Gf Securities Co.,Ltd. has released a research report suggesting that with the rapid growth in demand for generative AI, large language models, and high-performance computing, data centers are evolving from traditional general-purpose server architectures towards rack-level systems characterized by high density, high power, high interconnectivity, and a high proportion of liquid cooling.

AI cloud data centers are no longer just about improving individual server performance; they now involve a systemic upgrade centered around GPU/XPU, switching, storage, power supply, cooling, and mechanical structures.

The high-capacity, high-speed, low-latency architecture simultaneously elevates the importance of interconnectivity, power, thermal management, and structural support.

The report recommends focusing on the systemic opportunities arising from AI rack-level architecture upgrades, including high-speed interconnects, CPO/NPO, polarization-maintaining fiber, liquid cooling, power supplies, and structural components.

Key Drivers of the Upgrade Cycle

High-speed interconnect and liquid cooling are identified as core variables in this round of architectural evolution.

Data center connection solutions can be broadly categorized into four main links: high-speed connections, liquid cooling, power supply, and fiber optic transmission.

With the advancement of solutions like 224G high-speed connectors, copper cables, and CPO/NPO, the bottleneck for AI computing power is gradually shifting from single-chip performance to system connection efficiency.

Concurrently, the increase in AI rack power density is driving a transition in cooling from air cooling to cold plate and immersion liquid cooling, thereby raising the importance of liquid cooling connectors, piping, trays, and related structural components.

Impact on Optical Communication

CPO/NPO technologies are expected to boost demand for optical communication and specialty fibers.

CPO integrates the optical engine with the switch or compute chip within the same package, shortening the electrical signal transmission path from centimeters to millimeters compared to traditional pluggable optical modules, which reduces signal attenuation, power consumption, and latency.

NPO serves as a transitional solution, enhancing bandwidth density and energy efficiency while leveraging the mature pluggable ecosystem.

Since CPO internal optical engines often employ silicon photonics, and silicon waveguides are sensitive to the polarization of input light, polarization-maintaining fiber is typically required for internal fiber connections when using an external light source to ensure signal stability.

NPO may also use polarization-maintaining fiber, though generally in smaller quantities than CPO.

Power Supply Evolution

On the power supply side, the rapid rise in AI rack power density is expected to drive AIDC electrical architectures towards an 800V HVDC upgrade.

The trend towards low voltage and high current is also pushing onboard power supplies from discrete solutions towards modularization, vertical power delivery, and high-density integration.

NVIDIA's whitepaper notes that generational GPU performance improvements and the expansion of NVLink interconnect domains are significantly increasing rack power density, necessitating the relocation of power components away from the core compute area where possible.

This change not only reshapes power supplies, PCBs, and connectors but also imposes higher demands on internal rack space, load-bearing capacity, maintenance, and cabling.

Rising Importance of Structural Components

Regarding structural components, the upgrade to rack-level AI architecture is elevating the importance of mechanical infrastructure.

Systems like NVIDIA's MGX consist of compute trays, NVLink switch trays, and supporting mechanical, electrical, and piping infrastructure.

Modular racks provide structural integrity and maintainability, while sliding rails facilitate the smooth installation and maintenance of rack-mounted equipment.

With the introduction of heavier GPU configurations, large power shelves weighing hundreds of kilograms, mixed 1U-8U architectures, and liquid cooling systems, rails/slides are evolving from simple support components into critical elements ensuring the safety, thermal management, and operational efficiency of high-density computing equipment.

Server slides possess attributes like high load-bearing capacity, slim profiles, maintainability, and strong certification.

According to Dell Technologies, server rack rails are used to mount and secure servers within cabinets, optimizing server performance, improving airflow, and simplifying management and upgrades.

Compared to static rails, sliding rails allow a server to be pulled from the front of the rack even with equipment installed above it.

Industry barriers stem from patent protection, customer certification, and long-term customer stickiness, with certification cycles for major clients often being lengthy.

Benefiting from the expansion in server numbers and the trend towards heavier individual units, slides are upgrading towards higher load capacity, quick-release mechanisms, corrosion-resistant materials, precision ball bearings, and customization, potentially driving the slide market into a new cycle of volume and price increases.

Investment Recommendations

The report reiterates the suggestion to focus on systemic opportunities from AI rack-level architecture upgrades, including high-speed interconnects, CPO/NPO, polarization-maintaining fiber, liquid cooling, power supplies, and structural components.

It highlights companies focused on optical interconnect and polarization-maintaining fiber, such as AVIC Jonhon Optronic Technology Co.,Ltd., Changyingtong, Yangtze Optical Fibre and Cable Joint Stock Limited Company, Raycus Fiber Laser Technologies Co.,Ltd., and Changjin Photonics.

Investors should also monitor the trend towards high-density integration in tertiary power architectures, with examples including Zhongfu Circuit, Boke New Material, and Sinorelay.

Additionally, companies related to the HVDC trend in data centers, such as Liangel and Minshida, warrant attention.

Companies with capabilities in high-load-bearing slides, server rails, precision mechanical component design and manufacturing, customer certification, and patent portfolios are poised to benefit, with suggestions to follow Chuanhu Technology and Haidar.

Associated Risks

The report concludes by outlining potential risks, including slower-than-expected iteration of new technologies and products, intensifying market competition, and fluctuations in raw material and product prices.