The Backlash Codifies: California Signs the Nation's Toughest Data-Center Rules

Capability & Integration

• xAI shipped Grok 4.7 on September 21: a 2.1-trillion-parameter model, up from 1.5 trillion in Grok 4.6, built on a longer reinforcement-learning run and priced the same as its predecessor. xAI calls it the model’s “strongest safety guardrails yet” without publishing comparative eval results. It’s already live in GitHub Copilot, per SiliconANGLE.
• Anthropic released 36 optimization kits for more than 30 open-source biomolecular modeling tools on September 17, reporting roughly 4x average speedups and a low-memory mode that lets a single GPU predict structures beyond 10,000 tokens — infrastructure work that expands what outside labs can run rather than a new model, per Anthropic.

Robotics

• UBTech began home deliveries of its UWorld U1 companion robot in China on September 16 — a real but narrow deployment: 13,000-plus preorders from $16,500 to roughly $146,000, positioned as an emotional-AI companion rather than a labor substitute, per TechNode.

  Review

  “$16,500 to roughly $146,000” — the top-end figure looks off; independent pricing trackers and coverage of the same launch consistently put the flagship U1 Ultra at roughly $135,000–$136,500 (990,000 yuan), not $146,000 — suggested action: verify against the TechNode source and correct the ceiling figure.

• Global humanoid shipments hit roughly 19,100 units in H1 2026, up 272% year over year — but Chinese vendors took 97% of that volume, and industrial or commercial use now accounts for more than 70% of shipments, up from about half a year earlier. “Deployment” still mostly means factory floors, not homes or offices, per TechNode.

Hardware & Supply Chain

• Update: Nvidia’s fiscal third-quarter guidance still assumes zero data-center chip revenue from China, and USTR Jamieson Greer said after Sunday’s prep meeting that semiconductor export controls are not on this week’s Trump-Xi summit agenda — even though roughly ten Chinese firms hold approval to buy H200 chips and none have shipped since the deal was authorized last December, per The Motley Fool.

• Micron is reportedly aiming to roughly double its HBM production capacity to about 100,000 wafers a month by year-end, narrowing the gap with Samsung and SK Hynix as it ramps 12-high HBM4 output for Nvidia’s Vera Rubin platform, per DigiTimes.

  Unverified

  The 100,000-wafer figure traces to industry sourcing; Micron has not officially confirmed it as a production target.

Environmental & Cultural Impact

• Update: Governor Newsom signed seven data-center bills on September 21 — more than the three pending as of yesterday — covering electricity cost allocation, mandatory disclosure of water and power use, and local environmental review, calling it the most comprehensive state-level data-center oversight package in the country, per Governor of California.

• The legislation follows public sentiment, not the other way around: a Gallup poll released May 13 found 71% of Americans oppose building an AI data center near their home, with 48% strongly opposed — compared with 53% opposed to a nuclear plant in the same survey, per Gallup.

  Review

  “Gallup poll released May 13” — the release/survey-window date is inconsistent across available reporting on this same poll (survey windows and publish dates found in research range from March through July); the 71%/48%/53% figures themselves check out, but the May 13 release date could not be confirmed — suggested action: verify the exact publish date against the Gallup link before publishing as-is.

AI in the Wild

AI-generated images showing President Trump kissing White House aide Natalie Harp on a golf course spread across social platforms this week, racking up millions of views before the Washington Post confirmed on September 19 that both circulating versions were fabricated — one carried a visible OpenAI-tool watermark, the other showed the extra-finger distortion detection tools have flagged for years. The White House called it “clearly a fake photo,” per Washington Post. What its spread reveals: the tells that are supposed to catch AI fakes were sitting in plain sight and stopped nothing.

Takeaway