• Sources: primary, primary, primary
  • Summary: Google added two eighth-generation TPU systems on 2026-11-09: TPU 8tx, built on TPU 8t silicon with custom AMD CPU hosts for confidential inference and small-scale training, and TPU 8i Max, which packs two TPU 8i chips for double the compute, SRAM and HBM bandwidth. Google also open-sourced TPU Sync for KV cache tiering and RL weight sync, and made Cloud Pathways and GKE dynamic slicing generally available on GKE. Google says dynamic slicing cuts job startup times by up to 5x.
  • Why it matters: Teams running inference or RL on Google Cloud get confidential computing on TPUs and GA tools for multi-slice jobs, and PyTorch and vLLM code can target TPUs through the TorchTPU stack.
  • Background: TPUs are Google's custom accelerators for machine learning. Confidential computing keeps data and model weights encrypted and isolated in memory while they are processed, not only at rest and in transit. Pathways is the single-controller runtime Google DeepMind uses to drive many TPU slices as one program.