NVIDIA H20 Chips Return to Chinese Market
The U.S. government approved NVIDIA's resumption of exports to China of its specially designed AI chips H20 and RTX PRO. Featuring "reduced compute power + retained memory bandwidth" to comply with export controls, these chips alleviate China's AI computing shortage. Their NVLink 900GB/s high-speed interconnect enables 768GB pooled memory for 8-card clusters, supporting training for 70B-parameter models.
Lisuan Tech Launches First Fully Self-Developed GPU "7G100" Series
Chinese GPU firm Lisuan released its high-performance graphics GPU chip and "Lisuan eXtreme" graphics card, based on its self-developed Tiantu architecture. It supports NRSS (a DLSS-equivalent image enhancement tech), achieving >70 FPS avg. in Black Myth: Wukong (1080P high settings) and a FireStrike score of 26,800. Mass production starts in September for cloud gaming, smart cockpits, and more.
Innoscience Becomes NVIDIA’s Sole Chinese Partner for 800V Power Architecture
Innoscience was selected as an 800V DC power architecture supplier for NVIDIA, boosting single-rack power density beyond 300kW and propelling AI datacenters into the megawatt era. Its Hong Kong shares surged nearly 50% on the news.
China Energy Group Launches Domestic GPU Server Tender
The state-owned giant is procuring 4 Haiguang-based domestic inference GPU servers capable of high-concurrency runs for DeepSeek-R1/V3 (671B-parameter) models. Key hardware must be Chinese brands.
OpenAI Plans 1 Million GPU Deployment by Year-End
CEO Sam Altman announced a $40B investment in compute clusters to achieve "100x scaling" and break current large-model training bottlenecks. This could reshape global AI infrastructure competition.
AMD Develops Dedicated NPU Cards to Replace GPUs
Leveraging Xilinx tech, AMD is creating specialized NPU cards optimized for generative AI and LLM inference to reduce GPU load. OEMs like Lenovo, Dell, and HP are evaluating prototypes.
South Korea Invests $1.1B to Procure 13,000 GPUs
The government will distribute NVIDIA B200/H200 chips to Naver Cloud, NHN Cloud, etc., cutting reliance on overseas compute power and boosting national AI research.
NVIDIA H20 Unban Reveals New Supply Chain Strategy
Analysts note top cloud providers are adopting a "triple-track strategy" (NVIDIA + Huawei Ascend + in-house chips) to counter "special-edition chip traps" that widen tech gaps. Ecosystem development remains key for domestic alternatives.
Domestic GPU Ecosystem Barriers Rapidly Eroding
Huawei’s Ascend 910B cuts distributed training failure rates to 1/5 of H20’s via heterogeneous architecture, but CUDA migration still requires 30% code refactoring. Open-source communities now enable 90% auto-operator conversion, boosting framework compatibility.
Policy-Driven Compute Network Expansion
China’s "East-West Computing Resource Transfer Project" synergizes with domestic chips, with Ascend and Cambricon claiming 67% of government cloud share. The U.S. plans new curbs on interconnect bandwidth, accelerating output in the Yangtze River Delta cluster (target: 58% of 2025 capacity).
Weekly Insight: Global computing competition now runs on dual tracks—"tech sovereignty" and "ecosystem breakthroughs." While giants (OpenAI, South Korea) accelerate hyperscale deployment, China focuses on full-stack R&D (Lisuan GPUs) and supply-chain resilience ("triple-track"). Policy-market synergy is reshaping the industry landscape.