From November 7 to 8, money and rules tightened around AI at the same time: Anthropic is pushing ahead with a pre-Thanksgiving listing despite the selloff, Nvidia's H200 export checks were found wanting, and data centers face resistance after the midterms — while open-weight models closed in on closed ones in cyber offense, and Grok text started carrying a default watermark.
Anthropic pushes ahead with pre-Thanksgiving IPO
On November 7, Anthropic pressed ahead with its IPO despite the selloff: investors cited by the Financial Times say the company plans to list before Thanksgiving, raising $80–100 billion at a $1.25–1.75 trillion valuation — below the $2 trillion the most optimistic investors had hoped for, but above its current $965 billion. Annualized reported revenue is about $81 billion, and the roadshow starts next week.
Nvidia H200 export checks face questions
On November 7, a New York Times investigation found that the US Commerce Department cleared about two dozen Chinese companies to handle Nvidia H200 chips — including New Matrix, whose listed Beijing office was empty, with owners tied to Wise Road Capital, blacklisted by the US in 2024. Another approved firm, Maginfra, also had a vacant office and shares staff and patents with Inspur, which is on the Entity List. The reporting does not show chips reaching these firms.
Data centers hit resistance after the midterms
On November 7, TechCrunch reported that data-center construction faces resistance after the midterms: of 41 local ballot measures, 40 backed restrictions or rejected development. OpenAI answered with a compute campus policy blueprint that promises to cover its data centers' electricity costs and drops non-disclosure requirements, an attempt to ease tensions with host communities.
Open-weight models close in on closed ones in cyber offense
On November 7, Vals AI's CyberOffense benchmark showed open-weight models nearly matching closed ones at network attacks: OpenAI's restricted GPT-6 Sol Daybreak completed 80.8% of attack objectives, while the open-weight DeepSeek V4.1 Flash reached 72.8% at about $3 per run, far below the roughly $22 for its rival. Once weights are public, no provider can control how they are used.
Arete-27B: fine-tuning strips political bias from a model
On November 7, Nous Research released Arete-27B: fine-tuning Alibaba's Qwen3.8-27B on 53 million tokens cut propagandistic answers from 22% to under 1%, while general capabilities stayed roughly intact. In the same test, GLM-5.3 and DeepSeek V4.1 Flash gave propagandistic answers in 15% and 16% of cases.
Grok API text watermarking takes effect on November 7
From November 7, text from the Grok API carries an invisible watermark by default: it covers grok-4.7, 4.6, 4.5, 4.3 and three grok-4.20 variants, including structured outputs and tool-call arguments, and the x-watermark: true response header shows when it was applied. Team admins can switch it off per model, but must confirm a legal exception and accept responsibility for marking.
