Particle.news

Google Adds Confidential and Long‑Context Variants to Its Split TPU 8 Family

The change aims to let Google sell compliance-ready hardware to regulated buyers via expanded cloud partnerships.

Overview

  • Google, which announced the new variants on November 9, 2026, expanded its eighth-generation TPU lineup with the TPU 8tx for confidential training and the TPU 8i Max for low-latency, long-context inference.
  • The TPU 8tx uses hardware-level encryption developed with AMD so sensitive training data can be protected at the chip level for regulated customers and government agencies.
  • The TPU 8i Max increases compute density and memory bandwidth and adds more on-chip SRAM to serve models that must hold large context windows with lower latency and cost.
  • Both eighth-generation chips use TSMC’s N3 process and Google’s Arm-based Axion CPUs and plug into Google’s AI Hypercomputer, and Google plans a phased multi-cloud rollout that includes Oracle Cloud Infrastructure as a channel starting in 2027.
  • The split between training and inference silicon is meant to cut cost and latency for real workloads and to make Google Cloud more competitive for enterprises that need compliance and long-context AI, a shift that pushed Alphabet shares higher in premarket trading.