突发，Meta放弃一颗自研芯片，拥抱谷歌TPU

Core Insights - Meta has faced significant challenges in the development of its custom chips, leading to the abandonment of both the Iris and Olympus training chips [1] - The company has entered into a multi-billion dollar agreement with Google to rent AI chips, intensifying competition with Nvidia and Google in the chip market [1] Group 1: Custom Chip Development - Meta's strategy to develop custom chips aims to overcome the limitations of existing AI accelerators, with projected R&D spending of approximately $50 billion and capital expenditures between $66 billion and $72 billion by 2025 [2] - The company intends to design its own CPU and XPU, while also pushing interconnect ASIC manufacturers to meet its needs, threatening to develop its own interconnect structures if necessary [2] - Meta has opted to use the open-source RISC-V architecture instead of the licensed but closed-source Arm architecture for its future computing engines [2] Group 2: MTIA Chip Series - The MTIA v1 chip, launched in May 2023, is designed for inference rather than training, with a manufacturing process of 7nm, a frequency of 800 MHz, and a TDP of 25W [3][4] - MTIA v2, set to be released in April 2024, features a significant performance increase with a frequency of 1.35 GHz, a TDP of 90W, and improved SRAM capacity, although it still does not support training [5][8] - Both MTIA chips utilize a RISC-V core architecture, with MTIA v1 deployed moderately in Meta's data centers and MTIA v2 expected to have a larger deployment scale [9] Group 3: Acquisition of Rivos - Meta's acquisition of AI chip startup Rivos in October 2025 is seen as a strategic move to enhance its capabilities in developing high-end RISC-V chips tailored for AI workloads [12][16] - Rivos, founded in September 2021, has a strong team with experience from major tech companies and has developed a 3.1 GHz processor compatible with CUDA, facilitating the transition to RISC-V hardware [13][15] - The acquisition is expected to provide Meta with a competitive edge in the AI chip market, allowing for customized solutions that can compete with Nvidia and AMD [16] Group 4: Market Dynamics - Meta's recent agreements with Nvidia and AMD for GPU transactions indicate a strategy to bolster its computational power while mitigating risks associated with its custom chip development [17] - The collaboration with Google on TPU rental services represents a significant shift in the competitive landscape, as Google aims to challenge Nvidia's dominance in the AI chip market [19] - Meta's ongoing efforts to develop training chips highlight the company's ambition to establish a foothold in the AI hardware sector, despite facing numerous setbacks [20]