Deep Dive — Nasdaq's $2.3 trillion selloff and the 2027 capex peak
Thursday was the worst day for US tech stocks since 2020, and the cause was not a missed earnings number or a bad inflation print. It was a set of county ballots.
The Nasdaq 100 fell 6%, the Nasdaq Composite 5.5%, and the index closed more than 10% below its recent peak — a correction, in the technical sense. The Philadelphia Semiconductor Index dropped 12.2%, its worst session since March 2020. Bloomberg put the one-day loss of market value at a record $2.3 trillion, according to Bloomberg. The trigger it names is political: Democrats won both chambers of Congress on Tuesday, and voters in the same election restricted data center development almost everywhere it appeared on a local ballot.
That sequencing matters more than the percentage. The market did not reprice AI demand on Thursday. It repriced the permission to build for it — and those are different assets with different owners, which is why the damage was so uneven.
What the ballots actually decided
The cleanest count comes from Redbud Consulting, a Washington polling and data science firm, which tracked 38 local ballot contests with complete counts and found voters siding against data center development in all but one, per Fortune's reporting. Other trackers have published different totals — 40 measures with 39 restrictions, in one count — because they disagree about whether advisory questions and tax-break conditions belong in the tally. None of them disagree about the direction.
The specifics are what make the number real rather than symbolic. In Piqua, Ohio, a city of about 20,000, residents recalled the mayor and two city commissioners who had backed a Meta project budgeted at $1 billion and up to 1.4 gigawatts. At least 14 Ohio municipalities passed bans or restrictions. In Texas — more than 400 data center facilities, with many more under construction, per the American Edge Project and Technology Councils of North America — Democrat Gina Hinojosa unseated Governor Greg Abbott after campaigning on a moratorium until the legislature writes new protections.
None of those votes cancels a signed land contract. All of them change what the next contract costs, how long it takes, and whether it gets signed at all.
We mapped this exposure before the vote, when the moratorium question was already live in 16 governor races — data center moratoriums are on the ballot in 16 governor races — and the useful thing about Tuesday is that it converted a campaign promise into a calendar. A county commission can now be voted out. A utility negotiation cannot.
Why a zoning vote repriced $2.3 trillion
Semiconductor and data center equities trade on terminal value, not next quarter. A chipmaker at a rich multiple is a claim on a decade of buildout, discounted back; when the market raises its estimate of the probability that the decade arrives on schedule, the discount does the damage, not the forecast of this year's revenue. That is the mechanical reason a local land-use vote can move more market value in a day than most earnings misses.
Dan Ives at Yorkville Ives made the transmission explicit in a note Thursday: he now expects the tech industry's capex growth to peak next year rather than in 2028 or 2029. His numbers are worth keeping straight, because they are narrower than the headlines. He still expects hyperscaler capex to "easily exceed $1 trillion in 2027." What changed is the shape after that, and the pipeline: he estimates 10% of the 800 to 1,000 AI-focused data centers planned in the US "will now be wiped off the roadmap," with up to another 20% facing "a much longer regulatory road," all per Fortune's account of the note.
Read that as a delay curve rather than a demand cut. A project pushed from 2028 to 2030 is not a lost order for a transformer factory, but it is a lost year of compound growth for anyone who borrowed against the 2028 date. And the borrowing is the part the equity market is slower to see.
The dispersion is the story, not the average
Within one session, the market sorted itself by how much each company depends on building new capacity versus running what it already has. Google fell 1.6%. Meta fell 3.4% and Microsoft 2.8%. Nvidia fell 8%. Micron and AMD each fell about 14%. CoreWeave fell 9%, extending a two-day decline to 18%. Those are Fortune's intraday figures from Thursday, and the ordering is the analysis: the further a balance sheet sits from an operating campus and a search-advertising business, the harder it was hit.
That ordering is consistent with something the credit market already knew. In September, roughly $18 billion of debt tied to Oracle's New Mexico campus slid into stressed territory, as the Financial Times reported, with the gas pipeline meant to power the site slipping to 2027 — the episode we covered in Oracle's $18 billion New Mexico data center debt turns stressed. Stressed paper is a forward-looking instrument. It was pricing pipeline risk before a single ballot was counted, and Thursday's equity move was the slower asset class catching up to it.
This is why "AI bubble pops" is the wrong frame. Bubbles pop when demand disappears. What broke on Thursday was a financing assumption: that political permission was free, and that construction would arrive before the debt came due.
The offshore hedge is already being discussed out loud
The most consequential sentence of the week came from Meta's president, Dina Powell McCormick, speaking at a Wall Street Journal conference on Wednesday. Building data centers in the US is "going to be harder," she said, and then: "I can also tell you that every hyperscaler is now thinking about 'Do we need to build outside the United States?' Because they're enormous investments, hedging those investments is a wise thing to do." Fortune quoted both lines.
Take that seriously and the geography of the buildout starts to move without the volume changing much. Capacity that would have gone to a greenfield site in Ohio goes to a brownfield expansion that already has neighbours and a permit history, or to the Gulf, or to Southeast Asia. Same GPUs, same memory, different substation. It punishes anyone whose business model was site acquisition and local politics, and it rewards anyone who already owns powered shells — which is the argument this publication made in the days before the vote, and the reason the dispersion above is not random.
The Asian session on Friday is where the read-through gets over-read. SK Hynix fell more than 8%, Samsung 7%, LG Innotek 10%, Advantest 12.2%, Tokyo Electron 7.3% and SoftBank 11.6%, CNBC reported. Those are the correct companies to sell if US capacity stops being built — memory and test equipment are the most levered to volume. But they are also the correct companies to sell on any risk-off morning in a market where the same five names are held by every momentum fund. Correlation is not a capacity forecast. Watch the memory contract prices over the next two quarters; if HBM allocations for 2027 get re-cut, that is a demand signal and not a sentiment one.
What the skeptics get right, and what they miss
The bullish case is not stupid. Thursday's move was a multiple compression, not an earnings revision, and the earnings have not moved. OpenAI spent the same week publishing a community compact that promises to cover full electricity and grid costs, which is what a company does when it intends to keep building and needs the politics to stop being an obstacle. Trump's response, per Axios, was to tell voters to blame themselves — a signal that the fight over permitting is about to become a federal-versus-state argument, and federal preemption of local zoning is the single fastest way to reverse Thursday's repricing.
There is also a legal runway. Several measures are advisory, several will be litigated, and a recall in a city of 20,000 does not bind a state legislature. Anyone short the sector on the assumption that 38 local votes are permanent is making a bet about courts and legislatures, not about ballots.
The bear case that survives all of that is narrower and harder to dismiss: the buildout is now financed on the assumption that credit keeps arriving faster than construction slips, and Thursday told lenders that the slip rate just went up for political reasons they cannot underwrite. Equity holders can absorb a year of delay. A project finance stack with a 2029 maturity cannot. That is the channel through which a moratorium in Ohio becomes a margin call in New York, and it is why the neoclouds fell twice as hard as the hyperscalers.
What to watch
Four things, in order of how fast they will print. Friday's October jobs report is the immediate test of whether Thursday was a tech-specific de-rating or the start of a broader risk event. Then the Q4 earnings calls in the next few weeks, where the only question that matters is whether hyperscalers re-route capacity or re-price it — a re-route keeps the 2027 capex number intact, a re-price does not. Then January, when the Texas legislature either writes Hinojosa's moratorium into law or waters it down, which decides whether the largest data center state in the country has a permitting regime or a campaign slogan. And underneath all of it, credit spreads on data center debt, which will tell you what the lenders think months before the equity market admits it.
The AI trade did not die on Thursday. It got a new line item — political risk — and the market spent $2.3 trillion discovering that nobody had put a price on it.
If a county can veto a data center, who should carry the cost of that veto — the developer, the utility, or the ratepayer? Write to us via our contact page.
