Google boosts its TPUs for confidential inference and long contexts

Google adds two configurations to its TPU generation for AI: TPU 8tx for confidential processing, TPU 8i Max for heavy memory loads. The goal is to cover more enterprise use cases.

On November 9, 2026, Google unveiled two new configurations of its eighth-generation Tensor Processing Units, the TPU 8tx and TPU 8i Max. The first targets confidential inference on sensitive data, the second addresses use cases requiring large memory, such as long contexts or analysis of large codebases. In both cases, the company aims to better meet the needs of businesses and administrations, where performance and security constraints must coexist without too much conflict.

Deux TPU, deux problèmes très différents à résoudre

The TPU 8tx reuses the silicon of the TPU 8t and adds fifth-generation AMD EPYC host processors, as well as hardware security mechanisms to protect data and model weights during in-memory use. Google presents it as a solution for sectors subject to regulatory requirements, to run AI models and autonomous agents on sensitive information. The challenge is not only to encrypt storage or network transport but also to reduce exposure during computation.

Meanwhile, the TPU 8i Max is designed for workloads that consume large amounts of memory. Google cites agents capable of analyzing large codebases, long video sequences, or contexts exceeding one million tokens. To achieve this, the company combined two TPU 8i chips in a single package, doubling compute capacity, SRAM memory, and HBM memory bandwidth compared to the TPU 8i.

Un choix matériel qui s’accompagne d’un travail réseau et logiciel

Google pairs the TPU 8tx with its Boardfly network, designed for low-latency communications. According to the provider, this architecture can halve latency compared to the TPU 8t, without significant performance loss. This should enable chaining on the same infrastructure training, fine-tuning on private data, model distillation, and reinforcement learning of AI agents.

At the same time, the company continues to open its software environment. TorchTPU, released as open source, allows running PyTorch code and Hugging Face models on its accelerators. Integrations also exist for the vLLM and SGLang inference engines, while a suite of TPU agents remains in preview to automate integration, diagnostics, and kernel optimization. Google also announces compatibility with Claude Code, Cursor, and Google Antigravity.

Pourquoi cette annonce compte pour le cloud IA

Since April 2026, Google had already introduced the TPU 8t, intended for training large models, and the TPU 8i, optimized for low-latency inference and reasoning. With these two new variants, the eighth-generation lineup expands to four distinct systems. The idea is clear: make the infrastructure more specialized, at the cost of a more complex offering for customers. The cloud sometimes loves long menus, especially when the bill remains understandable.

Google also adds TPU Sync, a software layer designed to unify management of HBM memory, host DRAM, and shared storage. The tool aims to facilitate transfer of KV caches used by generative models and, according to the company, reduce certain latencies related to these transfers and the generation time of the first token by over 90%. Finally, Cloud Pathways is now generally available on Google Kubernetes Engine, with the ability to orchestrate distributed workloads across multiple TPU sets from a single controller.

Key points

  • November 9, 2026: Google introduces TPU 8tx and TPU 8i Max.
  • TPU 8tx targets confidential inference on sensitive data.
  • TPU 8i Max doubles compute, SRAM, and HBM bandwidth.
  • Boardfly can reduce latency by up to two times.
  • TPU Sync aims for over 90% latency reduction.
  • Google had already launched TPU 8t and TPU 8i in April 2026.

By the numbers

  • 2 chips — combined in a single package for TPU 8i Max.
  • 2 times — latency reduction announced with Boardfly compared to TPU 8t.
  • over 90% — latency decrease announced with TPU Sync.
  • 4 systems — number of eighth-generation TPUs announced in 2026.
  • 1 million tokens — context threshold cited for certain TPU 8i Max uses.

To read