3 6 Ke·2025-09-15 10:06Core Insights - Google has begun selling its Tensor Processing Units (TPUs) to small cloud service providers, aiming to compete with NVIDIA in the AI computing market [1][2] - The competition between Google and NVIDIA is intensifying, with analysts predicting a significant reduction in NVIDIA's GPU sales due to the rise of TPUs [2] - Google has been developing TPUs since 2013, initially to address increasing computational demands for AI tasks [3][4] TPU Design and Features - TPUs are specialized ASIC chips designed for AI computing, focusing on high matrix multiplication throughput and energy efficiency [4] - The architecture of TPUs utilizes a "Systolic Array" design, allowing for high data reuse and reduced memory access latency [4] - Google has developed a series of TPU versions, with the latest, Ironwood, achieving peak performance of 4614 TFLOPs and supporting advanced computing formats [9][10] Market Position and Future Projections - By 2025, Google is expected to ship 2.5 million TPUs, with a total projected sales exceeding 3 million by 2026 [8] - The growing acceptance of TPUs reflects a shift in the market as companies seek alternatives to NVIDIA's GPUs for better cost-effectiveness and supply chain stability [15] - Analysts suggest that if Google merges its TPU business with DeepMind and spins it off, it could be valued at up to $900 billion [12] Competitive Landscape - Other tech giants like Meta, Microsoft, and Amazon are also developing their own ASIC chips, indicating a broader trend of moving away from NVIDIA's dominance [15][17] - The competition is not limited to Google; Meta plans to launch its first ASIC chip by late 2025, further intensifying the market rivalry [15][16] - NVIDIA is responding to this competition by introducing technologies like NVLink Fusion, which allows for mixed-use of its GPUs with third-party accelerators [17]