At the Google Cloud Next 2026 conference, Google officially rolled out its 8th-generation Tensor Processing Unit (TPU), signaling a pivotal transformation in its AI hardware strategy. For the first time, it has allocated AI training and inference tasks to two distinct chips—TPU 8t and TPU 8i. This strategic move is rooted in the divergent demands of AI computing workloads: training tasks emphasize throughput and scalability, whereas inference tasks are more attuned to latency and concurrency.

The TPU 8t, a product of joint design efforts by Google and Broadcom, is specifically engineered for training ultra-large-scale AI models. On the other hand, the TPU 8i, developed in collaboration with MediaTek, zeroes in on AI inference scenarios. Both chips are propelled by Google's custom-built Axion CPU and are fabricated using TSMC's advanced 2nm process, with mass production slated for the end of 2027.

From a software ecosystem perspective, these chips offer compatibility with mainstream AI frameworks. Moreover, Google has introduced the Gemini Enterprise Agent Platform and AI agent tools, with the ambition of diminishing the obstacles for enterprises to deploy AI agents and expedite the automation of business tasks.