---
title: Google launches TPU 8tx and TPU 8i Max processors for specialized artificial intelligence workloads
url: https://www.dataloco.com/en/google-launches-tpu-8tx-and-tpu-8i-max-processors-for-specialized-artificial-intelligence-workloads
published: 2026-11-12T19:23:20+00:00
language: en
section: AI
source: https://www.techtalkthai.com/google-launches-tpu-8tx-8i-max-confidential-inference-large-context/
organizations: Google
publisher: Dataloco
---

# Google launches TPU 8tx and TPU 8i Max processors for specialized artificial intelligence workloads

Google introduced two new processors in its eighth-generation TPU family on November 9. The TPU 8tx focuses on confidential inference with hardware-level data protection, while the TPU 8i Max combines two TPU 8i chips in a single package to support large context windows.

The TPU 8tx is designed for organizations in heavily regulated industries and government agencies that need to run models and artificial intelligence agents alongside sensitive data. Typically, data is encrypted during storage or network transmission but must be decrypted in memory during processing. The TPU 8tx embeds security capabilities at the hardware level to protect and isolate data and model weights even during computation. The system pairs with an AMD EPYC fifth-generation host CPU and uses the same low-latency Boardfly network as the TPU 8i, achieving latency up to two times lower than the TPU 8t with minimal performance impact. Organizations can use a single system to fine-tune models with internal data, distill large models into smaller ones, and train agents with reinforcement learning without moving sensitive data across clusters.

The TPU 8i Max addresses the memory bandwidth bottleneck that occurs when dozens of artificial intelligence agents work together on entire codebases or multi-hour videos. By combining two TPU 8i chips in one package, the system delivers double the compute, SRAM, and HBM bandwidth of a single TPU 8i. This allows data centers to fit twice as much silicon in the same rack space, increasing processing power per square foot and per watt.

The eighth-generation TPU family now divides into four specialized systems: the TPU 8t for model training, the TPU 8i for low-latency reasoning, the TPU 8tx for confidential computing, and the TPU 8i Max for large context windows and maximum compute density per rack. This marks the first time Google has launched four TPU systems in a single year. The TPU business currently serves nearly 100 customers including frontier laboratories, trading firms, retail businesses, and universities. Customers can now deploy TPU systems in their own data centers.

Google released TorchTPU as open source to run PyTorch code and models from Hugging Face directly on TPU without code modifications, along with backends for vLLM and SGLang. The company also opened an experimental TPU agent suite that works with Claude Code, Cursor, and Google Antigravity to enable TPU usage within minutes, analyze bottlenecks, and autotune custom kernels using natural language. The XLA Serving Engine allows training models in Python and serving them on a C++ runtime across TPU, GPU, and CPU systems. The Agent Compute Engine automatically suspends and resumes agent sessions waiting for responses to reduce infrastructure costs.

For managing large clusters, Google released TPU Sync as open source, which unifies HBM from chips, DRAM from hosts, shared storage, and data center networking into a single memory hierarchy. This reduces latency in sending KV cache and time to first token by more than 90 percent. Cloud Pathways is now generally available on Google Kubernetes Engine, allowing control of multiple TPU slices from a single controller as if they were one machine. GKE dynamic slicing, also now generally available, uses optical circuit switching to combine or divide TPU slices instantly without creating new node pools, enabling jobs to start up to five times faster and recover from failures up to 4.5 times faster.