华尔街点评GTC：在英伟达的定义里，算力即收入，Token是新的大宗商品

Core Insights - The core message from NVIDIA's annual GTC conference is that the commercial logic of AI computing power is undergoing a fundamental restructuring, with tokens becoming a new commodity and computing power equating to revenue [1] Group 1: Market Outlook - NVIDIA's management has significantly raised the visibility of data center sales from $500 billion (covering until 2026) to over $1 trillion (covering cumulative 2025 to 2027), indicating strong growth potential [1] - Morgan Stanley's report suggests that this new figure implies an upward potential of at least $50 to $70 billion compared to Wall Street's current consensus for data center revenue from 2026 to 2027 [1][2] - The high-confidence purchase orders for Blackwell and Vera Rubin systems have exceeded $1 trillion, doubling from the $500 billion reported in October 2025 [2] Group 2: Demand Structure - Demand is diversified, with approximately 60% coming from hyperscale cloud providers and the remaining 40% from CUDA cloud-native AI companies, NVIDIA cloud partners, sovereign AI, and industrial/enterprise customers [2] - The new $1 trillion outlook aligns closely with Wall Street's previous expectation of around $970 billion for the three-year data center revenue period [2] Group 3: Technological Advancements - NVIDIA emphasized the acceleration of traditional enterprise workloads, announcing collaborations with IBM, Google Cloud, and Dell, and introducing two new CUDA-X foundational libraries [3] - The integration of Groq 3 LPU with Vera Rubin is highlighted as the most important architectural release, enabling high throughput and low latency for advanced workloads [4][5] Group 4: Product Development - NVIDIA's roadmap extends to 2028, with a consistent annual architecture release schedule, including Blackwell (2024), Blackwell Ultra (2025), Rubin (2026), Rubin Ultra (2027), and Feynman (2028) [9] - The Vera CPU is projected to become a multi-billion dollar independent business, with capabilities that significantly enhance AI workloads [8] Group 5: Infrastructure Strategy - NVIDIA is pursuing both copper cable and co-packaged optics (CPO) routes simultaneously, confirming that customers can choose their preferred technology without being locked into a single option [7] - The architecture for Rubin Ultra and Feynman includes advanced features such as chip stacking and custom HBM, enhancing performance for AI workloads [9] Group 6: Market Positioning - Morgan Stanley believes NVIDIA's vertically integrated platform, spanning multiple chips and systems, is difficult to replicate and supports a more sustainable AI capital expenditure cycle than currently anticipated by the market [10]