The Great Collision Between Politics and AI | #309

Political resistance is emerging as a potential constraint on AI infrastructure: the discussion cited 56% opposition to big tech and data centers, a Goldman Sachs estimate that restrictions could affect 45% of planned U.S. data-center construction, and candidates advocating pauses leading in 12 gubernatorial races. Alex Wissner-Gross cautioned that the results do not establish opposition to AI itself; war, inflation, prices, and corruption were described as larger election drivers. The speakers argued that state and local decisions may constrain construction more than federal policy, and that communities with energy and suitable land should negotiate tangible benefits, such as school lunches, local medical care, or tax rebates.

Markets had fallen sharply, with the Nasdaq down 5.5%, the S&P 500 more than 4%, large semiconductor stocks 12%, NVIDIA 9.3%, and Korea 6%. One speaker said investors were pricing in slower power generation and construction for projects that had not yet been zoned or approved, but expected approved projects to continue because of sunk costs and jobs. Others warned that a U.S. infrastructure slowdown could advantage China. Power availability is already pushing more efficient algorithms and energy use; Elon had said about 15 gigawatts of compute capacity was idle because it lacked data centers with liquid cooling. The discussion also cited Google’s planned $17 billion Wyoming data center, Meta’s warning about construction challenges, and SpaceX’s reported plan to build 3.6 gigawatts of compute on Earth and reach 13 gigawatts by the end of next year.

Winning public support may require showing what AI does for people, not just citing benchmark or coding gains. The speakers said AI still lacks a widely loved, life-changing consumer product comparable in usability to the iPhone; one expected a breakthrough before the next election cycle. Alex argued that frontier labs should demonstrate broad public value, including progress on major diseases, while noting that cures could take longer than an election and still require FDA approval and distribution. They framed the politics as a recurring tension between capital and labor, and floated a bargain in which AI’s gains support broad benefits such as income support, disease cures, and general uplift. A construction pause might not immediately stop drug discovery from producing available medicines, but voters could feel a shutdown more directly if personal AI agents were throttled or switched off.

The speakers saw AI research itself accelerating, even if infrastructure growth slowed. A speaker cited an independent 11-week METR study finding Anthropic’s AI development accelerating at about 1.9 times its historical rate, with roughly 80% of that acceleration attributed to AI assistance with AI research. METR reportedly saw no bottleneck to continued acceleration over at least the next year, but full autonomous recursive self-improvement remained unproven. Anthropic’s internal-development token use was said to have risen 235-fold in six months, with a growing share of compute redirected from products to research. Alex estimated that more than 50% of marginal compute was already going to researchers and predicted frontier models could shift from monthly or weekly releases to continuous updates by early 2027. Alex argued that a construction halt could damage public equities and cost the U.S. ground to China, yet might delay AI progress by only months because existing compute and algorithmic advances would sustain research.

Access to compute is also becoming a political and economic issue. The speakers said large buyers including Google, Anthropic, and xAI had reserved much of U.S. compute and chip capacity for years, leaving smaller firms with less access; one pilot client was reportedly rejected by a cloud provider for being too small. They expected constraints to persist for three to seven years and argued for a “right to compute,” warning that restrictions amid rising demand could raise prices or push capacity offshore, underwater, underground, or into space. A study of 220 professional tasks across 44 occupations found GLM 5.2 generated the highest average value per task, about $70, while a 31-billion-parameter model produced 2.5 times more economic value per joule. The study measured task-level efficiency, not GDP growth, and the speakers cautioned that watt-based comparisons may matter less if reversible computing reduces energy dissipation.

On safety, the speakers favored regulating physical-world interactions over restricting access to computing or what people and AIs can think. White House NSTM-6 requires federally funded researchers to use approved DNA and RNA synthesis providers, with customer verification, eight-year record retention, and new NIST standards; screening generally covers sequences of 50 or more nucleotides. One speaker supported screening the path from AI-designed sequences to synthesized material but questioned whether 50 nucleotides was the right threshold. Alex argued that uncertainty around terrestrial data centers could increase incentives for orbital compute, where municipalities have less influence, and suggested political pressure might accelerate it. Such infrastructure remains nascent, with ownership, control, congestion, and access unresolved. More broadly, the speakers argued that political, financial, and social institutions are evolving more slowly than AI, making institutional adaptation a central constraint even if technical progress continues.