The AI selloff priced permission, not demand

On October 29, Morgan Stanley and its lenders pulled a planned high-yield bond sale meant to refinance a $15 billion loan for an Anthropic-linked data center in Hubbard, Texas. The problem wasn't the borrower's numbers. It was the ground. Governor Greg Abbott's hold on new data-center approvals had frozen the site's expansion, and the banks were left carrying the debt they had meant to offload ^[1]. Permission is the plain word for the approvals that turn announced spending into poured concrete, and all year both parties had been tightening it. New York and Texas announced data-center construction halts in mid-September, when analysts still read them as a temporary setback and expected chip and automation demand to hold ^[2]. Abbott paused roughly 1,800 projects. Pennsylvania and Massachusetts banned the nondisclosure agreements developers had used to keep their sites quiet, and Massachusetts Governor Maura Healey put the new rule in plain words ^[3].

Unless a community says yes to a data center, we’re saying no. — Maura Healey

The holds spread to the local level through the fall. Memphis paused data-center permits for 135 days; Salem, Oregon froze a project for four months; Pittsburg, California for forty-five ^[4]. The geography crosses both parties and runs against a public that held steady all year at 65 to 70 percent opposed to data centers in their own communities ^[5]. Before the votes, the credit file had already found the crack on its own. The week of October 20, Apollo's chief economist Torsten Slok warned the boom carries a core economic flaw — the cost of producing compute exceeds what it returns to users — and put the payoff outside the largest cloud giants at zero ^[6]. CoreWeave, a neocloud — a company that borrows to buy chips and rents out the computing power — had by October 26 lost roughly 60 percent from its May high on $35.1 billion of debt, with bond yields already at 15 percent ^[7]. The next day Nvidia's Jensen Huang defended CoreWeave by claiming its cost of capital was falling; within ten days its 2032 bond traded above a 16 percent yield and its credit default swaps — insurance that pays out if it defaults — spiked ^[8]. On October 30, before a ballot was cast, Treasury Secretary Scott Bessent refused the industry's request for a federal liability shield ^[9].

It is humans who are responsible, not the AI. — Scott Bessent

The agenda the new majority inherits was drafted weeks before the vote. On September 11, four House Democrats pressed Speaker Johnson to cut recess short over what they called a "conflagration of risk," and the House instead stayed home until after the midterms ^[10]. Hakeem Jeffries's promise that regaining Congress would mean accountability dates to August, and the probes now being prepared — into the reported use of Elon Musk's Grok chatbot to plan the Venezuela operation, and into Meta, Apple, and Musk's companies — carry subpoena power ^[11][12]. The January list is drafted: corporate liability for AI-driven harm, mandatory independent safety audits of the most advanced models before release, and shields for consumers against electricity costs driven up by data centers ^[13]. The state template is already on the books — Illinois passed the first mandate for independent third-party audits in May, California standardized evaluators in October, both closing the gap the White House's September accord left when it promised "independent" audits and never defined the word ^[14][15][16]. The market priced that agenda in one motion. CoreWeave's 2032 bond crossed 16 percent with its default swaps spiking, and Anthropic's IPO talks dropped roughly 40 percent — from a $1.5-to-$2 trillion range toward $1-to-$1.3 trillion — against $895.4 billion of aggregate compute-purchase commitments and a $23.3 billion quarterly loss ^[8][17]. Hedge funds shorted the builders that depend on raising outside capital — CoreWeave, Nebius — while long-only money bought the large cloud giants, Amazon, Microsoft, and Alphabet ^[18]. The dividing line is the balance sheet, not the product. The cash-rich hyperscalers — the cloud giants that own their own chips ^[19] — haven't cut a dollar: Nvidia's own projections put their spending at $800 billion this year and $1.3 trillion in 2027 ^[20][21], and Amazon's Andy Jassy says AWS compute is nearly sold out through 2027 ^[17]. Even the debt market hasn't shut: Goldman Sachs counts $67 billion in AI-related bond deals priced since October 29, even as spreads — the extra interest over government debt — widened ^[8]. ING's analyst holds the other side plainly: demand fundamentals are intact, the S&P 500's 2027–2029 earnings potential hasn't been derailed, and the right move is to wait until January 1, when the administration's approach and its new regulatory appointments become clear ^[22]. A separate cost belongs to rates, not politics: markets expect the Federal Reserve under Chair Kevin Walsh to keep hiking, which raises every borrower's bill regardless of what Congress writes ^[2]. January will decide the agenda's fate and the spread's direction — whether the liability law passes, stalls, or dies, and whether the extra yield gives back ground. None of that will rewrite what this week's prices disclosed. Lenders had been reading all year what the voters had been living: the approvals behind every new site, and the neighbors, 65 to 70 percent against, who can hold those approvals up. A cheaper price in January would not unprice it. What the credit file says is not on the ballot.