Google details inference-focused TPU variants planned for 2027

Google is preparing two TPU accelerator variants for installation in 2027, targeting simultaneous agent workloads and inference with larger amounts of context, The Next Platform reported. The TPU 8tx replaces the training-oriented interconnect with a smaller network whose average latency is about half that of the original topology. The TPU 8i Max combines two accelerator chips per socket, increasing memory capacity per socket by 33% while reducing memory per chip. Google has also extended TPU Sync software to TPU generations 5 through 8. It moves an agent's working memory off the accelerator while the agent pauses, freeing resources for other tasks. Pricing for the new variants and measured throughput gains from TPU Sync remain unresolved.