Guotai Haitong: GTC is coming soon, focusing on AI chips, computing power, and storage direction.

Guotai Haitong's research report stated that in 2026, the most worth paying attention to at the GTC is not the parameter refresh of a particular chip, but whether NVIDIA will release Rubin's systematic implementation, Feynman's roadmap, and upgrades in optical interconnection, power supply, and liquid cooling integration, officially pushing the industry from "buying GPUs" to "deploying AI factories" in a new stage. Areas to watch include AI chips, computing power, and storage. Event: The NVIDIA GTC will be held in San Jose, California from March 16-19, covering a wide range of topics including AI agents, AI factories, science-oriented AI, CUDA, high-performance inference, open models, physical AI, quantum computing, and more. Guotai Haitong's main points 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 consisting of CPU, GPU, interconnection, networking, and system components. NVIDIA is transitioning the delivery unit of AI infrastructure from individual boards to entire cabinet systems. With the Vera Rubin platform officially entering mass production at CES 2026, this GTC is likely to unveil an enhanced version of this architecture - Rubin Ultra. A Rubin Ultra cabinet will integrate 144 GPUs, building a Scale-up network with a bandwidth of up to 1.5PB/s, and individual chip bidirectional interconnect bandwidth reaching 10.8TB/s. To achieve such high-density interconnects, Rubin may adopt a double-tier network topology and implement "fiber in, copper out" 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 one of the first chips to use TSMC's A16 process, integrating Groq's LPU hardware stack for the first time. Production of Feynman is expected to start in 2028, with customer shipments possibly starting between 2029 and 2030. Feynman may introduce extensive integration with SRAM-based or 3D stacking technologies, with individual chip power consumption expected to exceed 5000W. Feynman's main appearance at this time is in the form of a roadmap or architecture preview. Its value may not lie in short-term commercial delivery, but in demonstrating how NVIDIA understands the AI computing demands of the post-Rubin era to the market. Additionally, NVIDIA may showcase a new inference chip integrating Groq's "Language Processing Unit" (LPU) technology, signaling NVIDIA's active layout in the inference computing field to meet the market's demand for efficient and cost-effective computing solutions. Data center infrastructure restructuring driven by optical interconnection, power supply, and liquid cooling. In terms of interconnection, CPO and silicon photonics are becoming important directions for large-scale AI systems. In the future, data center interiors will gradually move from traditional copper interconnects to higher bandwidth density, lower loss optical connection systems. With power supply, solutions such as 800V HVDC, high-integration modular power supplies, and vertical power delivery reflect that the key factor limiting the expansion of AI systems is no longer just chip manufacturing capability, but also whether power can be efficiently and stably delivered to each computing node. On the cooling front, air cooling is losing its adaptability for ultra-high power consumption computing platforms, and liquid cooling is moving from an optional solution to a standard configuration, driving upgrades in cold plates, thermal interface materials, and cabinet-level liquid cooling systems. Risk warning: AI development falls short of expectations; technological iteration and industry applications fall short of expectations; intensified industry competition.