Garrison Lovely argues anti-AI populism will shape 2028 after local datacenter measures go against the industry in 40 of 41 contests

This is a policy story. No model or version was released. The concrete capability limits are in the Sanders–Casar bill:

• a permanent ban on superintelligence
• federal approval before advanced systems can be deployed
• bans on models that help design nuclear or biological weapons, evade human oversight, or deceive people
• prison terms of up to 20 years for some violations (Boston.com)

Even with a Democratic Senate, Democrats lack the 60 votes needed to pass AI regulation. Maria Cantwell is pushing for federal testing of models (Tech Times).

• Rules based on what models can do: The bill targets specific behaviors (weapons uplift, escaping oversight, deception) rather than training compute. Compliance would therefore rest on evaluations of dangerous capabilities and of whether models can be controlled. The reviewed sources do not say how those evaluations would be defined.
• Exposure to physical capacity: Cleanview puts 158.6 GW, or 41% of the planned US datacenter pipeline, in states whose governors have imposed or promised pauses. SemiAnalysis gives about 30 GW as a ceiling for 2028 capacity at risk. Ballot measures directly blocked only about 1.6–1.9 GW (Luminix).
• Automated AI research: Lovely says labs plan to automate AI research and AI safety research. That is his characterization, not a documented lab plan in the reviewed sources.

• AI's weight with voters: 23% of voters called AI and datacenters crucial and 33% a major factor (Unite.AI). That still ranked last of five issues (The Decoder).
• Which party is better on AI: Lovely says 30% chose Democrats. Unite.AI reports 29% Democrats, 27% Republicans and 43% neither. Lovely's 67–30 Democratic edge among "crucial" voters was not corroborated in the reviewed sources.
• Musk's spending: reported as $474M (CNBC) and $446M (TIME).
• Altman quote: the linked clip shows Altman saying AI will "most likely sort of lead to the end of the world" (00:00). The captions are automatic and may contain errors.

• The 99–1 Senate vote (Florida Phoenix) was bipartisan, which complicates Lovely's framing that "Senate Republicans wisely" stripped the measure.
• Leading the Future, funded partly by $25M each from Marc Andreessen and Ben Horowitz, wants one national rulebook that overrides state laws (TNW).
• Anthropic gave more than $40M to Public First Action, a PAC on the pro-safety side (TIME). The industry's political money is therefore split, not one bloc.
• Lovely's book Obsolete argues for stopping AGI development.

• Super PAC losses vs. wins: Lovely highlights the loss of Republican Darline Graham to Annie Andrews in South Carolina. Fast Company shows Leading the Future winning more often than the crypto PAC Fairshake (86% vs 77%) (Yahoo).
• Graham spending figures differ: $4.8M (TNW) vs $2.3M (Fast Company).
• Measure counts differ: Lovely says more than 30 localities in nine states; Ballotpedia-based counts say 41 measures in 10 states.
• Undocumented incidents: Lovely's references to summer "rogue AI hacking incidents" and criminal conduct by AI agents are not documented in the reviewed sources.
• Genre: the piece is an opinion column by an advocate, not neutral reporting.

• Siting risk is real: 76% of voters oppose a local datacenter, and governors-elect in Texas and Georgia have promised pauses. Capacity plans for 2027–28 need political contingencies.
• Preemption looks unlikely: state preemption failed 99–1, and Democrats now control Congress, so a single national rulebook is improbable soon.
• Hearings are coming: expect subpoena-backed oversight and pushes for federal model testing, and prepare documentation of dangerous-capability evaluations.
• Political money has limits: big super PAC spending did not guarantee wins this cycle, though it was not uniformly punished either.