Nvidia introduced a new superchip system Vera Rubin with enhanced specifications. The company positions itself not only as a GPU supplier but also as a CPU provider for agentic AI systems, emphasizing improvements in performance and energy efficiency.
NVIDIA
Intel Corporation
OpenAI
NVIDIA unveiled Spectrum-6, a 102.4 terabit Ethernet switch that doubles the bandwidth of the previous generation. It will be part of the NVIDIA Vera Rubin platform and used in gigascale AI factories. Early customers include CoreWeave, Microsoft, Nebius, SpaceXAI, and Tesla.
NVIDIA
Google DeepMind announced three new AI models: Gemini 3.6 Flash, 3.5 Flash-Lite, and a specialized 3.5 Flash Cyber. These models aim to boost efficiency, reduce latency, and improve performance in agentic workflows. Gemini 3.6 Flash uses 17% fewer output tokens and is cheaper than its predecessor. 3.5 Flash-Lite is the fastest model in the 3.5 series, while 3.5 Flash Cyber is designed for cybersecurity.
Google/DeepMind
Google DeepMind
Baidu ERNIE agent model topped the PinchBench v2 ranking, surpassing GPT-4 and other models. This achievement underscores Baidu's leadership in AI agents.
Baidu
CrowdStrike researchers discovered a worm attacking the AI software supply chain. The malware steals credentials, gains access to systems, and destroys files, masquerading as legitimate AI agent activities, making detection extremely difficult.
CrowdStrike
In a special report by Xinhua, it is highlighted that China has put forward an initiative aimed at jointly shaping the future of global artificial intelligence. This initiative is designed to give a powerful impetus to the development of AI worldwide, promoting international cooperation and coordination in this field.
Information about a mysterious model called 'Kivine' has surfaced online, which is believed to be an early version of Moonshot AI's Kimi K3. The model demonstrates the ability to process up to one million context tokens, impressing users worldwide.
Moonshot AI
Amazon has introduced the Self-Distilled Reasoning (SDR) method for supervised fine-tuning (SFT) of Amazon Nova 2 models. SDR generates chain-of-thought (CoT) reasoning paths for datasets lacking them, using the model itself. This addresses the problem of reasoning suppression and reduces catastrophic forgetting, improving target performance while retaining general capabilities.
Amazon Web Services
After leaving Intel, Pat Gelsinger became a partner at venture firm Playground Capital, which invests in deep technology. He is betting on startups in the field of lithography to overcome the physical limitations of semiconductors and revive Moore's law. Gelsinger believes that the key to progress lies in improving light sources for lithography and supports the company xLight.
ASML
NVIDIA
Groq
Cerebras Systems
Gathering data for physical AI in the real world is a slow, expensive, and risky process. Simulation enables generating large volumes of photorealistic physical data through GPU parallelism. This article reviews popular simulation engines: MuJoCo, MuJoCo Warp, NVIDIA Isaac Sim and Isaac Lab, as well as the new physics engine Newton, developed by NVIDIA, Google DeepMind, and Disney Research.
NVIDIA
Google DeepMind
MuJoCo
MSI has introduced the EdgeXpert MS-C931, a compact version of the NVIDIA DGX Spark (Project DIGITS) powered by the Grace Blackwell GB10 SoC, delivering up to 1000 TOPS in FP4. The device features 128 GB LPDDR5x, a 4 TB NVMe SSD, 10GbE, and an optional 200GbE ConnectX-7 for clustering. The review includes performance tests, power consumption, and capabilities for working with large language models.
NVIDIA
Micro-Star International (MSI)
Realtek
MediaTek
Taiwan-based Wistron has launched a 324,000-square-foot factory in Texas to produce NVIDIA GB300 Grace Blackwell Ultra and Vera Rubin superchips. The $700 million investment has created over 500 jobs, with plans to hire 1,000 employees by year-end. The project is part of NVIDIA's commitment to building AI infrastructure in the United States.
NVIDIA
Poolside released Laguna S 2.1, an open-weight model with 11.8 billion parameters and 8 billion active per token, designed for agentic coding. Based on a Mixture-of-Experts (MoE) architecture, it supports up to 1 million tokens of context and can run on a single NVIDIA DGX Spark. Laguna S 2.1 ranks first among open models with known size on Terminal-Bench 2.1 (70.2%) and SWE-Bench Multilingual (78.5%), outperforming many larger competitors.
Poolside
DeepSeek
NVIDIA
Tencent
Moonshot AI
Alibaba/Qwen
Anthropic