Google Restricts Meta's Gemini Usage — AI Demand Strains Infrastructure Capacity

Google has placed restrictions on Meta's usage of Gemini, according to a report by the FT (Financial Times). The move comes as demand for AI services expands rapidly, with infrastructure capacity believed to be nearing its limits.

As more companies tap AI infrastructure such as Google Cloud, the supply-demand balance has broken down. Google introducing usage limits for a major customer is a development that symbolizes the infrastructure crunch facing the industry as a whole. Competition over AI compute resources is intensifying not only in model development but also in terms of supply capacity.

Pentagon Signals Expanded AI Role in Military Targeting

The U.S. Department of Defense (the Pentagon) has signaled that it intends to further expand the role AI plays in setting military targets. Bloomberg reported.

While incorporating AI into the target selection process promises faster decision-making, ethical and legal concerns over automation have also been raised. The use of AI in the military domain is advancing in many countries, and beyond the United States, nations such as China and Israel are said to be actively adopting the technology. For now, the "human-in-the-loop" principle—where humans make the final judgment—is believed to be maintained, but international debate continues over how far AI involvement should be extended.

Oracle Shares Post Steepest Drop Since 2001 on AI Financing Concerns

Oracle's stock has recorded a sharp decline, its worst since the collapse of the dot-com bubble in 2001, driven by investor concerns over its AI-related financing. CNBC reported.

Oracle has been accelerating data center investment for AI infrastructure, but the market is worried about the toll that capital demand is taking on its financial health. It's a sign that companies that have ridden the AI boom are facing tougher scrutiny of the balance between investment scale and monetization. It could also trigger a broader valuation reset across AI-related stocks.

Text-to-Image Model "Sana 1.6B" Compressed to 1/8.6 Its Size — Ternarization Achieves Quality Close to FP16

Clark Labs has announced on Hugging Face that it compressed "Sana 1.6B," a Transformer-based text-to-image model, using ternary quantization, shrinking it down to 374MB.

The result is 1/8.6 the size of the original FP16 version (3.21GB) while reportedly maintaining quality close to FP16. The team adopted a ternarization approach that represents each weight with approximately 1.85 bits (a drastic cut from the conventional 16 bits). A hybrid design keeps roughly 5% of the parameters—such as conditioning and projection layers—at higher precision.

The work demonstrates that high-quality image generation is now feasible even on edge devices and in resource-constrained environments, and the fact that it has been released as open source (Apache-2.0) is another point worth noting.

U.S. Government Imposes Export Controls on "Claude Mythos 5" Over Its Zero-Day Mass-Discovery Capability

It has emerged that the U.S. government has imposed export controls on Anthropic's AI model "Claude Mythos 5," following a string of discoveries of long-hidden vulnerabilities, including a DoS vulnerability that lay dormant in OpenBSD's TCP stack for 27 years and a 16-year-old bug in FFmpeg.

The case has thrown into sharp relief the double-edged nature of the technology: while AI is a powerful tool for security research, the same technology could become a means of serious cyberattacks if misused. The government's regulation is seen as part of a broader, full-scale move toward managing the security risks posed by AI.

---