Nvidia announced the acquisition of AI chip startup Groq for $20 billion. Groq's CEO and President will join Nvidia, aiming to integrate Groq's LPU (Language Processing Unit) architecture, which significantly outperforms GPUs in AI inference speed and energy efficiency. Valued at only $6.9 billion three months ago, this acquisition represents a nearly 3x premium. This move is seen as a defensive consolidation by Nvidia against market share erosion from custom AI chips from companies like Google and Amazon, signaling a new phase of competition in the AI hardware ecosystem.

Chinese AI startup Z.ai released the open-source large model GLM-4.7, focusing on coding capabilities. It achieved 73.8% on the SWE-bench benchmark, becoming the first Chinese lab to break the 70% barrier, surpassing the performance of DeepSeek-V3.2, Kimi K2, and Claude Sonnet 4.5. GLM-4.7 is now open-sourced on Hugging Face and is compatible with mainstream AI programming agents like Claude Code. Having just passed its Hong Kong IPO hearing and expecting $300 million in funding, Z.ai demonstrates that China's open-source models are competitive on the global frontier.

OpenAI posted an executive position for 'Head of Preparedness' with an annual salary of $555,000. Responsibilities include addressing severe risks posed by frontier AI models capable of self-improvement. The role directly manages four major risk domains: AI self-improvement, cyber offense/defense, biological threats, and manipulation, and holds veto power over model launches. This marks the first time Altman has explicitly mentioned in an official job description that 'systems can improve themselves,' indicating that recursive AI capabilities have moved from theory into practical R&D and governance stages, highlighting the urgent need for safety governance around AI's closed-loop self-enhancement.

AI development tool company Cursor announced the acquisition of code review platform Graphite. The two will integrate AI code generation with automated review capabilities to create an intelligent development loop from IDE to PR. Graphite had previously raised $52 million in Series B funding, serving major clients like Shopify and Snowflake. Post-acquisition, Cursor will compete head-to-head with GitHub across platform-level stages like code generation, review, and deployment, driving deep changes in AI-powered engineering collaboration and development processes.

The Poetiq team, by adding a reasoning system on top of OpenAI's GPT-5.2 X-High model, achieved approximately 75% on the ARC-AGI-2 public set, significantly surpassing the 60% human baseline and all comparable models. This result shows AI already possesses capabilities surpassing humans in areas like abstract reasoning and complex task decomposition. However, the official verification score remains lower than the public set, reminding the industry to pay attention to the gap between benchmark testing and actual general intelligence.

AI evaluation agency METR released an analysis showing that Anthropic's Claude Opus 4.5 model completed a nearly 5-hour continuous autonomous task without human intervention, setting a new record for the longest autonomous working duration among current large models. This demonstrates a significant enhancement in AI's capabilities for long-chain tasks and sustained reasoning.

Alibaba-backed MiniMax released the M2.1 model, specifically optimized for various programming languages and mobile/Web application development. It enhances AI's code generation and toolchain integration capabilities in practical software engineering scenarios, further enriching the competitiveness of Chinese AI models in the engineering application field.

Chinese AI companies are accelerating their catch-up across multiple dimensions, including open-source large models, AI chips, and capital markets. Firms like Z.ai and MiniMax have secured large-scale funding and broken through key technical bottlenecks. Models like GLM-4.7 are on par with top US and European models in international mainstream benchmarks. With progress in Hong Kong listings and AI chip self-sufficiency, China is expected to achieve comprehensive breakthroughs in open-source models and hardware ecosystems by 2026.

Following Nvidia's acquisition of Groq, Groq's original SRAM inference patents and technical team will be integrated into Nvidia. Industry expects Nvidia will use Groq's patents to build a 'patent moat' in the SRAM inference domain, accelerating the trend towards customized and specialized AI inference hardware. Competition for patents and high-end chip talent is expected to further intensify.

A Jevons Paradox phenomenon is emerging in AI reasoning and knowledge work: as AI dramatically reduces the cost of knowledge work, the primary scenarios for future AI compute and token consumption will be new types of tasks and applications that do not exist today. AI will stimulate a large volume of knowledge work demand that was previously infeasible due to high costs, driving sustained industry expansion.