NVIDIA (NVDA)：我们回归工作，从Jensen的CES主旨演讲和分析师问答开始

NVIDIA (NVDA): It's back to work for us, starting with a JensenCES keynote and analyst Q&A Yesterday NVIDIA opened their 2026 Consumer Electronics Show with a Jensen keynoteas well as a financial analyst Q&A with Jensen and CFO Colette Kress. Note that we will bemeeting with the company 1 on 1 today. Rubin in full production, and looks like a monster...Vera Rubin is now in full production,and is still expected to begin shipping in 2H26,with the platform boasting materialperformance uplifts vs Blackwell (5x more inference performance, 3.5x more trainingperformance, 2.8x more memory bandwidth, and 2x more NVLink bandwidth per GPU),accomplished with only 1.6x more transistors given a high level of co-design across thevarious components (CPU, GPU, NIC, DPU, NVLink switch, and Spectrum-X CPO). Remainder of the keynote was more incremental, but announcements included1)Alpamayo open-source models for reasoning-based autonomous driving; 2) new openmodels for physical AI, and 3) new storage infrastructure for long-context inference. Unlikemost past CES keynotes, gaming GPUs were not even mentioned. NVIDIA implied that competitors would find it extremely difficult to challenge theVera Rubin platformgiven the company’s co-otimized system approach and full-stackinvestments, especially as competitors mostly remain focused on single chip offerings. Physical AI appears set for an inflection with Autonomous Driving leading thecharge,with NVIDIA leveraging their GB300 platform for training, Orin and Thor chipsfor inference, and RTX pro chip for simulation. The first NVIDIA powered MercedesAutonomous Vehicle (the CLA) is set to hit production later this year. Management believesthe effort (across both edge autonomy and datacenter) is already a $5-$10B business, andbelieves it will be “very large” by the end of the decade. Management is still awaiting clarity from the government on China H200 exports,but noted strong China demand and “plenty" of supply for ALL countries (i.e. they indicatedany need for China supply would not interfere with shipment needs to the ROW). Gross margins remain under control this year; long term margin sustainabilityremains dependent on the company’s continued ability to bring value(i.e. to drivethe most performance possible, at the lowest cost, with the highest throughput in a givenpower footprint). That being said, we didn’t hear anything to suggest the company believedtheir ability to do this would wane (quite the opposite). NVIDIA values Groq’s efforts on utlra-low latency inference, noting that highthroughput and low latency are conjunctional, and in a future where response timescould trend towards being instantaneous, NVIDIA was willing to pay for the opportunity todevelop a unique solution in this area. They do not see any products coming out of the dealimpacting or replacing their current business. (continued on next page) Finally, Jensen harped (numerous times!) on the sheer magnitude of currentdemand.We suspect 2026 is going to be a very good year for the company as a strongGB300 ramp transitions to Rubin, and valuation appears very attractive. We rate the stock Outperform, $275 PT. BERNSTEIN TICKER TABLE INVESTMENT IMPLICATIONS NVDA (Outperform, $275):The datacenter opportunity is enormous, and still early, with material upside still possible APPENDIX - FINANCIAL FORECASTSU.S. SEMICONDUCTORS