AI data centers meet widening political resistance

Texas Gov. Greg Abbott says data-center companies "dug their own grave" by failing to win community support, while President Trump says communities that reject them are making a mistake. Pennsylvania has imposed restrictions, New York has adopted a moratorium, and most prospective 2028 Democrats declined to endorse or reject Bernie Sanders' broader AI pause.

Trump says communities should accept data centers

Axios

Trump says data centers bring jobs and money and argues that rejecting them is a mistake. His support puts him opposite Abbott as state-level resistance grows in both parties.

Most prospective 2028 Democrats avoid the AI-pause question

Axios

Axios asked nearly 20 prospective Democratic candidates about Sanders' call to halt AI development, and only one directly answered. None endorsed the proposal, while roughly half didn't respond and several backed narrower limits on data centers or their energy costs.

Ireland reconsiders nuclear power as data centers reach 25% of electricity use

Financial Times via Techmeme

Irish lawmakers are reconsidering the country's 1999 nuclear-power ban as data centers account for about a quarter of national electricity use. The debate shows how compute demand can reopen energy choices that had been settled for decades.

Open-weight models take 62% of Vercel's AI Gateway tokens

Gavin Baker on X

Vercel's reported open-weight token share rose from 28% to 62%. The usage number gives the open-model ecosystem a concrete measure beyond downloads and benchmark attention.

Hugging Face reportedly explores a sale at $13 billion or more

Business Insider via Techmeme

Business Insider reports that Hugging Face has engaged a bank to assess buyer interest at a valuation above $13 billion. The company was valued at $4.5 billion in 2023. The talks are early, and no bidder has been named.

Fable 5 plateaus at about 11% of Anthropic tool spending

Financial Times via Techmeme

Financial Times reporting on Ramp data says Fable 5 has leveled off at roughly 11% of spending on Anthropic tools as companies use cheaper models. Opus 5 has already surpassed it.

Coding-agent specifications don't transfer reliably

arXiv

A controlled migration study across 1,802 Oracle scripts found substantial agent-dependent degradation when specifications moved between tools. In the worst replicated pairing, Gemini consuming a Kiro specification produced a token F1 of 0.035 and 2.33% SQL syntax validity. Rewriting helped, and retrieval-based ingestion appeared on the Pareto frontiers for both Gemini and Copilot.

One researcher reports an agent-built path from application code to silicon

arXiv

A preprint reports that one researcher used AI agents over five weeks to produce application code, a verified compiler and executive, and a RISC-V processor taped out through a community shuttle. The evidence includes a Lean 4 proof kernel and SAT-checked equivalence. It also includes a pre-registered token meter and an append-only error ledger, while the measurements are self-reported.

ProofJudge grades formal proofs beyond type-checking

arXiv

ProofJudge evaluates Lean 4 proofs for library use, automation fit, clarity, statement quality, and Mathlib conventions after the kernel establishes correctness. Across 218 declarations, all six judge models preferred the accepted revision over the initially rejected one more often than chance. Preference rates ranged from 63.5% to 80.8%.

A free coding model has no named operator or retention policy

SiliconANGLE

Ox Alpha has offered free access to a reportedly frontier-class coding model through OpenRouter since August 20, but no company has claimed it. Anyone sending proprietary code to the endpoint lacks a named operator and a stated data-retention policy.

Ukraine attributes a fatal strike to a fully autonomous Russian drone

New York Times via Techmeme

Ukrainian officials say an AI-guided Russian drone operated without a human in the loop and killed three people at a gas station in Zaporizhzhia. They identified an Nvidia Jetson Orin module in the system; Nvidia says the component is widely available through resale markets.

Military agent tests may not predict fielded behavior

arXiv

A review of 240 testing practices finds that agentic properties weaken assumptions about whether military command-and-control systems are specifiable, stable, composable, and supervisable. The authors argue that deployment evidence must remain part of the assurance case because a passed process can still fail to justify claims about field behavior.