AI Industry Development & Policy Dynamics

Zhipu AI officially open-sourced GLM-5.3-Flash, a large language model with 320 billion parameters but activating only 18 billion per inference. Model weights are now available on Hugging Face under MIT license, supporting native text, image, and video

According to Adam Fry, OpenAI's ChatGPT product lead, a routing bug recently directed approximately 3% of Pro and Thinking requests to GPT-5.5-mini instead of the selected GPT-5.6 model. The issue has been fixed. Users reported noticing response speed increases and quality degradation. Network requ

According to MSX.COM data, the U.S. stock market opened lower on August 26, with the Dow Jones Index down 0.12%, S&P 500 down 0.14%, and Nasdaq Composite down 0.17%. The VIX volatility index rose 1.75%. AI-related stocks declined broadly, with SAP dropping 3.08%, Palo Alto Networks falling 2.94%,

According to Odaily, MiniMax's To B business now accounts for approximately 80% of annual recurring revenue (ARR), up significantly from around 30% a year ago, with To C business declining to 20% from 70%. The shift reflects accelerating adoption among enterprises and developers. MiniMax

According to Google Cloud, Gemini Enterprise will now offer a pay-as-you-go pricing model, allowing enterprises to avoid fixed subscription fees and set monthly spending caps. Once budget is exhausted, AI agents will pause operations. Long-term commitment customers receive 10% to 20% discounts,

According to Techcrunch, AI Agent testing infrastructure startup Arga Labs raised $10 million in seed funding today, with General Catalyst leading the round. BoxGroup, Emergence Capital, Gradient Ventures, and SV Angel also participated. The company provides testing and training infrastructure for e

According to PANews, MiniMax disclosed during its earnings call on August 26 that the M3Pro model will increase its parameter scale to approximately 3T and expand reinforcement learning and long-context task training to improve model generalization and intelligence capabilities. CEO Yan Junjie

OpenAI announced its first AI chip, the Jalapeño, on Tuesday, stating it delivers industry-leading speed and efficiency for inference workloads. The chip is being developed with Broadcom and will be deployed within OpenAI's compute infrastructure by the end of the year. Industry analysts told CNBC t

According to multiple analysts interviewed by CNBC, OpenAI unveiled its Jalapeño AI chip on Tuesday, marking the first step toward reducing reliance on Nvidia's GPUs. The custom-built chip, designed for inference workloads, delivers industry-leading speed and efficiency comparable to Nvidia's

According to BMO Capital Markets, Broadcom has received an Outperform rating and a $455 price target following the investment bank's initiation of coverage. Separately, Broadcom is in talks to raise over $60 billion in debt financing for AI chip offerings to clients including Anthropic, with

According to PANews, citing MiniMax's mid-year earnings call on August 26, the company is advancing the adaptation of its M3 and H3 models to domestic chips. A large-scale domestic GPU cluster will be deployed shortly and will gradually assume real production

According to B.AI, on August 25, the platform's daily new API users reached 23,432, marking an all-time high, while daily token throughput surged to 594.71 billion tokens, also setting a new record. During the same period, B.AI accumulated 79,500 new API users, processed 74.108 million token

According to Wccftech, Tesla CEO Elon Musk recently shared progress on the company's AI5 inference chip during a discussion with hedge fund manager Ron Baron. The AI5, which completed tape-out in April and is now being produced by Samsung Electronics, delivers 2-3x better power efficiency per watt c

Chinese AI startup Moonshot AI is in early talks with Microsoft, Amazon, and Google over revenue-sharing agreements for its Kimi K3 model. The Alibaba-backed startup is seeking up to a 30% share of revenue generated from K3-related services running on Microsoft Azure, Amazon Web Services, and Google

Mirae Asset Asset Management's TIGER AI Semiconductor Core Process ETF holds Samsung Electro-Mechanics at a 23.96% weighting, the highest allocation to the component maker among domestic semiconductor supply chain ETFs, according to data as of the 19th. Jeong Eui-hyeon, head of Mirae Asset's ETF

According to Yonhapinfomax, on August 26, Mirae Asset's TIGER AI Semiconductor Core Process ETF holds Samsung Electro-Mechanics (23.96%) as its largest position, followed by LG Innotek (17.03%) and I.S. WIS (16.75%). The ETF head noted that MLCC prices rose approximately 10% in June and surged

According to PANews, Alibaba's Qwen launched Qwen3.8-Flash today, an early preview of the Qwen4 architecture. The multimodal mixture-of-experts model features 125 billion parameters and 51 billion N-gram embedding parameters, with per-token activation of only 6 billion parameters, delivering high co

According to Beating AI, Alibaba Qwen released Qwen3.8-Flash today, with open-source weights also available under Qwen3.8-Flash-Next, featuring the next-generation Qwen4 architecture. The multimodal MoE model has 125B total parameters with only 6B activated, plus 51B N-gram Embedding for memory

According to Odaily, Nvidia launched the Jetson Orin Nano 2 on August 26, targeting edge AI applications with 78 TOPS computing power, 8GB memory, and an eight-core Arm CPU. The new chip delivers double the inference performance compared to its predecessor Jetson Orin Nano Super while reducing power