughly a 69 percent chance on a quarter-point hike at the December 8-9 meeting. Long-duration capital projects are the assets most sensitive to a higher discount rate, and AI infrastructure is the longest-duration project in the economy.
The uncomfortable conclusion cuts both ways. The bears who declared a bubble on Thursday are overreading one session. The bulls who called it noise are underreading a real change in the rules of the game. Nothing about the technology got worse this week. Something about the permission to build it got harder, and permission is what a construction cycle runs on. The companies that come through this in good shape will be the ones that treat a zoning hearing with the same seriousness they treat a silicon allocation.
