Google will increase office space at its AI infrastructure R&D center in Taipei's Shilin district by 60%, Google CTO and Senior Vice President of AI Infrastructure Amin Vahdat said during his opening keynote at SEMICON Taiwan 2026 on Sept. 2. The announcement comes in Google Taiwan's 20th anniversary year, as the site has become Google's largest hardware engineering R&D base outside the United States. Vahdat described Taiwanese companies as evolving from manufacturing partners into "co-design partners" working "shoulder-to-shoulder" with Google. He said the rise of Agentic AI (software agents interacting autonomously) is likely to expand computing demand sharply. Google's TPU (Tensor Processing Unit) development cycle has shortened from about three years to one year, with two products launched in the latest year, and could eventually move to six-month or shorter intervals. The company is pursuing "Square Wave Deployment," a production model designed to move validated chips rapidly from initial units to full-scale capacity. Its latest TPU-8 provides a further 2x improvement in performance per watt, while multiple generations have built an approximately 100x advantage over traditional CPU computing in performance, power consumption and cost efficiency. Taiwan's supply-chain role spans 3D packaging, server motherboards, power delivery, liquid cooling, networking and other infrastructure components. Google said its systems can synchronize 9,600 TPUs in a single training system, nearly 100,000 through its data-center network and approximately 1 million across data centers using wide-area networks and submarine cables. Vahdat identified energy supply as the biggest constraint on data-center expansion, with Google's key metric shifting toward Delivered Value per Watt. Google's current data center PUE (power usage effectiveness) is approximately 1.09, meaning cooling and power-delivery overhead adds about 9% to IT energy consumption. Future development is expected to focus on larger AI packages, closer compute-memory integration, better data transmission, thermal management and power delivery, and greater optical communication integration.