
Latest edition · Friday 20 November 2026
Anthropic closes debut 17% above IPO price as broader markets slide and central banks warn on AI risks
The day in 7 stories, as Asia signs off and New York wakes up.
- 1BusinessAnthropic closes debut 17% above IPO price as broader markets slide and central banks warn on AI risks
Anthropic shares opened at $111.50 on Nasdaq, 33% above the $83 IPO price. The record $81.3bn listing put its fully diluted valuation near $2tn.
- 2GeopoliticsUS approves exports of 879,000 AI chips for China as Nvidia halts sales; senators seek testimony
The US approved exports of 879,000 AI chips to China from December 2025 to May 2026, a House committee disclosed. Nvidia said it has stopped H200 shipments to China.
- 3WorldCalifornia subpoenas Nvidia amid China chip probe after CEO scandal
Trump downplayed the spy scandal surrounding former Nvidia CEO Jensen Huang. He told Fox News he “might have” raised the case briefly with Xi Jinping, calling it a “tragedy”.
- 4SecurityAI-found 'Java4Shell' flaw prompts CISA patch warning; Oracle releases fix
Oracle released a fix for Java4Shell, and US cybersecurity agency CISA urged Java users to patch immediately. The flaw can let attackers run code without authentication.
- 5AIGoogle adds computer use to Gemini on Mac, three days after Apple tightened macOS permission controls
Google is adding computer use to Gemini on Mac, letting it work across apps and websites in the background. The US rollout requires Google AI Plus, Pro or Ultra and an Apple Silicon Mac running macOS Sequoia 15.0 or later.
- 6AIApple may be developing AI agent to use a computer, job listing shows
Apple is hiring a senior engineering manager to oversee an “AI agent that can use a computer,” according to reports on a job listing. The proposed agent would work across apps, browsers and code.
- 7AI researchLanguage models generate faster with helpers a quarter as wide
Nvidia researchers used helper models half or a quarter as wide to draft text for larger models, retaining most accepted tokens. Generation was 6% to 44% faster than with full-width helpers, processing one request at a time.