中国科技与通信行业 英伟达与高通主题演讲核心要点

NVDA and Qualcomm Keynote Takeaways CITI'S TAKE Kyna WongAC+852-2868-7820kyna.wong@citi.com NVDA CEO Jensen Huang delivered a GTC keynote today where he noted: (1)Vera Rubin has reached full commercial production, seemingly confirmingthe 2H26 ramp for the NVDA supply chain (FII, VGT, WUS). (2) NVDA hasevolved from being a GPU-focused system company into a full-stack AIinfrastructurecompany,which may extend its partnerships topower/coolingsupply chains including Chinese players.(3)With theintroduction of the Vera CPU specifically for the "Agentic AI" era andreinforcement learning, we see positive implications forLenovo,which isalso one of ecosystem partners set to benefit from the new GPU serverplatform and Nvidia Windows–based laptops with higher ASPs. We alsowatch for the long-term potential demand impact toMontage’s memoryinterfacechips given Nvidia’s shift toward SOCAMM/LPDDR5.(4)Qualcomm’s new data center portfolio Dragonfly may provide an additionalsource to China ASIC datacenter solutions. Kevin Chen+852-2501-2125kevin.y.chen@citi.com Karen Huang+852-2501-2755karen.xw.huang@citi.com Key highlights from Nvidia GTC keynote —(1) Software should be presented as anagent; tokens are for revenue/profit generation. We see many Chinese software/ICTcompanies migrating to agent’s provider and token-based business model. (2) VeraRubin has reached full commercial production, which likely confirms the 2H26 rampfor the NVDA supply chain (FII, VGT, WUS); TFC was highlighted twice in the videopresentation. (3) Nvidia is now a full-stack AI infrastructure company and implied itwould expand its supply chain control beyond L10 to include L11 (rack) and L12(cluster). (4) All about CPU—Nvidia sees Vera CPU as the next growth driver not onlyfor its data center business but also edge devices with the launch of the RTX Sparklaptop and new lines of PCs. (5) Agentic AI and Physical AI—The companyhighlighted autonomous driving and industrial and humanoid robotics. Key highlights from Qualcomm Computex keynote —(1) Qualcomm launchedDragonfly, its new product brand dedicated to data center environments. (2)Snapdragon platforms are evolving into native execution layers for third-party AIagents. (3) Qualcomm is extending edge silicon into high-growth mobility andindustrial automation frameworks like vehicles, robotics, and 6G wireless. Implications —Vera Rubin has reached full commercial production, which shouldbe positive to the NVDA supply chain, i.e.,FII, VGT, WUSon 2H26 ramp. We also seepositive implications toLenovo,which is also an ecosystem partner set to benefitfrom the new GPU server platform and the Nvidia Windows-based laptop withhigher ASP contribution. That said, we watch for the long-term potential demandimpact to Montage’s memory interface chips given that Nvidia’s Vera CPU and Sparkplatform adopt the SOCAMM/LPDDR5X memory architecture instead of RDIMMand hence could erode DIMM chip opportunities should Nvidia CPU claim a largershare of server processors. We believeQualcomm’s new datacenter portfolioDragonfly could provide an additional source to China ASIC datacenter solutions.NVIDIA’s and Qualcomm’s keynotes both signal that the industry has transitionedfromthe foundational LLM paradigm—characterized by static prompt-and-response interactions—into the agentic AI and physical AI era, suggesting apotentially faster-than-expected pace of AI edge proliferation in China. See Appendix A-1 for Analyst Certification, Important Disclosures and Research Analyst Affiliations. Key takeaways from Nvidia and Qualcomm keynote speeches From standalone GPU acceleration toward engineering of integrated and pod-scale systems —Nvidia emphasized that it is a full-stack AI infrastructure company.The centerpiece of its next-generation AI infrastructure is the Vera Rubin platform, amulti-rack supercomputing system that has reached full commercial production.Built on an advanced 3nm process, the platform integrates seven specialized chips.The system utilizes next-generation HBM4 memory architectures to overcomememory bandwidth bottlenecks. Large scale-out and scale-up capabilities aredriven by the unified integration of NVLink 72, CX9 SuperNICs, and Bluefield-4DPUs. To sustain operational power requirements exceeding 5,000 amps, the Rubinarchitecture implements advanced 45°C liquid cooling loops. It also features acable-less midplane design, removing physical connectivity failure points andincreasing systemic resiliency. The Nvidia Vera Rubin DSX AI Factory referencedesign outlines how to design, build, and operate the entire AI factory infrastructurestack including NVL rack, Vera GPU rack, spectrum-6 SPX switch, Groq 3 LPX rack,and Vera BlueField-4 STX storage rack. NVDA’s first host processor Vera CPU designed specifically for autonomous AIworkloads —Built on the Olympus Core, the Vera CPU achieves the industry'shighest instructions per clock (IPC), capable of decoding 10 instructions per cycle. Itis the first enterprise CPU t