Workflow
v8e
icon
Search documents
谷歌TPU,卖爆了
半导体行业观察· 2025-12-16 01:22
Core Insights - Google is significantly increasing its orders for Tensor Processing Units (TPUs) from MediaTek, with the order volume exceeding initial plans by multiple times [2] - MediaTek's first TPU, the v7e, is set to enter risk trial production by the end of next season, and it has also secured orders for the next-generation TPU, v8e [2][3] - The collaboration with TSMC is expected to boost MediaTek's production capacity for Google projects, with a projected sevenfold increase in CoWoS capacity by 2027 [2][3] Group 1 - The demand for Google TPUs is driven by strong client needs, leading to an increase in CoWoS capacity from 10,000 to 20,000 units annually for the v7e project [3] - MediaTek's ASIC business is anticipated to contribute significantly to its revenue, with estimates suggesting that the v7e could add over two times its equity in profits by 2027 [2][3] - MediaTek's CEO expressed confidence in the growth of ASIC revenue, targeting $1 billion in cloud-related ASIC revenue by 2026 and potentially reaching several billion by 2027 [4] Group 2 - Meta is exploring a partnership with Google to utilize TPUs for its AI projects, which could challenge NVIDIA's market dominance [6][7] - If successful, Meta plans to start renting TPUs from Google Cloud in 2026 and deploy them in its data centers by 2027, marking a significant shift in its AI infrastructure [6][7] - The potential collaboration has led to a drop in NVIDIA's stock price, reflecting investor concerns about Meta's future chip orders [7] Group 3 - Analysts predict that Google plans to double its TPU production by 2028, with TSMC expected to produce 3.2 million TPUs in 2024, increasing to 5 million by 2027 and 7 million by 2028 [10] - Morgan Stanley estimates that Google could generate up to $13 billion in revenue for every 500,000 TPUs sold to external clients [9] - Google's vertical integration strategy aims to enhance its technological advantages and profitability by developing its AI hardware and software [8][11]