Google unveils 8th-gen TPUs to power agentic AI era

Google Cloud announced a strategic overhaul of its custom artificial intelligence chip lineup at the Cloud Next 2026 conference on Wednesday, introducing the eighth-generation Tensor Processing Unit (TPU) in two distinct variants: the TPU 8t for training and the TPU 8i for inference.

This marks the first time Google has split its flagship accelerator into specialized chips, moving away from the one-size-fits-all model to address the specific demands of the "Agentic Era".

Specialized architecture for a new workload

TPU 8t (Training): Engineered for massive scale, the 8t allows a single cluster to scale up to 9,600 chips with 2 Petabytes of shared High Bandwidth Memory (HBM). Google states this delivers nearly three times the compute performance compared to the previous Ironwood generation, cutting development cycles for frontier models from months to weeks.

TPU 8i (Inference): Designed to break the "memory wall," the 8i prioritizes low latency. It features 288GB of HBM and 384MB of on-chip SRAM – three times more than its predecessor. This allows an AI model's active working set to reside entirely on the chip, significantly reducing response times for complex multi-agent tasks.

Efficiency and ecosystem

Google emphasized that both chips run on its in-house Axion ARM CPUs and utilize fourth-generation liquid cooling. The company claims an 80% improvement in price-performance for the 8i and up to a 2.8x gain for the 8t, allowing enterprises to serve twice the customers at the same cost.

In a related announcement, Google CEO Sundar Pichai reiterated the company's massive infrastructure push, revealing plans to invest $175 billion to $185 billion in capital expenditure this year.

(Cover via VCG)