Google adds two systems to its eighth-generation TPU portfolio

By Amara Osei

Google’s eighth-generation TPU portfolio now includes four systems after Google added the TPU 8tx and TPU 8i Max. The two offerings extend the TPU 8t training and TPU 8i inference designs that Google first unveiled in April 2026.

What does the expanded portfolio change?

By describing the products as “TPU systems” rather than chips, Google said it reconfigured compute, networking, storage and packaging around the TPU 8t and TPU 8i. TPU 8tx adds a confidential-computing configuration; TPU 8i Max combines two inference chips in one package for higher compute and memory density.

Amin Vahdat, chief technologist and senior vice president for AI and Infrastructure at Google, wrote: “As enterprises bring autonomous agents into production, we have created two new configurations of these foundational systems to solve two urgent new challenges: hardware-level data security and million-token context scale.”

The result is four systems tied to training, reasoning, protected inference and high-density inference.

How does TPU 8tx protect data while models run?

The TPU 8tx uses the same silicon as TPU 8t and pairs it with custom AMD CPU hosts. Google optimized the system for “secure and confidential inference,” but it can also support small-scale training. Google says the system best suits regulated enterprises and governments running frontier or proprietary models and autonomous agents alongside sensitive data.

Encrypted data normally requires unlocking in memory when stored or sent over a network for processing. TPU 8tx builds security into the hardware layer, keeping sensitive data and model weights protected and isolated even during active use.

Google’s Boardfly network supports real-time reasoning, which the company says keeps multi-step agents moving. TPU 8tx can deliver up to 2x lower latency than TPU 8t, with a near-zero performance penalty.

Organizations can run and customize models simultaneously on one system. Those functions include:

• Fine-tuning models with private data.
• Distilling large models into smaller ones.
• Training agents through reinforcement learning.
• Completing those steps without migrating sensitive data between separate clusters.

TPU 8tx moves protected inference and model adaptation into one system, reducing the need to transfer sensitive data among clusters.

How does TPU 8i Max address the memory wall?

Google designed TPU 8i Max for the “memory wall,” the gap between processor speed and memory bandwidth. Two TPU 8i chips share one package, allowing Google to scale compute and memory together. Compared with TPU 8i, the Max system provides twice the compute, twice the SRAM and twice the HBM bandwidth.

That balance keeps required data available to processors and prevents large workloads from stalling. The system also fits twice the silicon within the same rack footprint, so Google said data centers can deploy significantly more computing power per square foot and per watt.

TPU 8i Max doubles compute and memory bandwidth without requiring a larger rack footprint.

Which workloads do the four systems cover?

Vahdat said the systems provide a purpose-built accelerator for every stage of the agentic lifecycle while remaining flexible across workloads. He assigned them specific roles: “TPU 8t for scale-up training, TPU 8i for low-latency reasoning, TPU 8tx for confidential computing, and TPU 8i Max for massive context windows and the highest compute density per rack.”

He said computing is moving from human scale to agent scale. That shift increases business demand for accelerators that can reason across large AI-model context windows, while highly regulated customers are seeking specialized encryption capabilities for particular workloads. Google says the two new systems address those requirements.

TPU 8tx targets protected inference and agent latency; TPU 8i Max targets context scale and compute density.