China Galaxy Securities released a research report stating that the AI industry is gradually transitioning from the generative AI era to the inference AI and agent AI eras. Currently, inference and channel training have become the core computational demands driving growth in the AI sector. The firm anticipates that the value per cabinet for related power supply manufacturers is expected to see significant growth. It believes that supporting the Vera Rubin platform will require liquid cooling systems capable of handling greater flow rates and newer cooling architectures, which should continue to drive demand for technical upgrades among liquid cooling solution and component suppliers. The main points from China Galaxy Securities are as follows:
Event: On March 17, Beijing time, NVIDIA's annual developers conference (GTC 2026) opened in California. Founder and CEO Jensen Huang delivered a keynote speech sharing AI industry trends, unveiling the company's latest Vera Rubin full-stack AI computing platform system, and providing outlooks on new technology directions such as cabinet power supplies and all-liquid cooling solutions.
Industry Developments: 1) Inference and agent demands represent industry trends: The AI industry is progressively moving from the generative AI era into the inference AI and agent AI eras. Inference and channel training are now the core computational needs fueling AI industry expansion. Over the past two years, this has driven AI computational demand to increase by approximately one million-fold. 2) Throughput efficiency and interactivity/inference speed are key metrics for AI factories: Tokens have become the core production factor in AI factories. The essence of throughput efficiency (token/watt) is power efficiency; interactivity/inference speed, i.e., the token generation rate, determines the "intelligence level" of AI. 3) NVIDIA designates 2025 as the "Year of Inference": The company is fully optimizing the entire AI inference process to reduce customer AI infrastructure costs, aiming to become the world's lowest-cost and most reliable AI infrastructure platform.
Revenue Expectations: NVIDIA holds an optimistic revenue outlook for its Blackwell and Rubin flagship chip product lines in the data center segment through 2027. It is projected that the combined revenue from these two product lines in computing and networking will exceed $1 trillion, a significant increase compared to the $500 billion expectation disclosed at the company's 2025 GTC conference for the period through 2026.
Product Developments: 1) Official launch of the Vera Rubin full-stack AI computing platform system: This includes seven computing and interconnect chips such as the Vera CPU, Rubin GPU, and Groq 3 LPU, along with five types of racks: NVL72 (computing), Groq-3 LPX (optimized for the decoding stage, i.e., token output generation performance), standalone Vera CPU, BlueField-4 STX (storage), and Spectrum-X (switching). A single Vera Rubin full-stack AI computing platform system configuration consists of 16 Vera Rubin NVL72 racks, 10 Groq-3 LPX racks, 2 standalone Vera CPU racks, 2 BlueField-4 STX storage racks, and 10 Spectrum-X switching racks. 2) Significant computing power improvement: The launch of the Vera Rubin AI computing platform system has increased computing power by 40 million times over a decade. 3) Smooth progress in production capacity: Microsoft Azure has deployed the first Vera Rubin rack; NVIDIA's supply chain has achieved a production capacity of thousands of racks per week, capable of supporting multi-gigawatt-level AI factory construction monthly, while simultaneously mass-producing both Vera Rubin and GB 300 racks.
Power Supply Developments: 1) Cabinet power: Based on the Vera Rubin NVL72 reference design, the firm expects the VR NVL72 to utilize 4 PowerShelves, each with a power rating of 110KW, consisting of six 18.3KW power supplies, resulting in a total power supply capacity of 440KW. This represents a substantial increase of over 60% compared to the 264KW power supply of the GB200/GB300 NVL72. The firm anticipates that the value per cabinet for related power supply manufacturers is likely to see considerable growth. 2) External cabinet power: NVIDIA announced plans to fully adopt 800V high-voltage DC power supply, CPO optical interconnects, and high-density PCBs, aiming to reduce data center PUE below 1.1.
Liquid Cooling Developments: 1) Vera Rubin adopts a 100% liquid-cooled design with cable-free tray architecture, using 45°C hot water cooling. This eliminates the compression step in mechanical cooling, leading to significant energy efficiency improvements, lower data center cooling costs, and freeing up more power for computation. However, correspondingly, cooling with higher water temperatures requires greater cooling flow rates to remove the same amount of heat. The firm expects that supporting Vera Rubin will necessitate liquid cooling systems capable of higher flow rates and newer cooling architectures, which should persistently drive demand for technical upgrades among liquid cooling solution and component manufacturers. 2) Vera Rubin deployment efficiency is greatly enhanced, reduced from the traditional 2-day installation time to just 2 hours, which is expected to significantly improve maintainability. 3) The newly launched Vera Rubin platform employs a relatively conventional liquid cooling solution. The firm believes that with the future release of higher-power chips like Rubin Ultra and Feynman, new technology directions such as microchannels, microfluidics, and diamond cooling are worth anticipating.
Risk Warning: Risks include potential underperformance of industry policies, slower-than-expected progress in new technologies, sharp increases in raw material prices, operational difficulties for enterprises, political instability overseas, and deterioration of the trade environment.