Subscribe
0%
Illustration for: Google adds TPU 8tx for confidential AI and dual-chip TPU 8i Max. What it announced, and what it didn't

Google adds TPU 8tx for confidential AI and dual-chip TPU 8i Max. What it announced, and what it didn't

Google's two new eighth-generation TPU systems target regulated customers who want hardware-isolated inference and agent workloads that are limited by memory bandwidth. Google gave no pricing, ship dates or comparisons with Nvidia.

Listen to this article0:00 / 4:23

Google announced two more eighth-generation Tensor Processing Unit systems on November 9. TPU 8tx is built for confidential AI inference and training, and TPU 8i Max puts two TPU 8i chips in one package. Amin Vahdat, Google's chief technologist for AI and infrastructure, said in a company blog post that they join the TPU 8t and TPU 8i launched at Cloud Next in April. That makes four new TPU systems in one year, which Google calls a first.

Why it matters: rather than designing new chips, Google is reworking the memory, networking and security around existing silicon to fit two buyer groups that general-purpose GPUs don't specifically target. One group is regulated organizations that won't send sensitive data or model weights to shared hardware. The other is agent-heavy workloads where, in Google's words, "memory bandwidth quickly becomes the bottleneck." ITdaily reports Google's argument that this approach takes months, where a new chip takes years.

The two new systems

TPU 8tx TPU 8i Max
Silicon Same as TPU 8t Two TPU 8i chips in one package
Host 5th-gen AMD EPYC CPUs —
Network Boardfly (from TPU 8i) —
Google's headline claim Up to 2x lower latency than TPU 8t, with a "near-zero performance penalty" for confidential compute 2x the compute, SRAM and HBM bandwidth of TPU 8i, in the same rack footprint
Aimed at Governments and regulated enterprises, private fine-tuning, secure agent execution Long-context and multi-agent inference

Google says TPU 8tx runs fine-tuning, distillation and reinforcement learning on one system, so sensitive data doesn't move between clusters. Decrypt notes a limit of confidential computing: it isolates data and model weights from the infrastructure, but it can't stop a badly designed agent from mishandling data it is allowed to access.

For context, Google says the April chips deliver up to 2.7x better training performance (TPU 8t) and 80% better inference price-performance (TPU 8i) than the previous TPU generation.

The software around them

Google spent much of the post on software meant to make TPUs easier to adopt for teams used to GPUs:

  • TorchTPU (open-sourced last month): runs native PyTorch and Hugging Face models on TPUs, with vLLM and SGLang backends.
  • TPU Sync (open source): moves the KV cache between chips, servers and storage. Google says it cuts KV transfer latency and time-to-first-token by more than 90%.
  • TPU agent suite (open source, in preview): integrates with Claude Code, Cursor and Google Antigravity.
  • XLA Serving Engine (open source): a C++ runtime that serves models across TPUs, GPUs and CPUs.
  • Cloud Pathways and GKE dynamic slicing, now generally available. Google claims up to 5x faster job startup and up to 4.5x faster fault recovery using optical circuit switching.

Google also says TPUs now have nearly 100 customers, including frontier AI labs, trading firms, retailers and universities.

What Google didn't say

  • No pricing and no availability dates for either system, as ITdaily and Decrypt both note. The blog post gives none either.
  • No named customers for the new configurations.
  • No head-to-head with Nvidia. Every performance figure compares Google's hardware to earlier Google hardware (Decrypt). Google cites results on SemiAnalysis's InferenceX and AgentX benchmarks, but without numbers in the post.

The launch comes during a rough stretch for Nvidia. Decrypt reports Nvidia shares fell 9.3% on Thursday and another 5% on Monday amid the controversy over CEO Jensen Huang. Decrypt also mentions an unconfirmed report that OpenAI plans to buy TPUs. TechMeasured hasn't verified that report.

Who should care

  • Regulated teams (finance, health, government) that have kept frontier-model fine-tuning on-prem for data-control reasons should ask Google Cloud for TPU 8tx's attestation model and timeline once details are out.
  • Teams running long-context agents on TPU 8i should watch for 8i Max pricing. Doubling HBM bandwidth per package could cut the number of chips a large-context deployment needs, if the price per package doesn't double too.
  • PyTorch shops can test TorchTPU and TPU Sync now, without committing to new hardware.

Sources

  1. Google — Our newest TPUs: 8tx and 8i Max, built for the agentic era blog.google
  2. ITdaily — Google expands eighth TPU generation with TPU 8tx and TPU 8i Max itdaily.com
  3. Decrypt — Google Takes Aim at Nvidia With New AI Chips Built for Privacy and Massive Memory decrypt.co

#Google#TPU 8tx#TPU 8i Max#Google Cloud#AI chips#confidential computing#AMD EPYC#Nvidia