Core Insights - The 2025 OCP Global Summit emphasizes "Scale Up" architecture as the core theme for AI data center infrastructure development [1][8] - The industry is shifting focus from general server components to core technology suppliers that support supernode architectures to meet the endless demand for computing power from AI [1][8] Group 1: Scale Up Architecture - "Scale Up" architecture aims for higher density single-node computing, leading to revolutionary changes in cabinet design [2][5] - AMD, in collaboration with Meta and Wistron, introduced the Helios cabinet, which features an Open Rack Wide (ORW) specification that is twice the width of traditional cabinets [2][3] - The Helios cabinet is expected to start shipping in the second half of 2026, with major clients including Meta, Oracle, and OpenAI [3] Group 2: Power Supply Innovations - The 800V direct current (VDC) power supply architecture is highlighted as a key technology for the next generation of gigawatt-level AI factories [4][5] - Compared to traditional 50V systems, the 800V solution can transmit over 150% more power on the same copper cable and improve power usage efficiency (PUE) by approximately 5% [6] - Delta Electronics has showcased mature solutions, including a 1.2MW solid-state transformer and 800V electronic fuses, with the new 800V solution expected to debut with NVIDIA's Rubin Ultra platform in the second half of 2027 [6] Group 3: Cooling Solutions - Large-scale liquid cooling systems are becoming a focal point, with a clear technological path evolving from hybrid cooling to fully liquid cooling [7] - The GB300 computing tray, which has entered mass production, utilizes a hybrid cooling solution, while the next-generation VR200 platform will be fully liquid-cooled [7] - Google has open-sourced its 2MW cooling liquid distribution unit (CDU) design, supporting high-pressure applications for advanced cooling solutions [7] Group 4: Network Technology Enhancements - Enhancements in network technology are crucial for meeting AI demands, with Ethernet solutions (ESUN) and CPO switch technology being widely adopted for optimizing AI data networks [8][9] - The reliability and cost of these products remain key factors affecting their widespread adoption [8][9] - Active cables (AEC) are emerging as a cost-effective solution, gaining market share in network expansion, as seen in Meta's GB300 cabinet [9]