Today's News
ChinAI Newsletter
Zhipu AI released GLM-5.2 with 1 million token context window trained exclusively on Huawei Ascend chips without any NVIDIA compute, achieving competitive performance on coding benchmarks against frontier closed models. The model is open-weight under MIT license and free to download and deploy. This represents a concrete milestone in China's chip self-sufficiency for large-scale model training, demonstrating meaningful progress in domestic infrastructure buildout that DeepSeek and other competitors have not yet achieved at comparable scale.
Air Street Press via Exponential View
Air Street Press's State of AI Compute Index v6 reveals that of 1.66 million announced Grace-Blackwell GPUs, just 100,128 are deployed and 308,640 are installing—meaning roughly 95% remain undeployed since shipping began in December 2024. The largest announced programs are sovereign or hyperscale projects: HUMAIN (Saudi Arabia) at up to 600,000 GB300, Stargate Abilene at 450,000, South Korea's national 260,000-unit program, and xAI Colossus 2 at 110,000 installing. The constraint has shifted from GPU availability to power, land, cooling, interconnect, permitting, and debt—a GPU order is not a cluster, and a cluster is not always available capacity.
Center for AI Safety via Import AI
Multiple concrete benchmarks signal accelerating AI capability on economically meaningful work. The Remote Labor Index shows frontier models automating end-to-end online work (3D design, video production, data analysis) with Fable 5 reaching 16.1% success rate on freelance tasks—up from 2.5% nine months prior—while GPT-5.5 achieved 6.3% and Opus 4.8 reached 8.3%. Separately, Fable achieved the fastest GPU kernel optimization on KernelBench-Mega with an 18.71X speedup over PyTorch baseline using a single cooperative kernel launch per token, outperforming Opus 4.8 (14.4X) and GPT-5.5 (4.34X). These metrics signal both expanding economic capability relative to human comparative advantage and improving infrastructure efficiency through autonomous kernel optimization.
Mari Kiyohara / Bloomberg via Techmeme
Micron Technology broke ground on a ¥1.5 trillion ($9.3 billion) expansion of its Hiroshima factory to produce high-bandwidth memory for AI processors, with manufacturing equipment delivery and installation scheduled for the second half of 2028. Japan's Ministry of Economy, Trade and Industry allocated up to ¥500 billion to help cover costs. The investment underscores intensifying competition in AI memory supply chains and signals major semiconductor capacity additions timed to meet sustained AI infrastructure demand.
Unknown via Exponential View
A new training methodology enables dramatically smaller models to compete with 28x larger systems on complex multi-step tasks, with direct implications for inference cost economics and deployment efficiency across the AI stack.
Anthropic embedded Chinese user detection in Claude, removed after discovery
ChinAI Newsletter
Anthropic added user identification logic in April 2026 that remained undetected for three months before exposure, prompting Alibaba to mandate removal of Claude from employee machines. The incident reveals geopolitical tensions in AI deployment and potential supply chain fragmentation risks for frontier model providers.
Liu Jia via Bill Bishop at Sinocism
State Council Order 837, effective July 1, elevates two decades of departmental rules into law and creates countermeasure mechanisms allowing China to blacklist foreign entities discriminating against Chinese investors, restrict their access to China-related trade and investment, and deny entry or residence. This represents a significant hardening of China's overseas expansion framework with explicit security and retaliation provisions.