Author: Yilong Bao The 2025 OCP Global Summit has sent a clear message: the "Scale Up" architecture has become the core theme of AI data center infrastructure development. On October 20, Morgan Stanley's Asia-Pacific team released a research report emphasizing that to meet the endless demand for computing power driven by AI, the entire industry is racing toward larger-scale, higher-density "Scale Up" architectures. The report highlights the need for investment focus to shift from general server components to core technology suppliers capable of supporting supernode architectures. The summit clearly identified four key technology trends and their primary beneficiaries:
- Larger Racks: The AMD Helios ultra-wide rack architecture was showcased, driving upgrades in rack components, with Wistron and Wiwynn as major beneficiaries. - Higher Power: The 800V Direct Current (VDC) power supply has emerged as the next-generation standard, set to revolutionize the power distribution architecture of data centers, with Delta Electronics and BizLink leading the charge. - Enhanced Cooling: 2MW Liquid Cooling Distribution Units (CDUs) took the spotlight, particularly Google's Deschutes solution. - Faster Networks: AI-optimized Ethernet (ESUN) and CPO switch technology are on the rise, presenting upgrade opportunities for networking equipment providers like Accton.
In summary, the entire industry is gearing up for the impending rollout of gigawatt-scale AI data center clusters in the coming years, and companies capable of providing higher density, more efficient solutions are poised to take center stage in the next growth cycle.
Double-width Racks Initiate a New Era of Scale Up
The "Scale Up" approach aims for higher density in single-node computing, leading to a revolutionary change in rack formats. AMD, in collaboration with Meta and Wiwynn, introduced the Helios rack, which features the Open Rack Wide (ORW) specification, doubling the width of traditional ORV3 racks (21 inches). With the high floating-point computing performance (FLOPs) density of current high-performance chips, connecting more computing cores in low-latency environments necessitates placing them within the same scale-up domain. Given the limitations of current copper wire technology, this can only be achieved through larger backplanes or mid-plates, resulting in larger racks. Meta envisions a future with decoupling; while rack power density will continue to increase in the short term, it is expected to decline eventually due to the adoption of optical technologies, freeing it from copper interconnect limitations. Helios racks are set to begin shipping in the second half of 2026, with primary clients including Meta, Oracle, and OpenAI. Supply chain investigations indicate that Wiwynn is a major ODM partner for Meta, while Wistron serves as the primary ODM partner for GPU modules, substrates, and switch trays, with most PCBs requiring M9-grade CCL materials. Moreover, these ultra-wide heavy-duty racks place higher demands on mechanical components such as chassis and rails, benefitting suppliers like Chenbro and King Slide.
800V DC Power Architecture Leads the Next Generation of Efficient Gigawatt AI Factories
As rack power density soars, traditional power supply architectures are becoming unsustainable. The spotlight is now on the 800V DC (VDC) power supply, heralded as the critical technology driving the next generation of gigawatt-scale AI factories. Compared to traditional 50V architectures, the 800V DC solution can transmit over 150% more power on equivalent copper cables and improve power usage effectiveness (PUE) by around 5%. In terms of advancements, Delta Electronics has showcased mature solutions, including 1.2MW solid-state transformers (SST, mass-produced, with designs for over 3MW in development), 800V electronic fuses (eFuse), 90kW DC-DC power supplies, and 12kW distribution boards. The new solution is expected to more than double the power supply value per watt compared to current designs. Power interconnect suppliers like BizLink will also benefit from the increased demands for liquid-cooled busbars and other stringent specifications. The report states that the 800V DC solutions are anticipated to debut in the second half of 2027 alongside NVIDIA's Rubin Ultra platform.
Large-scale Liquid Cooling Systems Become the Focus
Cooling is the lifeline that determines whether computing power can be reliably delivered. The technological path showcased at the summit is clear, evolving from hybrid cooling to full liquid cooling. Specifically:
GB300 Status: The mass-produced GB300 compute trays utilize a hybrid cooling approach (85% liquid cooling / 15% air cooling), with only six quick-disconnect (QD) connectors per compute tray. Yield rates are no longer a market concern. VR200 Outlook: The next-generation VR200 platform will feature complete liquid cooling, increasing the number of quick-disconnect connectors to 14 per compute tray. It is currently in rack-level production and testing, with deliveries expected by the end of Q3 2026. CDU Standardization: Google has open-sourced its 2MW CDU design, supporting pressures of up to 80 PSI, facilitating advanced cold plate designs. BOYD, Cooler Master, Delta Electronics, and Invec have all showcased related products.
The report cites predictions from Promersion, which suggest that while cool plate technology will remain the market leader until 2030, the tipping point for immersion liquid cooling is expected to arrive in 2028.
Networking Technology Continues to Optimize to Meet AI Demands
In addition to strategies for scaling up architecture within nodes, enhancing high-speed interconnectivity between nodes is also vital for maximizing AI cluster performance. The report notes that the Ethernet solutions (ESUN) and CPO switches launched to improve network performance have been widely applied in optimizing AI data networks. However, reliability, maintainability, and cost issues remain critical factors affecting their widespread adoption. Specifically, recent developments include:
Accton and Others: Accton and Tianhong have both demonstrated their latest 1.6T network switch products based on Broadcom’s Tomahawk 6 ASIC, anticipated to start early applications by the end of 2026 or early 2027. Accton also showcased proof-of-concept for CPO switches based on Tomahawk 6 ASIC and IRIS wavelength switches. Meta’s Research: Meta disclosed findings showing that its 51.2T CPO (Co-Packaged Optics) switch has an annualized link failure rate (ALFR) of just 0.34%, significantly better than the 1.58% rate for pluggable optical modules, indicating a distinct reliability advantage, although cost and maintainability remain key to widespread adoption. Meanwhile, Active Electrical Cables (AEC) are emerging as a cost-effective solution, steadily gaining market share in scale-out networking. Meta’s GB300 racks incorporate AEC, a trend expected to continue benefiting suppliers like BizLink.
In conclusion, the 2025 OCP Global Summit has delivered an extremely clear message: the arms race for AI infrastructure has entered a "giantization" phase, with scale-up emerging as the central theme throughout the event.