Summary of Key Takeaways from APAC's OCP Summit Industry Overview - The report focuses on the AI server development industry, particularly the challenges and innovations discussed at the APAC's Open Compute Project (OCP) Summit held in Taipei on August 5-6, 2025. Core Points and Arguments 1. **AI Datacenter Clusters and Power Innovations**: Current large-scale datacenters consume between 50-100MW and are not fully compliant with AI server hosting standards. Companies like Meta are planning to build Titan AI clusters, with the first 1GW+ AI datacenter (Prometheus) expected to be operational by 2026. Another datacenter, Hyperion, will have a capacity of up to 5GW. Key beneficiaries include Delta, Vertiv, and power plant vendors [1][12][31]. 2. **Capex Increase for US CSPs**: The top four US Cloud Service Providers (CSPs) are expected to increase their capital expenditures (Capex) by 55% and 25% year-over-year in 2025 and 2026, respectively, indicating a robust multi-year investment in AI infrastructure [1][12]. 3. **Modular Design Trends**: The DC-MHS (Datacenter Modular Hardware System) is gaining traction as a modularized computing platform that enhances scalability and cost efficiency. Intel is a key player in this trend, which is expected to see high adoption rates starting with its next-generation server platform [9][12]. 4. **Customized Liquid Cooling Solutions**: There is a rising demand for customized liquid cooling solutions among the top four US CSPs. Amazon is shifting to its proprietary thermal solution, IRHX, while Google is developing a customized L2L CDU design. Delta is identified as a key ODM for these solutions [12][31]. 5. **HVDC Power Technologies**: The report discusses bifurcated HVDC (High Voltage Direct Current) power technologies, with a focus on the ±400V architecture being more favorable due to its compatibility with existing datacenter practices. This shift could benefit suppliers of SiC or GaN components [18][21]. 6. **Increased Content in Power Rack Designs**: CSPs are expected to introduce side power rack designs to accommodate higher power density. This will lead to significant content increases for system vendors like Delta, as the complexity of power delivery systems rises [21][23]. 7. **Vertical Power Delivery Adoption**: Vertical power delivery is anticipated to become more common in next-generation AI accelerators, driven by the need for power efficiency. This trend is expected to benefit power vendors like Delta and Infineon [29][31]. 8. **Networking Equipment Upgrades**: The demand for high-end networking equipment is increasing due to the growth of AI clusters. Ethernet technologies are being explored as alternatives to Infiniband for scale-up networking solutions, which could benefit Ethernet switch vendors [31][32]. Other Important Insights - The introduction of OpenBMC and Caliptra is expected to enhance security in server designs without necessarily lowering entry barriers for BMC chip vendors [7][12]. - The report highlights the importance of customized solutions in the evolving landscape of AI server infrastructure, indicating a shift towards more tailored approaches in hardware design and cooling solutions [12][31]. This summary encapsulates the key insights and trends discussed at the OCP Summit, emphasizing the ongoing innovations and investment opportunities within the AI server development industry.

Core Insights - The recent GTC conference highlighted the update of the AIDC power architecture, marking a significant development in NVIDIA's supply chain [1][2] - Key focus areas include the independence of power units, increased usage of power shelves, and the importance of components like BBU, supercapacitors, and PDU [1][2][3] - The new power architecture is expected to enhance the value and importance of server power suppliers, with Megmeet positioned to benefit as a pioneer in the Power rack solution [1][3] Power Unit Independence - The GTC conference confirmed that Power rack will become the new generation power technology solution, addressing the increasing power demands of GPUs [1][2] - The new architecture involves moving Power shelf, BBU shelf, and supercapacitor shelf out of the IT rack to form an independent Power rack [2][3] Increased Usage of Power Shelves - Concerns about the sidecar solution reducing the usage of Power shelves were alleviated, as the new architecture actually increases their usage due to an additional power circuit [4][5] - The architecture allows for a three-tier voltage conversion process, enhancing the overall power supply reliability [4][5] Market Potential and Value - The Power rack is projected to have a total value of approximately $216,365, with significant contributions from various modules [6][7] - The introduction of the Power rack is expected to create new revenue streams for Megmeet, transitioning its role from a simple AC/DC supplier to an integrated power solution provider [3][6] Implementation Timeline - Short-term plans involve using AC PSU solutions for GB200 and GB300, while mid-term upgrades will adopt the 800V HVDC sidecar power architecture [8] - The market anticipates that by 2025, Megmeet and other suppliers will begin to see new business opportunities from BBU and supercapacitor products, with significant growth expected in 2026 [8][9] Competitive Advantage - Megmeet is currently the only supplier to have released a sidecar integrated solution, positioning it to capture a significant market share in the Power rack segment [9] - The anticipated growth in AI computing power in China is expected to further boost Megmeet's domestic business, laying a foundation for future market expansion [9]

Core Viewpoint - The GTC conference highlighted the transition from Power shelf to Power rack as the next-generation power solution for AI data centers, indicating a significant upgrade in the importance of power suppliers in the industry [2][4]. Group 1: Power Rack Development - The Power rack integrates multiple components such as Power shelf, BBU shelf, Super cap shelf, and PDU, enhancing the overall value and competitive barriers in the AI power supply industry [2][3][6]. - The shift to Power rack architecture addresses the increasing power demands of advanced computing nodes while optimizing space utilization within data center racks [4][6]. Group 2: SST Technology - SST (Solid State Transformer) is identified as the future power supply technology, capable of converting high voltages (10-34.5KV) to 800V DC, facilitating modular design and voltage adaptability across different regions [3][12]. - The implementation of SST technology is expected to reduce power loss significantly by increasing voltage levels, thus minimizing current and associated losses [15][19]. Group 3: Super Capacitor Benefits - The use of super capacitors can significantly reduce power supply volatility, with Delta's data showing a reduction in supply fluctuation from 73% to 6% [20][24]. - Different super capacitor technologies exhibit varying characteristics, with EDLC offering longer cycle life and LiC providing higher energy density, which can enhance the dynamic load capacity of GPUs in AI data centers [21][22]. Group 4: Future Roadmap - NVIDIA anticipates the gradual application of the Rubin and Feyman architectures between 2026 and 2028, which will further accelerate the adoption of Power rack solutions [11][12]. - Google's roadmap for data center power architecture evolution indicates a transition towards higher voltage levels and modular designs, aligning with the industry's shift towards SST solutions [13][15].