巨头加速抛弃英伟达

Core Viewpoint - Major tech companies, including Microsoft, are accelerating efforts to reduce dependence on NVIDIA's GPUs, which dominate 90% of the AI chip market. Companies are developing custom chips to enhance efficiency and lower costs, while NVIDIA is transforming into a "full-stack AI" infrastructure provider to maintain its market leadership [2][4][7]. Group 1: Microsoft's AI Chip Development - Microsoft has launched its commercial AI chip "Maia 200," which is designed for high-performance AI inference, claiming it is three times more efficient than AWS's latest AI chip and offers 30% better performance within the same budget [5][6]. - The Maia 200 chip utilizes TSMC's 3nm process and integrates SK Hynix's HBM3E memory, with plans to support OpenAI's latest models [5][6]. - Microsoft aims to shorten the production to deployment timeline for its chips, indicating a potential reduction in reliance on NVIDIA [5][6]. Group 2: Other Companies' Custom Chip Initiatives - Google is using its custom Tensor Processing Units (TPUs) for training and running its Gemini AI models, which outperform GPUs in certain tasks while reducing operational costs [6]. - AWS has released its Trainium3 AI chip, boasting a fourfold increase in computing performance and a 40% reduction in energy consumption compared to its predecessor [6]. - Meta is exploring the use of Google's TPU in its upcoming data centers, while OpenAI is collaborating with Broadcom to develop a custom chip set for release later this year [6]. Group 3: NVIDIA's Market Position and Strategy - Despite the rise of custom chips from competitors, NVIDIA continues to expand its business into AI models and robotics, aiming to maintain competitiveness in a diversifying market [7]. - NVIDIA is also venturing into CPU supply, recently announcing a $2 billion investment in CoreWeave to deploy its CPUs, challenging Intel and AMD [7]. - The company is actively developing AI models and platforms, including an open-source weather forecasting AI model and the Omniverse platform for robotic simulations [7]. Group 4: NVIDIA's Growth Projections - NVIDIA is expected to surpass Apple as TSMC's largest customer this year, with projections indicating that 22% of TSMC's revenue in 2025 will come from NVIDIA, compared to Apple's 18% [8].