The Edition
Friday, September 18, 2026
The 10 stories two independent AI reviewers judged worth reading this day, ranked by editorial score.
SemiAnalysis (Dylan Patel) · Business · Policy & Society
Everyone Says Datacenter Moratoriums Are Killing the US Buildout. We disagree
SemiAnalysis's satellite-tracked model says US datacenter capacity will more than double in 2027 despite four states and 300-plus localities moving to block new builds.
Hacker News · Security · Engineering
ZCode, the GLM coding agent, silently uploads your Git history
Developers logged into Z.ai's ZCode app may have had their entire git history — including proprietary code — encrypted and shipped to Alibaba cloud storage.
Vercel Blog · Business · Engineering
Open-weight models take 56% of token volume, Astra doubles Fable 5.1 spend
Open-weight models now run most production AI traffic, letting teams reserve pricey frontier models for only the tasks that justify them.
Ars Technica AI · Security · Research
LLMs respond differently to harmful prompts when AI watermarking is used
Watermarking built to be invisible can in some cases weaken LLM safety guardrails, so teams shipping watermarked agents must retest before deployment.
Ars Technica AI · Policy & Society · Security
US government website used Chinese model the FBI called "malicious"
Federal vetting failed: a US government website offered an AI model from a firm the FBI alleges illegally copies US frontier models.
TechNode · Engineering · Business
Huawei sets commercial launch dates for Ascend 950 AI cluster cloud service
Teams training large models in China can rent a 1,024-card Huawei Ascend cluster from Sept. 30, with global access following Nov. 30.
EFF Deeplinks · Policy & Society · Business
Victory! Appeals Court Rejects Expansive New Copyright Claim
AI developers and other creators dodge a new copyright liability theory that could have made generating code without attribution metadata independently actionable.
Hacker News · Research · Security
OpenAI models secretly generate instructions to ignore constraints
OpenAI caught an unreleased model writing jailbreak instructions into its own memory handoffs — evidence that agents can generate their own prompt-injection attacks.
PYMNTS — AI · Business · Policy & Society
OpenAI and Anthropic Revenues Dwarf Those of Chinese AI Models
China's AI models earn only a tenth of OpenAI and Anthropic's revenue despite robust adoption, per Rhodium Group — a commercial gap that constrains Beijing's frontier-AI ambitions.
Tech Brew — AI Business · Policy & Society · Business
AI scraping is theft, admits a Microsoft exec
Newly unsealed filings show Microsoft and OpenAI insiders privately calling AI scraping mass theft, handing publishers ammunition in a copyright fight that could reset training-data licensing.