Google announces TPU 8tx and TPU 8i Max

Google has announced two additions to its eighth-generation TPU portfolio: TPU 8tx, aimed at secure and confidential inference, and TPU 8i Max, designed for larger context windows and higher compute density. The company says the pair brings the number of eighth-generation TPU systems introduced in a single year to four. ¹

What Google is announcing

TPU 8tx is positioned as Google’s security-focused system for inference workloads involving sensitive data and model weights. Google says it provides hardware-level protection while information is in use, uses the same silicon as TPU 8t and fifth-generation AMD EPYC CPU hosts, and can deliver up to twice the latency performance of TPU 8t with what it describes as a near-zero performance penalty. Those performance figures are Google’s own claims, not independent benchmarks. ¹

TPU 8i Max takes a different approach. Google says it combines two TPU 8i chips in one package, doubling the compute, SRAM and HBM bandwidth available from the standard TPU 8i configuration. The result is intended to handle massive context windows while increasing compute density within the same rack footprint.

The announcement also covers a wider software and access push. Google says TorchTPU is now open source, a TPU agent suite is available in preview, and XLA Serving Engine and TPU Sync are part of the updated platform. Cloud Pathways on Google Kubernetes Engine is generally available, while dynamic slicing is intended to give customers more flexible access to TPU capacity. ¹

Google’s announcement places the two systems inside its AI Hypercomputer strategy, alongside TPU 8t for large-scale training and TPU 8i for real-time reasoning. The company says the portfolio now serves different stages of the AI lifecycle, from training and inference to workloads that need stronger security or unusually large working context.

Availability, pricing and customer access details for TPU 8tx and TPU 8i Max are not fully set out in the announcement. Google also says its comparisons and energy or cost claims are based on its own testing, so buyers will need independent figures before making a serious infrastructure decision. ¹

Our opinion

The interesting part of this announcement is not simply that Google has added two more accelerators. It is the attempt to split AI infrastructure into more specialised jobs: training, fast reasoning, confidential inference and context-heavy workloads. That is a more practical pitch than promising one universal chip will do everything equally well.

The caveat is that the useful comparison remains missing. Until Google publishes clear pricing, availability and independently reproducible performance data, TPU 8tx and TPU 8i Max are platform announcements rather than straightforward alternatives to the latest Nvidia or AMD systems. The hardware direction is credible; the buying case still needs numbers.

The official Google Cloud broadcast linked from the announcement provides the accompanying infrastructure context. ²

¹ https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/² https://youtube.com/live/pabDkNR7jIs