Guotai Haitong Securities released a research report stating that the most noteworthy aspect of the 2026 GTC is not the parameter upgrades of any single chip, but whether NVIDIA will advance the industry from "purchasing GPUs" to a new phase of "deploying AI factories." This transition is expected to be driven by the systematic implementation of the Rubin platform, the release of the Feynman roadmap, and integrated upgrades in optical interconnect, power supply, and liquid cooling. The firm recommends focusing on AI chips, computing power, and storage. The NVIDIA GTC conference will be held from March 16 to 19 in San Jose, California, covering areas such as agent-based AI, AI factories, scientific AI, CUDA, high-performance inference, open models, physical AI, and quantum computing. Key insights from Guotai Haitong are as follows: The mass production and systematic implementation of the Rubin platform. Rubin is no longer just a single GPU product but an integrated AI supercomputing platform comprising CPUs, GPUs, interconnects, networking, and system components. NVIDIA is shifting the delivery unit of AI infrastructure from individual boards to full cabinet systems. With the Vera Rubin platform confirmed for mass production at CES 2026, this GTC is likely to unveil an enhanced version—Rubin Ultra. A single Rubin Ultra cabinet will integrate 144 GPUs, establishing a Scale-up network with a bandwidth of up to 1.5 PB/s, and each chip will achieve a bidirectional interconnect bandwidth of 10.8 TB/s. To enable such high-density interconnects, Rubin may adopt a two-layer network topology and transition from copper to optical connections within the cabinet. The forward-looking disclosure of the Feynman architecture is expected to be the most strategically significant highlight of the conference. Feynman may be among the first chips to adopt TSMC's A16 process and integrate Groq's LPU hardware stack. Production of Feynman is anticipated to begin in 2028, with customer shipments likely between 2029 and 2030. Feynman may introduce extensive SRAM-based integration or 3D stacking technology, with single-chip power consumption projected to exceed 5,000W. Feynman is expected to be presented primarily as a roadmap or architectural preview, with its value lying not in short-term commercial delivery but in illustrating how NVIDIA envisions AI computing needs in the post-Rubin era. Additionally, NVIDIA may showcase a new inference chip incorporating Groq's "Language Processing Unit" (LPU) technology, signaling its active expansion into the inference computing domain to meet market demand for high-efficiency, low-cost computing solutions. The restructuring of data center infrastructure driven by optical interconnects, power supply, and liquid cooling. In terms of interconnects, CPO and silicon photonics are becoming critical for large-scale AI systems, with future data centers gradually shifting from traditional copper interconnects to optical systems offering higher bandwidth density and lower loss. On the power supply front, solutions such as 800V HVDC, highly integrated modular power delivery, and vertical power supply reflect that the key constraint on AI system scalability is no longer just chip manufacturing capability but the efficient and stable delivery of power to each computing node. In cooling, air cooling is becoming inadequate for ultra-high-power computing platforms, with liquid cooling increasingly transitioning from an optional solution to a standard configuration. This shift is driving simultaneous upgrades in cold plates, thermal interface materials, and cabinet-level liquid cooling systems. Risks include slower-than-expected AI development, delays in technological iteration and industry adoption, and intensified industry competition.
Comments