英伟达最大客户，彻底变心？

Core Insights - Microsoft is transitioning its AI workloads from GPUs to self-developed accelerators, aiming for better performance per dollar [2] - The company has launched its first AI accelerator, Maia 100, but it still lags behind competitors like NVIDIA and AMD in performance metrics [2] - Microsoft plans to develop a second-generation Maia accelerator, expected to be more competitive in computing, memory, and interconnect performance [3] Group 1: Transition to Self-Developed Chips - Microsoft has purchased a significant number of GPUs from NVIDIA and AMD but intends to shift most of its AI workloads to in-house chips [2] - The driving force behind this transition is the "performance per dollar" metric, which is crucial for large-scale cloud service providers [2] - Microsoft CTO Kevin Scott confirmed the long-term goal is to primarily use self-developed chips in data centers [2] Group 2: Current and Future Developments - The Maia 100 AI accelerator was introduced in 2023, enabling the migration of OpenAI's GPT-3.5 to Microsoft's chip, freeing up some GPU capacity [2] - The Maia 100 features 800 teraFLOPS of BF16 performance, 64GB of HBM2e memory, and 1.8TB/s memory bandwidth, which is significantly lower than NVIDIA and AMD offerings [2] - Microsoft is also developing a self-designed CPU named Cobalt and various platform security chips for cryptographic processing [3] Group 3: Competitive Landscape - Despite Microsoft's efforts, it is unlikely to completely replace NVIDIA and AMD chips in its data centers, as many customers still require these GPUs [3] - Google and Amazon have deployed thousands of their own accelerators, but still rely heavily on NVIDIA and AMD for large-scale deployments [3]