Nvidia’s board is reported to have launched an independent investigation and hired outside legal counsel to determine whether CEO Jensen Huang has become a target of foreign intelligence agencies. Huang remains CEO; there is currently no evidence that company secrets or national security have been compromised, and the investigation is ongoing.

Anthropic is reported to have purchased $3.4 billion worth of Caterpillar gas-fired power generation equipment, with a total capacity of 1.7 gigawatts. The deal involves taking over orders from OnePWR Solutions that had not yet secured financing. The equipment is expected to be delivered within two years, helping the company secure power for future computing needs.

China had more than 700 million generative AI users in the first half of 2026, with a penetration rate above 50%. Intelligent Q&A was the main use case, accounting for 76% of users; image and video generation accounted for 47.8%, and text processing for 37.6%. Usage of general-purpose AI assistants and productivity tools both grew by more than 100% year over year.

Microsoft showcased Mage-VL-Cowork, a 4-billion-parameter on-device vision-language model released under the Apache 2.0 license. It can choose to use a mouse, keyboard, or terminal to complete tasks. On 361 tasks in OSWorld-Verified, the model scored 66%; each task was limited to a maximum of 50 steps, and Google Drive tasks were excluded. In a spreadsheet demo, it completed a task using both a graphical interface and a terminal in 39 seconds; the comparison model used only the graphical interface.

DeepSeek’s open-source agent application Harness reportedly has around 1.54 million daily active users and processes approximately 92 trillion tokens per day. It has about 35,000 third-party plugins. The application reached this user scale within three months, and its plugin ecosystem includes not only productivity tools but also interface customization, virtual pet collection, and launch animations.

Google open-sourced the Elephant-Goldfish Model plugin under the Apache 2.0 license. An Elephant session drafts a specification, which a stateless Goldfish subagent then independently reviews in five stages before any code is written. The plugin supports tools including Antigravity, Claude Code, Codex, and ChatGPT. Tests completed end-to-end execution, although some host environments had issues with token consumption and isolation.

Researchers from Tsinghua University and Moonshot AI introduced TCSExam to evaluate frontier models on theoretical computer science research. GPT-6 Astra proved 70% of research-level theorems in Lean formal proof tasks, but did not solve any of 30 open conjectures. The evaluation covers different task types, including formal proofs and open problems.

Google DeepMind chemistry researchers used Antigravity to extract compound data from patents and papers, analyze factors that affect molecular potency, and generate candidate molecules and suggested synthesis routes in one hour. The system also provides an interactive viewer for researchers to explore molecules, inspect their properties, and examine the relationship between structure and potency.

In tests with Qwen3-30B-A3B, NF12 weight compression reduced VRAM usage from 56.87 GiB with BF16 to 43.00 GiB, a decrease of 24.4%. On an RTX PRO 6000 Blackwell Server Edition with an input length of 8192 tokens, throughput increased from 127.1 to 149.7 tokens per second. The method can exactly reconstruct the original weights, but different kernels are not guaranteed to produce identical outputs; vLLM support is still at the proposal stage.

The Hugging Face team released an MoE training guide drawing on approximately 800,000 GPU-hours of experiments. It covers cluster and data preparation, expert configuration, routing design, load balancing, inference deployment, and reinforcement learning post-training. The guide includes 55 interactive diagrams, runnable code, and performance data, and is free for developers to use.