英伟达AI工厂破局物理极限 新技术激起A股千层浪

Core Insights - The increasing demand for AI computing power is pushing single data centers to their physical limits in terms of power, cooling, and space [2] - NVIDIA has introduced the Spectrum-XGS Ethernet technology, which integrates geographically dispersed data centers into a unified AI super factory to address these limitations [2][3] - The future of computing power competition is shifting from individual data centers to cross-regional computing networks, with NVIDIA aiming to create a global "AI factory" network [5] Technology Overview - The Scale-Across model is emerging as the third pillar of AI computing, alongside Scale-Up and Scale-Out methods, allowing for the integration of data centers across different geographical locations [3] - NVIDIA's Spectrum-XGS technology has improved data center integration efficiency by 40% for its client, Coreweave [3] - The technology addresses high latency and performance uncertainty in long-distance data center interconnects, significantly enhancing distributed AI training efficiency [2][3] Market Implications - The introduction of GW-level AI supercomputing centers is expected to drive upgrades in optical communication infrastructure, increasing demand for optical modules and hollow-core fibers [6] - High-end PCB manufacturers are likely to benefit from the need for advanced circuit boards to support high-speed signal transmission in these new data centers [6] - The trend towards "powering computing" through grid-like scheduling and delivery of computing resources is anticipated to accelerate with the establishment of these AI factories [6] Industry Trends - The data center industry is evolving from traditional Internet Data Centers (IDC) to AI Data Centers (AIDC) due to the exponential growth of AI technology and AIGC [5] - Hollow-core fiber technology is expected to see rapid growth in demand due to its advantages in low latency and high capacity, making it suitable for long-distance data transmission [7][8] - The application of hollow-core fibers is still in its early stages, with challenges in mass production and cost optimization needing to be addressed before widespread commercial use [8]