Core Viewpoint - NVIDIA is shifting its focus entirely towards AI, as evidenced by its absence of gaming graphics cards at CES 2026 and the introduction of new AI products and architectures [2][10]. Group 1: AI Product Launches - NVIDIA unveiled the next-generation Rubin architecture GPU, which boasts inference and training performance that are 5 times and 3.5 times better than the Blackwell GB200, respectively [4][17]. - The company introduced five new product families targeting various AI applications, including the NVIDIA Nemotron for Agentic AI, NVIDIA Cosmos for physical AI, and NVIDIA Alpamayo for autonomous driving [6][8][39]. - The Vera Rubin NVL72 architecture was officially launched, featuring six core components designed to enhance AI data center capabilities [14][15]. Group 2: Performance Metrics - The Rubin GPU achieves an inference performance of 50 PFLOPS and a training performance of 35 PFLOPS under the NVFP4 data type, significantly surpassing its predecessor [17]. - Each Rubin GPU is equipped with 288GB of HBM4 memory and offers a bandwidth of 22 TB/s, supporting the high computational demands of modern AI models [18]. - The overall architecture of the Vera Rubin NVL72 can deliver 3.6 exaFLOPS of NVFP4 inference performance and 2.5 exaFLOPS of training performance [37]. Group 3: Networking and Connectivity - The introduction of NVLink 6 enhances interconnect bandwidth to 3.6 TB/s per GPU, with a total bandwidth of 260 TB/s across the entire NVL72 rack [20][21]. - The Vera CPU integrates 88 custom Arm cores and features a bandwidth of 1.8 TB/s for NVLink C2C interconnect, facilitating efficient communication between CPU and GPU [22]. Group 4: AI Model Developments - The Alpamayo model, a large-scale open-source visual-language-action model for autonomous driving, was launched with 10 billion parameters [41]. - The Nemotron series expanded to include specialized models for speech recognition, visual-language processing, and safety, enhancing AI applications across various sectors [49][51]. - The Cosmos model for robotics was upgraded to generate synthetic data that adheres to real-world physical laws, aiding in the development of AI agents [54][58]. Group 5: Industry Impact and Future Outlook - NVIDIA's comprehensive approach to AI, integrating models, data, and tools, is expected to strengthen its competitive edge and ecosystem lock-in [10]. - The company plans to begin mass production of the Vera Rubin NVL72 in the second half of 2026, indicating a strong commitment to advancing AI infrastructure [38].

Core Insights - NVIDIA has introduced the Alpamayo family of open AI models, simulation tools, and datasets aimed at enhancing the development of safe, reasoning-based autonomous vehicles (AVs) [1][12] - The Alpamayo models are designed to address the challenges of operating AVs in complex and rare driving scenarios, known as the "long tail" [2][12] - The integration of reasoning capabilities into AV decision-making is expected to improve safety, explainability, and scalability in autonomous driving [3][4] Group 1: Product Features - The Alpamayo family includes chain-of-thought, reasoning-based vision language action (VLA) models that enable AVs to process and respond to novel scenarios step by step [3][12] - Alpamayo 1, a key model in the family, features a 10-billion-parameter architecture that utilizes video input to generate driving trajectories and reasoning traces [13] - AlpaSim is an open-source simulation framework that supports high-fidelity AV development, providing realistic sensor modeling and scalable testing environments [13] Group 2: Industry Support and Collaboration - Major mobility leaders such as Lucid, JLR, and Uber are expressing interest in the Alpamayo models to develop reasoning-based AV stacks that facilitate level 4 autonomy [8][12] - The open-source nature of Alpamayo is seen as a catalyst for innovation within the autonomous driving ecosystem, allowing developers to adapt the technology for specific needs [9][12] - The release of Alpamayo is viewed as a significant advancement for the AV research community, enabling unprecedented scale in training and development [12][13] Group 3: Market Implications - The introduction of reasoning capabilities in AVs is anticipated to enhance their ability to navigate complex environments and make safe decisions in unpredictable scenarios [4][12] - The shift towards physical AI emphasizes the necessity for AI systems that can reason about real-world behavior, moving beyond mere data processing [9][12] - The comprehensive ecosystem provided by Alpamayo, including models, datasets, and simulation tools, is expected to accelerate the deployment of safe, reasoning-based autonomous vehicles [4][12]

Open Model Ecosystem - NVIDIA leads the open model ecosystem [14, 100] - 80% of startups are building on open models [10] - 1-in-4 OpenRouter tokens are generated by open models [10] AI Performance and Benchmarks - NVIDIA's Llama Nemotron Nano VL 8B achieves 70.2% in Text 4 Recognition, 69.1% in Text 4 Referring, 61.8% in Text 4 Spotting, 81.4% in Relation 4 Extraction, 39.2% in Element A Parsing, 31.9% in Mathematical 4 Calculation, and 73.1% in Visual Unders A [20] - nvidia/canary-gwen-2.5b achieves an average WER of 5.63 [26] New NVIDIA Technologies - NVIDIA announces Alpamayo, an open reasoning VLA for autonomous vehicles [61, 65] - NVIDIA ships full-stack AV on 2025 Mercedes Benz CLA [68] - NVIDIA Vera CPU features 88 custom Olympus cores, 176 threads, 1.8 TB/s NVLink-C2C, 1.5 TB system memory, 1.2 TB/s LPDDR5X, and 227 billion transistors [120] - NVIDIA Rubin GPU offers 50 PFLOPS NVFP4 Inference (5X Blackwell), 35 PFLOPS NVFP4 Training (3.5X), 22 TB/s HBM4 Bandwidth (2.8X), 3.6 TB/s NVLink Bandwidth per GPU (2X), and 336 billion transistors (1.6X) [122] - NVIDIA ConnectX-9 Spectrum-X SuperNIC provides 800 Gb/s Ethernet, programmable RDMA, line-speed encryption, and 23 billion transistors [125] - NVIDIA BlueField-4 offers 800G Gb/s DPU, 64 Core Grace CPU, 6X Compute, 2X Networking, 3X Memory BW, and 126 Billion Transistors [127] - NVIDIA NVLink 6 Switch scales up fabric with 3.6 TB/s per-GPU bandwidth and 108 billion transistors [131] - NVIDIA Vera Rubin NVL72 achieves 3.6 EFLOPS NVFP4 Inference (5X Blackwell), 2.5 EFLOPS NVFP4 Training (3.5X), 54 TB LPDDR5X Capacity (3X), 20.7 TB HBM Capacity (1.5X), 1.6 PB/s HBM4 Bandwidth (2.8X), 260 TB/s Scale-Up Bandwidth (2X), and 220 Trillion Transistors (1.7X) [134] - NVIDIA Spectrum-X Ethernet Co-Packaged Optics scales to 102.4 Tb/s with 200G silicon photonics and 352 billion transistors [136]