Zhi Tong Cai Jing·2025-11-25 03:12Core Insights - Google is significantly upgrading its competition with Nvidia in the AI chip market, with Meta emerging as a potential multi-billion dollar customer [1][2] - Google is now marketing its custom Tensor Processing Units (TPUs) for deployment in customer data centers, marking a strategic shift from its previous model of limiting TPU use to its own cloud data centers [1][2] - Meta is reportedly in discussions to invest billions to integrate Google's TPUs into its data centers starting in 2027, while also planning to rent TPUs from Google Cloud as early as next year [1][2] Industry Impact - The explosive growth in AI computing demand and Nvidia's continued dominance in the supply chain indicate that Google's push for TPUs signifies a more competitive phase in the AI chip war [2] - Google Cloud executives have suggested that expanding TPU adoption could help the company capture 10% of Nvidia's annual revenue, representing a multi-billion dollar opportunity [2] - If the deal with Meta is finalized, it would validate Google's hardware ambitions and establish TPUs as a viable alternative to Nvidia's chips [3] Market Reactions - Following the news, Google's stock rose over 3% in after-hours trading, while Nvidia's stock fell by more than 2% [3] - Analysts suggest that Meta's potential use of Google's TPUs indicates that third-party large language model providers may consider Google as a secondary supplier for inference acceleration chips [3] - Related Asian stocks saw significant gains, with IsuPetasys in South Korea rising 18% and MediaTek in Taiwan increasing nearly 5% [3] Technological Advancements - Google's TPU technology, particularly the seventh-generation "Ironwood," is designed to handle a variety of computing tasks from training large language models to real-time chatbot operations [4] - The Ironwood chip can connect up to 9,216 TPUs in a single rack, effectively eliminating data bottlenecks for complex models and enabling customers to run and scale the largest AI models [4]