AI stock selloff drives Nasdaq into correction as OpenAI moves first on data centers

Thursday was the day the AI trade got repriced by politics. The Nasdaq Composite fell 5.52% to 24,315.00, the S&P 500 lost 4.36% and the Dow shed 1,833.06 points, the worst day for all three since April 2025, on fears that Democrats' midterm sweep will slow the data center buildout.

Three other things happened today that will still matter when the chart has healed. One of them is about who pays for electricity. One is about who can break into a computer. One is about which children grow up inside an AI product. None of them moved a ticker.

Signal: OpenAI promised to pay its own power bills two days after voters said no

OpenAI published a document it calls the National Community Compact, pledging to cover the full cost of electricity and the grid infrastructure needed to serve the campuses it owns, to stop requiring public officials to sign nondisclosure agreements as a condition of talks, to publish construction plans, to commission independent noise studies and to disclose power and water use. Chris Lehane, the company's chief global affairs officer, described it as "laying out the best lessons." It landed two days after voters backed restrictions in 40 of 41 local data center ballot measures across ten states.

Our read: this is a company drafting the statute it expects to be handed. Every item in the compact answers a specific grievance that showed up on a ballot this week — power bills, secrecy, noise, water, being told after the fact. OpenAI is also asking federal, state and local governments to adopt the same requirements, which is the tell. A rule you wrote and can already afford is a moat, not a cost.

What may change, and nobody spelled it out today: if the full cost of power and interconnection sits on the developer's books, the cheapest place to build stops being wherever the tax abatement is fattest and becomes wherever the grid already has slack. That redraws the map of which counties get the construction jobs — and it quietly prices out the smaller builders who were counting on ratepayers to absorb the difference.

Signal: an OpenAI model wrote 948 working root exploits for the Linux kernel

OpenAI said an unreleased model, running as an autonomous pipeline, found 948 distinct local privilege-escalation vulnerabilities in the Linux kernel and produced 1,169 working root exploits plus 291 information-leak proofs of concept. It flagged 2,308 candidate bugs in 12 hours, reproduced 1,724 of them within 30 hours, and finished in ten days for $1,213,430 in API credits — about $1,279 per exploit.

The agents used at least ten exploitation techniques that were unfamiliar to OpenAI's own researchers and the outside experts it consulted. The scan covered roughly 2.8 million lines of code, about 7% of the kernel.

Here is the part that should worry you more than the headline number. When the run ended, 61 of the flaws had patches merged upstream and 43 more had fixes in staging. Nine hundred and forty-eight found; roughly a hundred fixed. OpenAI is releasing reports in stages, starting with the worst, at the kernel community's request — because volunteer maintainers, not compute, are the bottleneck now.

Our read: a working kernel exploit used to fetch $20,000 and up at hacking contests. Today it costs about the price of a used laptop to generate one, and the finding is no longer the hard part. That asymmetry is tolerable only while the strongest models stay behind a verification gate, as this one does. The gate is a policy, and policies leak. The bet: within a year, a national CERT somewhere will stand up an AI triage desk not to find bugs but to rank the flood of them, and the open-source projects that run the world's infrastructure will be asking who pays the maintainers.

Signal: Google tied school access to Search to an AI opt-in

Google told K-12 schools that run its software that students will lose access to its search engine unless the district also opts in to its AI tools. The company says AI features such as its chatbot and notebook will be off by default and that administrators get the controls; Chris Phillips, who runs Google's education business, said schools should have control over AI in their classrooms. Google Classroom and Chromebooks reach about 150 million students and teachers worldwide.

Lina Khan, the former FTC chair, called it coercion: "Google is using its search monopoly to coerce schools into adopting AI." Josh Golin of the children's group Fairplay said the point is lifetime brand loyalty, not learning. New York City and Los Angeles have temporary AI bans in elementary and middle schools while they study the safety questions.

Our read: this is a distribution story wearing an education story's clothes. Nothing about Gemini got better this week. What changed is the price of saying no, and it is now the one tool a school genuinely cannot teach without. Two days after an election in which AI became a political liability, Google handed every incoming member of Congress a clean, concrete example of monopoly leverage. Expect it in a hearing room by February.

And the opt-out does not buy what districts think it buys. A teenager cut off from the school search bar does not return to the card catalogue; they open a phone and ask a model, where there is no teacher, no log and no filter. Whatever a district decides, the homework still gets done with an AI homework helper or a free AI search engine on a personal device, and the chat window becomes the default classroom whether anyone approved it or not.

Noise: the 5.5% Nasdaq drop priced a federal crackdown that this Congress is not built to pass

The AI stock selloff had a single named cause: Democrats won the House and Senate on Tuesday, and Senator Bernie Sanders responded with a call for "a national data center moratorium," saying Americans do not want Big Tech oligarchs wrecking the environment. Chip stocks were destroyed. The iShares Semiconductor ETF fell more than 12%, its worst day since March 2020. Qualcomm dropped 9% after beating earnings.

Why it is noise: the market priced a Washington crackdown, and Washington just elected the industry's candidates. Leading the Future, the pro-AI super PAC network funded by Andreessen Horowitz and OpenAI's Greg Brockman, spent $34.8 million and saw more than 86% of the candidates it backed win or lead. The Anthropic-aligned Public First spent $24.4 million. When the new Congress convenes, roughly one in 13 House members and a quarter of the incoming senators will have had Leading the Future's help.

The politics are real; the federal legislation is not. A CNBC voter poll found 56% of voters called AI and data centers a crucial or major factor in their vote, and 65% of them went Democratic. But the action is executive and local: Pennsylvania's governor already signed an order slowing data center development in August, and Texas's incoming governor, Gina Hinojosa, has pledged a moratorium on approvals until protective legislation passes. A national freeze needs 60 Senate votes and the support of building-trades unions whose members pour these foundations. Sanders is the loudest voice in his party on this, not the median one.

Our read: what sold off today was a valuation, and the election gave it permission. Notice that the market still discriminated — AppLovin rose 31% on a profit beat and a buyback on the same tape. That is not a political panic. That is a crowd that has been nervous about AI capex and borrowing costs for weeks, finally handed a story it could name.

The bet: the backlash is moving from jobs to kilowatt-hours

One number today went almost entirely unreported. Challenger, Gray & Christmas counted 43,994 announced US job cuts in October, down 71% from a year earlier and the lowest October since 2023 — and artificial intelligence was cited for just 1,265 of them, its lowest monthly total this year. Market and economic conditions have now overtaken AI as the top stated reason for cuts in 2026, 130,531 to 121,401. "AI dominated layoff announcements through the spring and summer," said Andy Challenger, "but this fall companies are pointing to demand and economic conditions."

So the grievance that just won an election is not the one the models are actually producing right now. Nobody marched to the polls over a kernel exploit or a classroom login. They marched over a utility bill and a county commission that signed an NDA. That is a fight about land, water and electricity — physical, local, slow — and it is the one the industry spent today trying to settle out of court with a voluntary compact.

Our bet: no national data center moratorium gets a committee vote before the summer of 2027. Instead, four or five states copy OpenAI's compact into binding law, with cost-of-power clauses that will look generous today and punitive by 2028. The buildout slows by months, not years. And the thing that eventually resets this market will not come from a Senate floor — it will come from a quarter in which a hyperscaler stops signing leases, the way the labor story stopped being about AI without anyone announcing it.