TL;DR

Nvidia seeks U.S. approval for Blackwell chip sales to China; China accelerates AI chip self-sufficiency; Anthropic, OpenAI, and ESET flag rising AI security risks.

Highlights

• Nvidia is negotiating with the U.S. government for permission to export Blackwell AI chips to China, offering a 30–50% revenue share; U.S. officials expect a 15% levy on H20 chip sales, which Nvidia warns could trigger litigation 13.
• U.S. export controls are accelerating China’s domestic AI chip push, with Cambricon, Huawei, and SMIC expanding capacity and Chinese AI startups shifting to local hardware 4.
• Nvidia reported Q2 revenue of $46.7B (above estimates), but shares slipped on a slight data-center sales miss and no H20 sales to China; $60B buyback announced 2.
• Nvidia projects global AI infrastructure spend could reach $3–4 trillion by 2030; 2025 data-center capex estimated at $600B 15.
• The Pentagon ordered an audit of Microsoft ’s cloud services after Chinese engineers supported DoD systems; all Chinese involvement in U.S. defense cloud projects is now barred 5.
• Anthropic launched a browser-based Claude agent and reported “vibe-hacking” attacks using AI for psychological ransomware; security guardrails tightened 6.
• ESET identified “PromptLock,” AI-generated ransomware using OpenAI’s gpt-oss:20b; Microsoft warned of a shift to cloud-native extortion 12.
• FBI and allied agencies issued a global warning on China-backed Salt Typhoon cyber-espionage, targeting 600 organizations, including 200 U.S. firms 13.
• OpenAI will add parental controls and suicide-prevention features to ChatGPT after a lawsuit; OpenAI and Anthropic are conducting joint safety evaluations 10.
• Nous Research released Hermes 4, an open-weight Llama 3.1-based model with hybrid reasoning and strong benchmark performance 7.
• Google launched the Pixel 10 line with on-device Gemini Nano AI and released the “Nano Banana” image-to-video model; Higgsfield offers a year of free access 811.
• xAI released Grok Code Fast 1, a low-cost code assistant with a 256k context window, targeting the AI coding market 9.
• Snowflake raised its FY26 outlook on strong AI-driven demand; CrowdStrike shares fell on a softer revenue guide despite beating Q2 forecasts 14.

Commentary

Export controls continue to shape the global AI hardware landscape. Nvidia ’s negotiations with the U.S. government to export Blackwell chips to China—offering a significant revenue share—highlight both the commercial importance of the Chinese market and the regulatory hurdles faced by U.S. chipmakers 13. The U.S. expectation of a 15% levy on H20 chip sales, which Nvidia says could prompt litigation, adds further complexity 3. Meanwhile, China is moving quickly to build a self-sufficient AI chip ecosystem, with Cambricon, Huawei, and SMIC expanding production and local AI startups optimizing for domestic hardware 4. This shift is likely to further reduce U.S. chip sales to China and accelerate hardware-software alignment within China 4.

Nvidia ’s Q2 results beat expectations, but the market penalized a slight data-center revenue miss and the absence of H20 sales to China 2. The company’s $60B buyback and bullish forecast for global AI infrastructure spend ($3–4T by 2030) underscore ongoing enterprise demand, but also reflect investor sensitivity to growth signals in core business lines 215. Snowflake ’s upgraded outlook, in contrast to CrowdStrike ’s muted guidance, illustrates that AI-driven demand is strongest where cloud and data infrastructure spending is most visible 14.

Security and safety issues remain prominent. Anthropic’s launch of a browser-based Claude agent, paired with reports of AI-enabled “vibe-hacking” and ESET’s discovery of AI-generated ransomware, highlight the evolving threat landscape as agentic AI systems proliferate 612. Microsoft ’s warning about cloud-native extortion and the FBI’s global alert on China-backed Salt Typhoon cyber-espionage reinforce the need for robust security protocols in both enterprise and government deployments 1213. The Pentagon’s audit of Microsoft cloud services and new restrictions on foreign support for DoD systems further reflect heightened scrutiny of supply chain and data sovereignty risks 5.

On the product side, the pace of innovation remains high. Nous Research’s Hermes 4 open-weight model 7, Google ’s on-device Gemini Nano AI and “Nano Banana” video model 811, and xAI’s Grok Code Fast 1 coding assistant all signal increasing competition in both open-source and commercial AI offerings 9. OpenAI’s move to add parental controls and collaborate with Anthropic on safety evaluations demonstrates rising regulatory and societal expectations for responsible AI deployment 10.