Apple(US:AAPL) 3 6 Ke·2025-12-17 02:51Core Viewpoint - Apple is developing its first AI server chip, codenamed "Baltra," in collaboration with Broadcom to avoid purchasing chips from Nvidia, indicating a strategic shift towards in-house chip development for AI applications [1][7]. Group 1: Chip Development and Collaboration - Apple is working with Broadcom to develop the Baltra chip, which is expected to be produced using TSMC's 3nm process and enter mass production in 2026, with deployment anticipated in 2027 [1][2]. - The Baltra chip is designed to meet Apple's significant AI inference needs, as the company has opted not to train large models internally but instead will pay approximately $1 billion annually to use Google's custom Gemini model [2][7]. Group 2: Technical Specifications and Design Focus - The architecture of the Baltra chip will prioritize latency and throughput optimization, utilizing low-precision computation designs such as INT8 precision, which is crucial for inference tasks [2]. - Apple may adopt a chip architecture similar to Nvidia's GB300, featuring 64 interconnected chips with high-bandwidth LPDDR memory, potentially reducing costs while meeting specific performance requirements [2]. Group 3: Future Directions and Innovations - Apple has submitted a patent for an "optical-based distributed unified memory system," suggesting a focus on advanced computing technologies that could influence the design of future chips like Baltra [4][6]. - The development of the Baltra chip is seen as a critical step for Apple to enhance its AI capabilities and regain competitiveness in the AI sector, especially after setbacks in large model training [7].