硬AI·2025-06-17 14:30Core Viewpoint - Meta is planning to launch several high-spec AI ASIC chips between the end of 2025 and 2026, with a total expected output of 1 to 1.5 million units, potentially surpassing NVIDIA's GPU shipments at some point in 2026 [1][2][5]. Group 1: Market Dynamics - Currently, NVIDIA holds over 80% of the AI server market value, while ASIC AI servers account for only 8-11% [4]. - By 2025, Google is expected to ship 1.5 to 2 million TPU units, and AWS's Trainium 2 ASIC is projected to reach 1.4 to 1.5 million units, compared to NVIDIA's AI GPU supply of over 5 to 6 million units [4]. - Supply chain research indicates that the combined shipment of Google and AWS's AI TPU/ASIC has reached 40-60% of NVIDIA's AI GPU shipments [5]. Group 2: Meta's MTIA Project - Meta's MTIA project is a significant case in the current ASIC wave, with the first ASIC chip, MTIA T-V1, set to launch in Q4 2025, designed by Broadcom and featuring a complex 36-layer PCB architecture [8]. - The MTIA T-V1.5, expected in mid-2026, will double the chip area and exceed NVIDIA's next-generation GPU specifications, while the MTIA T-V2 in 2027 may introduce larger CoWoS packaging and high-power rack designs [8]. Group 3: Challenges and Competition - Meta aims to achieve 1 to 1.5 million ASIC shipments by the end of 2025 to 2026, but current CoWoS wafer allocation can only support 300,000 to 400,000 units, indicating potential production bottlenecks [9]. - NVIDIA is not passive; it plans to introduce NVLink Fusion technology at the 2025 COMPUTEX, allowing seamless integration of third-party CPUs or xPUs with its AI GPUs, which is part of its strategy to maintain market share [12]. - Despite the rise of ASICs, NVIDIA remains ahead in chip computing density and interconnect technology, making it difficult for ASICs to catch up in performance [13].