System integration emerges as a new driver for AI server upgrades; Orient Securities recommends focusing on four major beneficiaries: Foxconn Industrial Internet, Lenovo Group, and others.
Recently, Orient Securities released a research report stating that advancements in process technology are driving improvements in chip performance, while advanced packaging has become another key driver of chip performance enhancement. Amid the slowdown in process technology upgrades, increasing the area of a single die can boost the number of transistors, thereby enhancing computational power. However, constrained by factors such as lithography machine design and reticle size limitations, the maximum feasible die size is approximately 800-900mm². NVIDIA’s H100 already approaches this limit. In its B200 series, NVIDIA has adopted an advanced packaging technique of combining two dies into one package, achieving an integration of 208 billion transistors within a single package, more than double that of the H100 (which contains 80 billion transistors). According to NVIDIA's technical roadmap, Rubin Ultra will integrate four dies into a single package, delivering a computing power of 100 PFLOPS FP4 per card.
System integration is becoming a new driver for performance improvement in AI servers. While advancements in wafer manufacturing processes and advanced packaging have met the performance upgrade needs of products like personal computers and smartphones, they may still fall short of the rapidly growing demand for AI computing power and the fast-paced development of AI server performance. We believe that system integration is now emerging as a new driver for performance enhancement. The number of GPUs in AI servers has increased from the typical eight GPUs per server to 72 GPUs per cabinet, and is expected to rise further to 144 GPUs in the VR Ultra NVL576 cabinet by 2027 (each GPU package contains four GPU dies, totaling 576 GPU dies). The increase in GPU count significantly raises requirements for thermal management, making system integration much more challenging. For instance, the production ramp-up of GB200 NVL72 has been constrained by the complexity of system integration. Industry-leading companies with technological advantages are expected to benefit from the higher barriers to entry in system assembly and improved competitive dynamics.
In terms of investment targets, maintain focus on companies related to AI server system integration:
Foxconn Industrial Internet: Significant optimization and improvement were achieved in Q2 compared to Q1 in the testing of GB200 series products; system-level cabinet debugging time was notably shortened, with automated assembly processes introduced, effectively improving overall production and delivery schedules. The company has expanded capacity across multiple global plants and deployed fully automated assembly lines. It expects GB200 shipments in Q3 to continue showing strong growth momentum. Major orders come from large North American cloud service providers, while projects for sovereign and branded customers are also advancing simultaneously. A quarter-by-quarter improvement trend is expected to be maintained throughout the year. The profit margin of GB300 per unit has the potential to surpass that of GB200 and could become a significant profit driver for the company’s AI server business next year.
Hygon Information Technology: Following the merger with Sugon, it is expected to form vertical integration capabilities encompassing CPUs, DCUs, and system integration.
Lenovo Group: NVIDIA previously stated that partners such as Lenovo are expected to launch various servers based on Blackwell Ultra starting in the second half of 2025.
Huaqin Technology: A core ODM supplier of AI servers for well-known domestic internet companies, providing full-stack shipments of switches, AI servers, and general-purpose servers, benefiting from downstream cloud capital expenditure expansion.
