Google today unveiled its eighth-generation TPU chips at Cloud Next '26, launching two new products optimized for different AI workloads: TPU 8i and TPU 8t. This breakthrough marks a brand-new stage for AI hardware inference.
TPU 8i is purpose-built for autonomous AI agents — capable of quickly completing reasoning, planning, and executing multi-step workflows, dramatically improving user experience. Its dedicated architecture addresses the performance bottlenecks that AI agents face on complex tasks.
TPU 8t, on the other hand, focuses on model training optimization, supporting the most complex models running on a single massive memory pool — dramatically simplifying the training workflow for large-scale AI models while improving compute efficiency.
Combined with Google's full-stack infrastructure, these two chips will popularize highly responsive agentic AI and set a new standard for AI inference performance. This specialized hardware design represents an important trend in AI infrastructure's evolution from general-purpose compute to purpose-built silicon — providing strong support for the next generation of AI applications.