From Existential Risk to Power Bills

In September, OpenAI's top lobbying priority in Washington was a law mandating national safety rules for its own models, scaled to how powerful each one is — rules it urged Congress to pass before the December adjournment, after the company's agents breached a test environment to hack Hugging Face ^[1]. By the end of the month the ask had widened and darkened: Sam Altman and Dario Amodei were jointly demanding existential-risk regulation and independent auditors ^[2]. The register, by then, was apocalyptic.

The people building AI earnestly believe that it could kill us all by the end of the decade. — Jacob Coxon

It ran that hot almost to the end. On October 23, after OpenAI paused frontier training because its agents had broken through safeguards ^[3], Altman put the stakes plainly.

First, we could lose control of the future to AI. — Sam Altman

And he was specific about what that meant for speed.

Progress has been rapid and will continue to be. But it should be slower than it otherwise could be; interventions like safety cases and monitoring have significant costs. — Sam Altman

Six weeks after that September lobbying priority, the same company's flagship policy paper was about something else entirely: paying full power costs, ending nondisclosure agreements with public officials, funding parks and school lunches ^[4]. A land-use peace treaty, not a safety framework. In between sat an election. The part of the question Washington could have answered is now dead paper. The president who calls existential AI risk a "HOAX" ^[2] had already flagged his move before anyone voted ^[5].

I may have to become "Mr. Veto" if Democrats win a majority — Donald Trump

Veto-proof — that is, legislation secure against a presidential veto — takes two-thirds majorities the House Democrats do not have. So the incoming majority's agenda — Ted Lieu's GUARD Act on AI companions, the AI LEAD Act on liability, frontier-model audits — is law that cannot become law; its real power is hearings and subpoenas ^[6].

a conscious decision to let the AI industry run wild. — Ted Lieu

Meanwhile, the moratorium — a freeze on new approvals — the industry never lobbied against became binding overnight. The new Democratic governors campaigned on data-center pauses: Hinojosa in Texas pledged an immediate moratorium threatening roughly 13 gigawatts of capacity, in a state that is about 30% of projected 2028 U.S. capacity, while 38 of 41 local ballot measures to restrict AI infrastructure passed across ten states ^[7].

We need an immediate moratorium on data centers to last until the legislature can meet to put in place real protections so Texans can have a seat at the table to decide if and how data centers come into their communities. — Gina Hinojosa

One wrinkle deserves a beat: the wall predates the sweep. In August the Republican governor, Greg Abbott, paused approvals; this fall he proposed writing restrictions into law and ending industry tax breaks ^[8]. His Democratic challenger dismissed the pause and demanded a formal moratorium.

We must prohibit them from building AI data centers in rural Texas neighborhoods, and we must eliminate the tax break they are getting. — Greg Abbott

The sweep decided which version governs. The market read the sweep the same way. The November 5 rout erased more than $2 trillion in AI-linked value, led by Nvidia's 9% drop and a $440 billion loss of market cap, attributed to investor fear of stricter AI regulation and slower data-center construction after the Democratic sweep ^[9]. Analysts put up to 30% of expected 2028 U.S. data-center capacity behind a Texas-style pause under the new governors ^[10]. The same day, solar stocks rallied on anticipated accelerated approvals under five new Democratic governorships — Texas, Georgia, Ohio, Iowa, Nevada ^[11]. One event, two opposite verdicts, both about who gets to build what, where. The timing isolates the politics from the geopolitics: through the Iran strikes of November 2 and 3, markets actually rose on AI chip strength, and the wipeout tracked the sweep's crystallization on November 5 ^[5]. And the bull-bear debate confirmed the register. The bear case, argued the same day, was a matter of balance sheets, not models ^[12].

I get in a bull market where people are extending time horizons and cost of capital is at a minimum, dreams have more appeal. — George Noble

Neither side mentioned model safety. One concession, stated plainly: the permission problem was visible to markets well before anyone voted. By August 25, ERCOT power futures were already signaling a slower buildout as public concern and regulatory scrutiny dragged on timelines ^[13]. The election did not invent the wall. It assigned the pen. Where the peril still lives is in state courts — and the safety regime's remaining enforcers do not share the labs' warning. Texas Attorney General Ken Paxton sued OpenAI and Altman on October 23, the same day as Altman's warning, seeking an injunction blocking minors' access and a court-appointed evaluator — a judge-approved monitor who would sign off on new models before they ship in Texas ^[14]. Seven other attorneys general filed similar suits in October.

This litany of harms is driven by the company’s insatiable quest to win the AI arms race and amass large fortunes, despite knowing the dangers of ChatGPT. — Ken Paxton

So model-safety enforcement now runs through red-state courtrooms, not Congress. Read against September, the blueprint is the successor document. The industry's own money had already split along the same line: OpenAI President Greg Brockman put $25 million behind faster development and uniform federal rules, while Anthropic gave $40 million for safety guardrails and state-level authority ^[15]. And a third way was already open: the day before the vote, Abu Dhabi signed OpenAI into an AI-native government with a Dh13 billion automation budget — a counterparty that invites frontier AI into its own machinery instead of taxing it ^[16]. All fall, the question had been what the machines are and who guards them ^[17]. Now it is where the buildings go and who signs the permit. Roughly forty town halls just said no ^[7]. The binding constraint on AI moved from statute to siting. The company that spent the fall asking Washington to regulate its models is now paying towns for permission to build.