Overseas AI news gathered on August 6 covered a wide range of developments: fundraising, data center investment, physical AI, content generation, model performance, and reactions from the developer community. The standout stories: DeepSeek resuming a major funding round, and reports of a massive UAE data center investment in Japan.
DeepSeek Resumes Funding Round at a Valuation of Around $74 Billion
According to reports, Hangzhou-based AI startup DeepSeek has resumed its second funding round, targeting roughly $8 billion (about ¥1.2 trillion). Investment firm Monolith Management is reportedly considering participating. The valuation is believed to be near 500 billion yuan (around $74 billion).
The company had suspended the process last month, reportedly out of frustration after remarks by its founder were leaked to investors; this round appears to be its restart. If the valuation of DeepSeek, which has drawn global attention with its open large-scale models, reaches the $74 billion level, it would rank among the largest ever for a Chinese AI company. The information comes from people familiar with the matter, and no official announcement has been confirmed at this point.
UAE-Backed Fund Considers $6.3 Billion AI Data Center in Akita Prefecture
Japan has picked up its own mega-investment headline. Mubadala Investment Company, a sovereign fund of the UAE (United Arab Emirates), is reportedly leading a plan to build an AI data center in Akita Prefecture at a scale of up to ¥1 trillion (about $6.3 billion). The project is said to involve around 500 megawatts of capacity, with other investors from Japan and abroad possibly taking part.
With semiconductor and power constraints much discussed, the location in northeastern Japan and the region's renewable energy potential look set to attract renewed attention. Information remains limited at this point, however, and neither the finalization of the plan nor its timeline is known.
NVIDIA Announces "Cosmos 3", an Open Model for Physical AI
NVIDIA has announced Cosmos 3, an open foundation model for "physical AI" spanning robotics, autonomous driving, and vision AI. It adopts an architecture NVIDIA calls Mixture-of-Transformers, enabling a single model family to handle visual reasoning, world generation, and action prediction. The lineup includes Super (64B) for high performance, Nano (16B) for efficient inference, and Edge (4B) for edge devices and on-board robotics.
It is offered under the Linux Foundation's OpenMDW 1.1 license, and a key strength is that companies can retrain it on their own data. NVIDIA says it has taken the top spot among open models on multiple benchmarks. The company plans to integrate Cosmos with "Isaac GR00T" for robots, "Alpamayo" for autonomous driving, and "Metropolis" for vision AI, while expanding its partner alliance, the "Cosmos Coalition", for Japan's robotics and manufacturing industries.
AI Now Generates Over 95% of New Chinese Micro-Dramas
AI adoption is accelerating rapidly in content production as well. Reports say that in Q1 2026, more than 95% of new Chinese micro-dramas (vertical short-form video dramas) were generated with AI. The figure highlights how, in the fast-growing micro-drama market, AI has already taken over scripting and video generation at an industrial scale.
At the same time, concerns have been raised about the impact on actors' work and the quality of mass-produced content.
Chinese LLMs Close In on the Frontier; Kimi K3 Leads on Cost Efficiency
Chinese players continue to make their presence felt in model performance. Alibaba's "Qwen3.8 Max" scored 56 on Artificial Analysis's Intelligence Index, up 10 points from the previous generation, reportedly putting it on par with Anthropic's "Claude Opus 4.8". Meanwhile, Moonshot's "Kimi K3" posted an even higher score at about 25% lower cost — a sign that competition is intensifying not just over raw performance but over cost efficiency as well.
Suno to Introduce Watermarking for Generated Songs, Responding to Mounting Copyright Lawsuits
Suno, the AI music generation company, has announced that it will begin rolling out a feature that embeds watermarks in the songs it generates. Facing multiple copyright lawsuits from the recording industry and others, the company is signaling that it will answer legal demands with provenance technology for its outputs. "Provenance" for generated content is heading toward standardization across other AI tools as well, making this an industry-wide trend to watch.
Developer Community Roundup: Handling AI-Generated Code and Local LLM Pitfalls
In open-source development, debate continues over how to handle AI-generated patches. The maintainer of the Linux wireless (WiFi) subsystem has laid out a strict policy of explicitly rejecting low-quality AI-generated patches (so-called "slop").
For local LLM users, cautionary reports have emerged around Gemma 4. On third-party quantized builds (such as those from Unsloth), multimodal features (image and audio recognition) have been reported to break on recent llama.cpp builds, outputting meaningless tokens; switching to the official GGUF restored functionality. On a more practical note, a developer blog post pointing out that having an LLM write email subject lines breaks Gmail's threading display has also been drawing attention.